A transparent overview of the analytical pipeline powering GEOPOL Analytics — from raw data ingestion to strategic intelligence synthesis. Grounded in quantitative historical analysis and causal mechanism detection, each stage is designed around academically validated approaches adapted for real-time geopolitical monitoring. The page describes the system as it runs today; every change of method is dated in the Methods Journal at the end.
The system draws inspiration from cliodynamics (Peter Turchin, Historical Dynamics, Secular Cycles) and quantitative history — the mathematical modeling of historical processes. However, it does not assume historical recurrence or cyclical determinism.
Instead, it operates at the edge of cliodynamics, in what might be called Macro-History: a framework where events never repeat, but causal mechanisms are timeless. The same strategic logic that drove the Delian League drives NATO. The same speculative mechanics that inflated Dutch tulips inflated subprime mortgages. The same dependency trap that forced Sully to cultivate mulberry trees in 17th-century France constrains Japan’s T-bond strategy today.
“History doesn’t stutter, but it sure has a tendency to rhyme.” — attributed to Mark Twain
The system classifies causal chains using a taxonomy inspired by Boyer’s crisis transmission mechanisms, adapted for real-time detection:
| Mechanism | Description | Historical Example |
|---|---|---|
| A — Speculative Bubble | Price-value disconnection, herd behavior, leverage amplification | Tulip Mania (1637), Assignats (1795), 2008 Financial Crisis |
| B — Trade Contagion | Supply chain disruption, trade dependency, competitive devaluation | 1930s Smoot-Hawley, 2020 supply chain crisis |
| C — Monetary Transmission | Interest rate divergence, capital flight, currency crisis | 1997 Asian crisis, negative real rates → gold rush |
| D — Liquidity Drain | Central bank tightening, credit freeze, repo market stress | 2019 Repo crisis, FED QT → emerging market stress |
| E — Strategic Dependency | Resource dependency, encirclement, alliance restructuring | Sully’s mulberry cultivation (17th c.), Japan T-bond dilemma, OPEC embargo 1973 |
The core of the system is not a Large Language Model. It is a causal intelligence engine that learns autonomously from three sources:
LLMs are used downstream for report generation and natural language synthesis — they consume the engine’s output, they do not produce the intelligence.
| Class | Origin | Description | Example |
|---|---|---|---|
| A | Literature-derived | Coefficient sourced from a peer-reviewed publication or established theory | Boyer mechanisms, Dempster-Shafer combination rule |
| B | Empirically calibrated | Parameter optimized on historical data with measurable fit | Factor Z corpus thresholds (V3.2), Oracle hit rate weights |
| C | Expert heuristic | Reasoned choice by domain expertise, not formally calibrated | tension = bel×0.6 + pl×0.4, confidence formula |
| D | Engineering constraint | Choice dictated by CPU, RAM, latency, or API quota limitations | TTL 6h, ±3-day lag window, batch sizes, concurrency limits |
Parameters marked (C) are candidates for future empirical calibration as more data accumulates. The system’s learning loop (§07) progressively promotes C-class parameters to B-class as validation evidence grows.
| Class | What it carries |
|---|---|
| Measured | Its source, its date and its age. |
| Estimated | Its method and its bounds: an estimate is an interval when its uncertainty is known, and the Dempster-Shafer engines treat the width of that interval as ignorance (§4.5). |
| Static | A sourced reference with its vintage (USGS, IEA…), presented as such. |
| Absent | None with its provenance — never 0, 50, 100, a mean, an invented share, the value
of another series or an old undated value. |
No numeric fallback enters a calculation: a default value never feeds a score, a composite, a conviction or an input of the Trade-Bot. A source that fails says so (failure with its reason, never an empty success or zeros); a view that is not wired says “not available” instead of showing plausible values; a label never calls a value measured or live when it is not. The data audit of 16 October 2026 lists the components not yet brought into line. Tests built on doubles (fake APIs, fake caches) prove the logic, not the real data.
Each section of this page describes its module as it runs today. When a method changes, its section is rewritten and the change is dated in the Methods Journal (§08).
The intelligence pipeline begins with continuous, multi-source data ingestion across four complementary collection channels: structured news feeds, social media platforms, institutional APIs, and real-time sensor streams. This layered approach follows the principle of source triangulation — no single data channel is trusted in isolation.
All collection is orchestrated by a distributed task scheduler running 180+ scheduled tasks across 6 specialized worker queues, ensuring that data freshness, API quotas, and system resources are balanced without manual intervention.
The RSS layer ingests from 900+ curated feeds covering news agencies, think tanks, institutional sources, and regional media in 5 languages (French, English, Russian, Chinese, Italian). Feeds are classified by quality tier (high/medium) and category (media, think tank, institution, social, data), allowing downstream analysis to weight source credibility accordingly.
Geographic coverage is deliberately asymmetric, emphasizing regions with higher geopolitical signal density: Francophone Africa, the Middle East, Eastern Europe, and the Indo-Pacific arc. Western mainstream media provides baseline context, while regional outlets surface early signals that international press typically picks up 24-48 hours later.
Social media collection operates across 8 active platforms, capturing public discourse from diverse linguistic and geopolitical spheres. The system uses a privacy-preserving architecture: where possible, collection routes through privacy-focused frontends or anonymization layers rather than direct platform APIs, reducing fingerprinting and rate-limit exposure.
Each platform connector implements adaptive resilience — instance rotation, blacklisting on error, configurable timeouts, and graceful degradation. If a source becomes temporarily unavailable, the pipeline continues with remaining sources rather than blocking.
| Platform | Method | Geo focus | Resilience |
|---|---|---|---|
| Privacy frontend + fallback | US, International | Instance rotation, anonymization fallback | |
| Telegram | Public channel preview | RU, Middle East, EU | Anonymized access, per-channel timeout |
| Bluesky | AT Protocol API | US, EU | Native API, session caching |
| VKontakte | Official API v5 | Russia, CIS | Rate-limited (3 req/s), anonymization fallback |
| YouTube | API + privacy mirror fallback | Global | Dual connector with automatic failover |
| Mastodon | Fediverse API | EU, Tech | Multi-hashtag, public timeline |
| Public AJAX endpoints | China | Trending + search, rate-limited | |
| Hacker News | Firebase API | Tech, US | Public, real-time |
Posts are fetched in parallel using a 6-thread pool with a 5-minute global timeout, then normalized to a common schema (id, content, author, engagement metrics, publication date, source country). Sentiment analysis is performed asynchronously in batches to avoid blocking the collection pipeline.
Beyond unstructured text, the system ingests structured quantitative data from 45+ institutional APIs spanning economic indicators, geopolitical event databases, security threat feeds, and environmental sensors. The vast majority operate on free or open-access tiers — the architecture prioritizes publicly available data to ensure reproducibility and independence from commercial providers.
Fertiliser flows by bloc. Once a week, UN Comtrade gives the imports of nitrogen (HS 3102), potash (3104) and compound fertilisers (3105), by supplier, for 2021 and the last two years. Each bloc has a core and a circle. The cores do not overlap: the 27 members of the European Union, the United States, and the members of the BRICS (Brazil, China, India, South Africa, Egypt, Ethiopia, Iran, the United Arab Emirates, Indonesia); a strategic alliance prevails over a trade agreement, so Brazil, Egypt and South Africa belong to the BRICS core. The circles may overlap, and every country that is also in another core is flagged: the European circle is the Commission’s official list of the 84 partners with a trade agreement in place (each with its agreement and status, list read on 25 October 2026); the American circle is the 20 partners of United States free-trade agreements, Japan and the European Union. The current composition applies to every year, so that years compare the same countries. The European core is read as one: Comtrade reports the Union’s imports from outside the Union by country of origin, so trade between members is not counted twice. Each member is still read on its own, with its take: its share of the members’ direct Russian and Belarusian imports divided by its share of their imports from outside the Union (above 1, a member takes more than its weight). Russia and Belarus no longer report their trade, so their sales are read from the importers’ declarations. For each importer the Russian and Belarusian share is an interval: its direct imports at the low end; at the high end, for each re-export hub (Belgium, the Netherlands), the smaller of the hub’s sales to that importer and the hub’s own Russian and Belarusian imports of the year; and an estimate that allocates the hub’s sales in proportion to its own Russian and Belarusian share. A circle’s share sums its members’ declared imports, with and without the countries shared with another core; the change since 2021 is computed on the countries that declared both years. A year not declared is absent; a supplier missing from a declared breakdown is a measured zero; a country that does not report to Comtrade (Kosovo, Liechtenstein, covered by Switzerland) is named, never estimated; a weight is published only when every country declares one, and weights estimated by Comtrade are flagged. A bloc whose requests all fail says so and keeps its previous reading, dated; a batch of countries that fails is named and the bloc is marked partial. An annual average hides turning points, so the cores and the large EU importers (the Union as a whole, the United States, the BRICS members, France, Germany, Italy, Spain, Poland, the Netherlands, Belgium) are also read month by month over the last 24 months: the direct Russian and Belarusian share over the three months ending at the last month a country has declared (all three declared, otherwise absent), against the same three months a year earlier, with the lag of the last declared month. Hubs are not untangled month by month, so the monthly share is the low end of the interval. The BRICS monthly share sums the members that declared the same three months in both years, and names them; the three months end at the members’ median last declared month, so that one member reporting late does not hold the bloc back, and the late members are named with their last month. A month that is not declared stays absent and is named; when it is the only missing month of the window and both its neighbours are declared, an estimated share is published beside the measure, never in its place: an interval in which the missing month’s share and volume range between the lowest and highest of its last twelve declared months (fertilisers arrive in cargoes, and one month can triple its neighbours), with the mean of the two neighbours as the central estimate.
Once a day (06:15), the system reads the official travel advisories of nine governments and converts each one to a common 1–4 scale (1 normal precautions, 2 increased caution, 3 reconsider travel, 4 do not travel). The national level is kept apart from advisories limited to a region of a country: a warning about one province does not raise the level of the whole country. An issuer that fails keeps its last levels, flagged as stale, for three days at most; if every issuer fails, nothing is overwritten. The daily snapshots are archived, and changes over 7 and 30 days are computed only for issuers that answered on both days.
| Issuer | Role | What a country’s absence means | How the level is read |
|---|---|---|---|
| United States, Canada, United Kingdom, Germany, Netherlands | Core (consensus basis) | Level 1 | Published levels, normalised to 1–4 |
| New Zealand | Extended | Level 1 | Published levels 1–4 |
| Japan | Extended | Level 1 | Four Japanese levels mapped to 2, 3, 4, 4; a national level only when the text covers the whole country, otherwise a partial advisory |
| China | Extended | No reminder in force (not “level 1”) | Inferred from the wording of dated safety reminders (e.g. “do not travel” = 4); a reminder is kept in force for one year (assumption) |
| France | Extended | Unknown (only red is readable) | Whole territory in red = 4; red zones only = partial advisory; read from the text of about 190 country pages |
Consensus, not maximum. The level used by OMEGA (number of countries advised against) and by the country risk score (5% component) is the second-highest level among the five core issuers, i.e. the level that at least two of them reach: a single issuer at level 4 does not set the score. An issuer that only publishes warnings (China) or only the red level (France) counts in a consensus only from level 2 upwards.
A deliberate choice: the consensus basis stays on the five core issuers.
The four newer issuers are computed in parallel (an extended block on every country)
but feed no score. Measured on the first complete reading (7 October 2026, nine issuers answering, 239 countries),
switching to all nine would raise the count of countries at consensus level 3 or more from 40 to 42,
move 10 countries up and none down, and change the country risk score by at most 1.5 points.
The change would come from allied issuers (Japan, New Zealand, France);
China, the only issuer outside the Western bloc, changes no consensus.
Widening the basis would therefore alter the series without making it more diverse.
The extended series accumulates so that the choice can be made with both measures in view.
A switch of basis would be dated in the Methods Journal and the earlier series treated as discontinuous.
Limits. Levels are not strictly comparable across issuers (different scales and definitions). China’s level is deduced from wording, not from a published scale; France’s reading depends on prose and is guarded by sanity checks (it is refused if the pages change format), not by a structured feed. Australia is not integrated yet.
Relations between States are stored as dated events, one row per target country, in one table
(relation_events; unique key: source, identifier in the source, actor, target, so that reading a text twice
never counts it twice). Five families of public sources feed it twice a day (06:40 and 18:40 UTC):
the US Federal Register (presidential documents, BIS, USTR, State Department), OFAC’s recent actions,
the EU sanctions map, and the United Kingdom and Canadian lists. Each source reports its own state
(read, discarded, new, suspect if nothing was read); a failing source never stops the others.
Bilateral GDELT. The system reads only the open GDELT event files (one file every 15 minutes; the paid “GDELT Cloud” and the rate-limited document API are not used), every hour, for 66 key countries. Events are aggregated per day and per directed pair: number of events, mentions, Goldstein scale, tone, four cooperation/conflict classes and counters of selected fact codes. The day of an event is the day it was added, never its stated date (about 1% of rows carry old dates copied from archives). Repeats of the same article, pair and fact are counted once (44% of rows in the key countries). A freshness check turns the source “late” after 90 minutes and “stale” after six hours; each day records how many files were expected and read, and a day that was poorly covered is ignored by the detector. History: 90 days, rebuilt automatically.
Shift detector. For an unordered pair and four measures (volume, cooperation, rupture, tone), the 7-day window is compared, on a logarithmic scale, with the 60 preceding sliding windows using the median and the median absolute deviation. A shift needs z ≥ 3.5 on at least 3 of the last 5 days, a weekly median of at least 14 events, and well-covered days. A GDELT deviation alone is only a “candidate” (kept for three days for checking and tuning). It becomes a published event only when a second, independent family corroborates it (state measures or travel-advisory changes involving the same pair).
Limits. GDELT codes events by machine from news text: it measures media attention to a relation as much as the relation itself. Region codes are never guessed to be countries. The detector measures a deviation from the pair’s own history; it does not say why.
Twice a day (07:50 and 19:50 UTC) the system reads the official documents of the UN Security Council:
the resolutions (S/RES/n(year)) and the verbatim records of meetings
(S/PV.n), published as PDF about two to three weeks after the meeting and not always in order.
A missing record is retried for 60 days, then dropped. Each resolution gives events against the States its text
concerns; each recorded vote on a draft resolution gives one event per State concerned, with a severity:
1 unanimity, 2 abstentions, 3 votes against, abstention of a permanent member
or rejection, 4 veto.
A vote is written only if it checks out. The lists “in favour / against / abstaining” extracted from the PDF must give exactly the counts announced aloud by the President (an independent check) and contain only known State names (193 members and observers); otherwise the vote is discarded and counted. An adopted draft takes the targets of its resolution; a rejected or vetoed one takes those of the meeting (invited States, agenda item, mentions before the vote). When a text names more than three countries with similar weight, the target is left general rather than guessed.
Limits. Text extraction from PDF can split a word; the checks above catch most failures but not all. A sanction or a vote is an act, not a measure of influence. Votes of the General Assembly (alignment scores) are a separate, annual source.
A dedicated streaming worker maintains persistent connections to real-time data sources via WebSocket protocols. These streams provide sub-minute updates for phenomena where temporal resolution matters — aircraft movements, maritime traffic, lightning strikes, and radio spectrum anomalies.
Stream data is snapshot every 10 minutes into GeoJSON structures cached with 5-minute TTL, ensuring map layers always reflect near-real-time conditions without overwhelming storage or bandwidth.
A stream held by a single connection (AIS, whose free key allows one; lightning) is held by one process of the worker only. That process renews its lock and publishes the stream every minute from its own thread, whichever process the scheduled task lands on; a process that does not hold the lock opens nothing and closes any connector it still has. A connector that has not yet received a position publishes its health only: its empty state would declare the strategic passages uncovered.
Transits at the strategic passages. The transit volume of a passage counts the
merchant ships under way (at least 1 knot, neither moored nor at anchor) in a counting zone drawn
tight on the strait: cargo ships (AIS types 70–79) and tankers (80–89) only. Fishing, pleasure
craft, passenger ships (the ferries that cross the strait), service vessels and ships whose type is not
yet known (the type only arrives with the static data message) are left out and published apart. The type
of a ship never changes: each type received is kept in a lasting register (180 days), so that a restart of
the connector forgets none. A passage where fewer than 80% of the ships have a known type has no variation,
and such readings never enter a baseline: a register still filling up would otherwise pass for a rise of
merchant traffic. Its
variation compares the count with the readings of the past 1 to 7 days of the same definition. A passage
from which no position has arrived for 30 minutes is without coverage: absent, never a fall of traffic.
Military watch. Ships declaring the military type (35) are counted apart in the wider
zone around each passage, present and under way, with their names. The watch aims at the movements of
large fleets, which sail with their transponders on; ships running dark escape it by construction. It
enters no index. A zone that receives no position has no reading (neither zero nor “no fleet”).
Two watch zones outside the passages cover the bases of the Baltic Fleet (the approaches of Kaliningrad and
the Gulf of Finland): only military ships are read there, and no transit is counted.
| Stream | Protocol | Refresh | Coverage |
|---|---|---|---|
| Air traffic | REST (OpenSky Network) | 10 min | Global — military & civil |
| Maritime AIS | REST/WebSocket | 10 min | 12 strategic passages & chokepoints |
| Lightning | WebSocket (Blitzortung) | 10 min | Global citizen sensor network |
| Radio spectrum | HTTP + WebSocket (KiwiSDR) | 10 min FFT | 71 stations worldwide |
The Opportunistic Ionospheric Radar is an experimental passive intelligence pipeline operating in the HF band (2–30 MHz). Unlike conventional radar systems that emit their own signals, this method exploits existing radio emissions (time signals, beacons, broadcasting) reflected by the ionosphere as signals of opportunity. By analyzing how these signals propagate through the F2 layer, the system performs ionospheric sounding, emitter geolocation, and propagation anomaly detection — all without ever transmitting. The approach draws from classical bistatic radar and signals intelligence (SIGINT) techniques, adapted to open-source civilian infrastructure (publicly accessible KiwiSDR receivers distributed worldwide).
GPS badge).The Port Occupancy module measures how full the berths of 19 strategic ports are, from Sentinel-1 radar and Sentinel-2 optical imagery (Copernicus CDSE, Process API, physical values requested on each port’s box only). AIS transponder counts are published beside it as an independent witness.
Collection tasks are distributed across 6 specialized worker queues, each dimensioned for its workload profile. A deduplication lock mechanism prevents task accumulation when the processing queue backs up — if a task is already running, subsequent triggers are silently skipped rather than queued, avoiding the cascade failures that plague naïve job schedulers.
The platform serves its own map tiles from a 128 GB PMTiles archive (zoom levels 0–15, full global coverage) hosted locally on the server. This eliminates all dependency on external tile providers (CartoDB, Mapbox, OpenStreetMap CDNs) and ensures zero third-party tracking on the map interface.
Every ingested document — whether a Reuters dispatch, a Telegram post, or a FRED data annotation — passes through a multi-stage NLP pipeline that extracts structured intelligence from unstructured text. The pipeline is designed for multilingual robustness: the same analytical chain handles French, English, Russian, Chinese, and Arabic without language-specific branching.
Processing is split between real-time (inline with ingestion) and batch refinement (asynchronous Celery tasks), allowing the system to provide immediate first-pass results while progressively improving accuracy through evidence accumulation.
Sentiment scoring relies on XLM-RoBERTa, a transformer model pre-trained on 100+ languages. The model was fine-tuned on multilingual Twitter data for sentiment classification, making it particularly suited to the short, informal, and multilingual nature of social media and news headlines.
Raw model output is then adjusted for geopolitical context using domain-specific lexicons in 6 languages (French, English, Russian, Chinese, Arabic, Hindi). Terms like "ceasefire" or "sanctions lifted" carry geopolitical valence that generic sentiment models often miss. The adjustment follows a weighted blend:
A pre-trained transformer excels at linguistic structure but lacks domain common sense. An article titled “The Dachshund: France's favourite sausage-dog on display at the Moulin Bestiary” was once classified as strongly negative — the model interpreted “sausage” and “dog” in the same sentence as alarming. To address these semantic blind spots, a lightweight PyTorch correction model learns continuously from human-validated corrections.
The corrector is deliberately minimal — 4 input features, 2 hidden layers, ~3k parameters — because it doesn't need to understand language. It only needs to learn when RoBERTa is wrong, based on surface-level signals (score, confidence, text shape). Every 20 accumulated corrections trigger a retraining pass, and the updated model is persisted to disk. The correction only overrides the base model when its own confidence exceeds 0.7, preventing the student from overruling the teacher on uncertain cases.
Corrections are sourced from active user participation. In the article browser, each article displays its current sentiment classification alongside four buttons — Positive, Neutral+, Neutral−, Negative — allowing any reader to flag a misclassification in one click. This human-in-the-loop design treats every user as a potential analyst whose domain intuition enriches the model over time.
Entity extraction uses SpaCy with a large French model, augmented by a curated geopolitical entity database of countries, international organizations, and armed groups. The model is lazy-loaded on first NER request (not at application startup), keeping memory footprint low until analysis is actually needed.
Extracted entities feed two downstream consumers: a co-occurrence network that maps relationships between actors, locations, and organizations across articles, and the conflict detector that uses entity presence to boost or reduce conflict classification confidence.
| Entity type | Examples | Downstream use |
|---|---|---|
| GPE / LOC | Countries, cities, regions | Map geolocation, hotspot detection |
| ORG | NATO, OPEC, Wagner Group | Actor network, sanctions cross-ref |
| PERSON | Heads of state, military leaders | Politico-Meter, influence mapping |
| NORP | Nationalities, ethnic/political groups | Demographic signals, tensions |
The entity network tracks co-occurrences across the article corpus — when two entities repeatedly appear in the same articles, an edge forms between them. Network density (edges / possible edges) provides a proxy for geopolitical interconnectedness in the current news cycle.
Each document is scored against 14 geopolitical themes using a weighted keyword system. Keywords carry individual weights (0–5) reflecting their specificity and diagnostic value. A term like "VBIED" is far more indicative of armed conflict than "tensions", and the weighting reflects this.
The confidence formula combines two signals: keyword score (how strongly matching terms signal the theme) and match breadth (how many distinct terms were found). This prevents a single repeated keyword from inflating confidence while rewarding diverse evidence.
The system supports synonym expansion and contextual metadata (geopolitical vs. economic framing), allowing the same event to be classified under multiple themes with independent confidence scores. Theme accuracy is tracked over time via usage statistics, enabling progressive calibration.
Raw NLP scores are refined through a Bayesian evidence fusion framework inspired by intelligence analysis methodology. Rather than trusting a single model output, the system accumulates evidence from multiple independent signals and updates its confidence sequentially.
Beyond generic NLP, specialized analyzers run in parallel to extract domain-specific intelligence from the same document stream. These enrichment layers operate independently of the core sentiment/theme pipeline, using dedicated lexicons and scoring models.
Individual data points — a VIX spike, a conflict event, a shift in social sentiment — are noise until correlated. This section describes the engines that transform isolated signals into systemic intelligence: the OMEGA Engine (delta detection), its causal graph and resonance, the direction it serves for each asset, the Factor Z (media-social dissonance), weak signals (deviations from a reference) and the Causal Pattern Engine (precedents).
OMEGA is the system’s heartbeat. Every 15 minutes, a Celery task collects a snapshot of 32 live metrics — 13 financial instruments (via yFinance) and 19 OSINT counters (conflicts, cyber threats, RSS volume, social posts…) — and compares it to the previous snapshot. Significant variations are flagged, classified by severity, and broadcast to every downstream consumer.
omega_timeseries), by its distance
to their median. MODERATE beyond 95 % of them,
HIGH beyond 99 %, CRITICAL beyond
all of them (unprecedented over a week), STABLE otherwise; the market regime widens or tightens the
moderate and high tails, never the critical one. A metric that has not moved for 7 days gives a move
without precedent but of unmeasurable size, capped at MODERATE. Each change publishes its
exceedance (share of past changes at least as large) and basis. With fewer
than 48 reference changes the severity is unknown, never judged by a fallback threshold.
Counts move by steps and on very different scales: a fixed threshold or a robust z-score turns their
ordinary moves into alarms, while a rank judges each change against the metric’s own history.unavailable_metrics, with the reason
given by their collector in unavailable_reasons.
| Domain | Metrics | Signal Type |
|---|---|---|
| Financial | VIX, WTI, Brent, Gold, BTC, S&P 500, Euro Stoxx 50, EUR/USD, Silver, Copper, NatGas, USD Index, 10Y Treasury | % change |
| Conflict | Conflict events, fatalities, earthquakes M4+, countries advised against (level 3+) by at least two core issuers | Absolute |
| Cyber | CVE, APT, IOC threats | Absolute |
| Environment | Thermal/fire detections (FIRMS) | Absolute |
| OSINT | RSS articles, negative sentiment, social posts, OSoME propagations, RAG queries | Absolute |
| Infrastructure | Aircraft over Europe (OpenSky snapshot), online WebSDR receivers, active weak signals, healthy sources (keys of the common data catalogue fresher than their expected age; stale and absent ones are published in sources_health) | Absolute |
| Strategic | New tightening sanction measures over 7 days (OFAC, UK list, EU sanctions map, Canada SEMA, Federal Register, from relation_events), strategic flux alerts | Absolute |
risk_level is not a simple count of critical metrics.
n_critical and n_high count only the changes that move toward risk: a
metric of the risk set, with risk polarity, going up. A fall, or a neutral count (press volume, social posts,
SDR receivers), does not raise the level, and neutral counts do not vote on the global direction. Three
amplifiers then apply: Archiviste precedents that confirm the pattern, the resonance score
(§3.2) when several domains vibrate together, and lagged correlations that announce a propagation.
The precedents and the anticipations follow the same rule: only a move toward risk is confirmed by a
precedent or announces a follower, and only a positive correlation announces a follower moving the same way.
A sharp fall of the VIX, even with a historical precedent, does not raise the level:
omega:latest while it rebuilds its graph: the level is then
None with the status pending. The final level follows, with the status
final, in omega:risk_level:last. Every reader of the level — the Trade-Bot,
the Investo-Meter, the Pattern Engine, the homepage — reads the final level; a missing level is shown as
not available, never as STABLE.
regime_modulation).
A VIX spike alone is noise. A VIX spike coinciding with rising conflict events, a surge in negative RSS sentiment, and a carry trade unwind — that is resonance. The resonance scorer detects when signals from different analytical domains converge on the same nodes of the OMEGA causal graph.
causal_indirect links, up to 3 per chain): the 2022 invasion of Ukraine
reaches natural gas directly, then energy storage and renewables. Attachment takes place
once the resonance series nodes exist (Brent, natural gas, materials risk, Factor Z,
media sentiment): this is where OMEGA measures the correlations through which distinct
chains cross — Hormuz → Brent ↔ natural gas ← Ukraine 2022.
perime or absent in its hotspots field). Each country of the
index carries the date of its ACLED data: when the download fails, the connector serves its last downloaded
copy with that date, or the index carries over the previous value with its date; data older than 45 days,
or a carried-over value without a date, makes no hotspot (field outdated).
inputs field.
value (normalized 0–1)
representing its current “vibration” intensity.amp(t) = amp₀ × 2−Δt/half_life.
Persistent signals lose weight; fresh shocks dominate.latency_h hours). The weight of a conflict towards
financial nodes (0.75, latency 2 h, half-life 5 days) is applied only as measured by the event
study of conflict channels (32 interstate crises and 15 major wars, 2003–2026, protocol fixed before
any computation): only a conflict carries it, other event types not being measured; a conflict sends no
resonance to the channels the study does not support (volatility, hydrocarbons, safe havens and world
equities, with their neighbouring nodes); the dollar is published without use and keeps it, and an
unmeasured channel keeps it as unmeasured. Whether a major war moves volatility, defence stocks and the Swiss franc,
as the 15 wars suggest, is tested on shocks to come under a protocol fixed on 15 November 2026: openings
of hostilities involving a nuclear-armed state or a major hydrocarbon producer, attacks on energy
infrastructure or on the shipping of the main straits (the infrastructure counts, not the dead), and
measured entries into war, each read apart once ten shocks have their twenty sessions elapsed.hub_boost = 1 + log₂(structural_degree) × 0.2.amplitude = 1 / (1 + e−2(raw−1)).
A Spearman correlation boost (max +20%) is applied when strong cross-domain correlations exist.omega:directions); the Trade-Bot reads it
(§6.1, §6.3). The direction
combines the signed evidence of the system’s sources, on a fragility base (combination below):
The core hypothesis: when institutional narrative diverges significantly from public sentiment, informational dissonance is building — a measurable signal that traditional indicators miss. Whether this dissonance translates into observable social tension remains an open empirical question — the Factor Z measures the dissonance itself, not its downstream consequences. Populations react, they don’t anticipate. The Factor Z captures this asymmetry by measuring the gap between what media says and what people feel, with a configurable temporal lag (default: 6 hours).
| Tension Level | |Z| Range | Interpretation |
|---|---|---|
| CONSENSUS | < 1.0 | Background noise — media and popular sentiment converge |
| DIVERGENCE | 1.0 – 3.0 | Normal inter-platform gaps — monitor |
| DISSONANCE | 3.0 – 6.0 | Notable polarization — possible information bubble forming |
| CONCERNING | 6.0 – 9.5 | Narrative fracture — potential manipulation or emerging crisis |
| SEVERE | 9.5 – 12.5 | Deep and persistent social cleavage |
| CRITICAL | ≥ 12.5 | Major narrative rupture — armed conflict / mass disinformation |
| Component | Weight | Market input | Sub-score |
|---|---|---|---|
| Volatility | 25% | VIX (live; FRED previous close as fallback) | 90 at VIX ≤ 12, 5 at VIX ≥ 40, linear between: 90 − (VIX − 12) × 85/28 |
| Momentum | 25% | S&P 500 daily change (%) | 50 + 17.5 × change, bounded to 5–95 |
| Safe haven | 20% | Gold daily change (%) | 50 − 17.5 × change (gold up = fear), bounded to 5–95 |
| Breadth | 15% | Share of the 8 major indices that are up (at least 3 observed) | 5 + 90 × share |
| Dissonance | 15% | Mean |Z| of the sector Factor Z | 80 − 7 × mean |Z|, bounded to 5–90 |
Missing components are dropped and the remaining weights renormalised (the coverage is published). Classes: below 25 extreme fear, below 45 fear, below 55 neutral, below 75 greed, otherwise extreme greed.
In this system a weak signal is the quantified deviation of a measure from its own reference: dated, attached to a subject (a country, a chokepoint, a sector, an index), with its known limits. It is neither a forecast nor a risk score, and its predictive value has not been established. Fixed thresholds or raw counts, without a statistical reference or a kept history, are not shown as signals: the page rests on its own measurement foundation.
| Step | What is done |
|---|---|
| Archive | Series that expire (Redis histories of 90 points living 7 days, SQLite tables purged after 7 days) are copied every six hours into a PostgreSQL table partitioned by year, unchanged and with their original dates. Re-reading never duplicates a point; a point rewritten during the day keeps its last value. PostgreSQL is the system of record; Redis and SQLite are buffers. A daily dump is kept on the server (an off-site copy is still to be decided). |
| Recording | Every hour, 19 measures are reduced to one value per subject and day (UTC; the median of the day’s readings, or the last value). A missing or non-finite value is never recorded as zero. |
| Reference | The 28 days preceding the day read (the day itself is never part of it). Below 14 values the series is “reference being built” and no deviation is shown. |
| Deviation | z = 0.6745 · (x − median) / MAD, with a floor on the MAD (the measure’s resolution) so that an almost constant series does not produce giant values. Counts are taken as log(1+x). A median below a minimum level (e.g. 20 vessels a day for a chokepoint) gives no deviation. |
| Levels | To watch: |z| ≥ 2.5 on the last reading. Active: |z| ≥ 3.5 on at least 2 of the last 3 readings, in the same direction. Extreme: active and |z| ≥ 5. |
| State measures | The pace of State measures is a daily count of distinct acts (a text aimed at several countries is one act), for the sanctions layer and the trade layer, all sources of measures together. Only business days are counted (official publications do not happen at weekends): an act dated on a Saturday or a Sunday counts on the Monday. The series starts the day the last source of the layer was in place, because the level before that is that of the sources already present. A source that stops answering lowers the count: a day without acts and a day without data cannot be told apart. |
| Press measures | Two series per theme, read only from the articles analysed by RoBERTa (articles awaiting analysis, which carry a default neutral sentiment, and those analysed by the fallback model are left out): the share of the theme among the analysed articles, in %, and the tone S = share of positive articles minus share of negative articles (original classes, before the harmonisation by clusters of similar articles). A share is used rather than a count because press volume drops by about a third at weekends and fell by 28% in September. Themes above 10 analyses a day are measured daily, themes between 4 and 10 over a sliding window of 3 days, rarer themes not at all. Theme tagging started on 19 August 2026: the series start there, and a day on which tagging did not run is left empty, never filled. Themes are set by weighted keywords: a theme is not a subject, and the corpus is that of the configured feeds, not the world’s press. |
| Dated facts | The panel Relations and decisions lists dated facts read from the same database of State measures: UN Security Council decisions (every resolution, and only the votes that dissent: a vote against, an abstention of a permanent member, a rejection, a veto), corroborated bilateral shifts and changes of travel advisories. They are facts, not deviations from a reference: they carry no level and no z. A vote is published two to three weeks after the meeting, so it appears late. |
| Independence | Indices that recombine other measures (OMEGA, logistics stress, regime and cascade indices) are marked derived: they are displayed but never counted as a confirmation of another signal. Two signals from the same corpus (for instance two press-based measures) do not confirm each other. |
Measured, not assumed: false alarms. On pure Gaussian noise with a 28-day reference, about 3% of readings reach “to watch” and about 0.05% reach “active” (against 1.2% and 0.05% for a reference known exactly): a small reference has heavier tails, because its MAD is itself noisy. Skewed or heavy-tailed series produce more. With a few hundred series read daily, expect on the order of ten “to watch” per day by chance alone, which is why they are shown in grey and why “active” requires persistence. Thresholds are starting values, to be tuned on the observed rate after four weeks.
Limits. A deviation says that something changed relative to its own recent past, not why, nor whether it matters. A reference of 28 days ignores seasonality. Several inputs have known defects, listed on each signal card (fixed thresholds in the original index, dead components, a global Z Factor that is not a z-score, a sign lost upstream). Every detection is stored so that its usefulness can be evaluated later against what actually followed.
The Pattern Engine goes beyond detection: it asks “have we seen this configuration before, and what happened next?” It captures system-wide snapshots at critical moments, measures the real outcomes at their horizon, and builds templates from validated chains.
engine_version = 2). Those of the first engine are kept but labelled
“legacy v1”: their vectors are rebuilt from the raw columns that were really
measured, and they are excluded from templates and injections (the reasons are dated in the Methods Journal).
Templates are built from version-2 data only. Until a template rests on enough chains validated on correct
data, no concordance is injected into OMEGA, reports or LLM contexts.
provenance). A dimension whose source is absent is None, not
a value. The snapshot records its coverage: the weighted share of measured dimensions.
Dimensions 21–25 weigh 0.5 while they ramp up. The OMEGA level is read from
omega:risk_level:last, the final level also read by the Trade-Bot: OMEGA first publishes
omega:latest with a provisional level (None, status pending) while it
rebuilds its graph (§3.1).
| # | Feature | Range | Source |
|---|---|---|---|
| 1 | BED Score | 0 – 100 | Bond-Equity Divergence |
| 2–3 | ICR FR / US | 95 – 115 | Real Cost Index (absent when no component is measured). FR/DE weights: Eurostat household budget survey by household type (imputed rents counted, mutual health insurance in health); components: HICP sub-indices of the last complete month weighted by official HICP weights (the flash estimate of a month carries the total only; its sub-indices arrive about mid-month after); food also reads the FAO Food Price Index annual change. Effort rate: survey spending revalued by HICP, divided by EU-SILC income, by income quintile, as an interval. US effort rate: BLS Consumer Expenditure Survey (spending and after-tax income of the same sample, by quintile), each post revalued by its CPI-U index, income carried forward by CPI-U or by private hourly earnings (interval bounds) |
| 4 | Carry Score | 0 – 100 | Carry Trade (cpi.score) |
| 5 | Cascade Score | 0 – 100 | Cascade Index (cci.score) |
| 6 | SPY-TLT Correlation | −1 – +1 | 21-day correlation matrix |
| 7 | LSI Score | 0 – 100 | Logistics Stress |
| 8 | Breadth Stress (BSI) | 0 – 100 | 100 = extreme concentration |
| 9 | Refuge Score | 0 – 100 | Refuge Flow (rfi.score) |
| 10 | Concentration Score | 0 – 100 | Market Concentration |
| 11 | Fear & Greed | 0 – 100 | Market sentiment composite |
| 12 | Energy Risk | 0 – 100 | Mean ESI of the large importers (DE, FR, IT, JP, GB) |
| 13 | VIX | 10 – 80 | FRED VIXCLS |
| 14 | S&P Change % | −10 – +10 | Daily change (absent without previous close) |
| 15–16 | 10Y / 2Y Yield | 0 – 8 | FRED DGS10 / DGS2 |
| 17 | OMEGA Resonance | 0 – 1 | Amplitude from omega:resonance:score (OMEGA’s first, provisional write carries none) |
| 18 | OMEGA Critical Count | 0 – 32 | Delta severity |
| 19 | Gold trend gap | −15 – +15 % | GC=F vs its 200-session mean (price archive) |
| 20 | BTC trend gap | −50 – +50 % | BTC-USD vs its 200-session mean |
| 21 | Semiconductor trend gap | −30 – +30 % | SOXX vs its 200-session mean |
| 22 | Dry-bulk freight, 5 sessions | −20 – +20 % | BDRY ETF (dry bulk, not containers) |
| 23 | PER / Forward spread | 0 – 20 | Nasdaq valuation divergence |
| 24 | Gold / 10Y correlation | −1 – +1 | 30-day rolling |
| 25 | Sector rotation | 0 – 100 | XLP/XLV vs XLK, 5 days (no score if one ETF is missing) |
Dimensions 19–22 are gaps to trend read in the price archive, not price levels: price levels cannot be compared across eras, and a bounded level that leaves its range freezes its dimension at 0 or 1 — constant dimensions inflate every similarity.
causal_snapshots
with engine_version = 2.direction_accuracy: 1, 0 or None) and does not
enter confidence. Classification rules fire only on measured values.hit_rate): the share of outcomes with the
same sign as the median outcome.A cosine sees only the shape of a vector: a calm situation at 0.2 everywhere and a crisis at 0.8 everywhere would be identical (similarity 1.0). The masked distance sees the level (similarity 0.4). The injection thresholds were fixed before any version-2 data. Every consumer uses the same gate: OMEGA, reports, the DS agent, the remote context and the Investo-Meter.
On synthetic random-walk markets the skill is about −0.05 at every horizon: that is the cost of estimating from 20 analogues only. A real skill must therefore beat zero despite this handicap. Nothing reads the analogues: OMEGA, reports, LLM contexts and the Trade-Bot do not see them, and a test checks it. The evaluation is the evidence (or its absence) on which a decision to use them will be taken. Evaluation (measured on 17 October 2026): on 2007–2026 the Brier skill is −0.004 at 7 days, −0.019 at 14, −0.025 at 30 and −0.076 at 90 (78 independent outcomes): no skill shown. The analogues are therefore descriptive context only, shown in the Quant-Lab (regime tab) with their evaluation, and outside every calculation. Known limit: adjusted closes are revised retroactively by later dividends, which slightly rescales past prices (returns and trend gaps are barely affected).
| Pattern Type | Snapshot Signature | Outcome |
|---|---|---|
| Valuation Divergence | PER/Forward spread > 8 | S&P −3% or more |
| Semiconductor Rotation | Sector rotation > 60 | S&P −1% or more |
| Chokepoint Disruption | Dry-bulk freight +10% in 5 sessions + Energy risk > 60 | — |
| Infrastructure Break | Cascade > 50 + Energy risk > 50 + OMEGA HIGH/CRITICAL | — |
| Gold/Rate Decorrelation | Gold/10Y corr > 0.3 | Gold +1% or more |
| Flight to Safety | Refuge > 50 | Gold +1% or more |
| Bond Stress Correction | BED > 55 | S&P −2% or more |
| Geopolitical Escalation | OMEGA CRITICAL + Cascade > 40 | — |
| Carry Unwind | Carry > 50 | VIX +5 pts or more |
| Liquidity Squeeze | LSI > 55 + BSI > 65 (concentrated market) | — |
| Contagion Spread | SPY-TLT corr > 0.6 + Cascade > 40 | — |
| Risk Appetite Return | F&G > 70 | S&P +2% or more |
| Generic Stress | Unclassifiable | Various |
These engines are not independent — they form a closed loop of increasing analytical depth:
Raw data and correlation scores are useless without interpretation. This section describes how the platform transforms structured signals into human-readable intelligence reports — combining Retrieval-Augmented Generation (RAG), multi-provider LLM orchestration, and Dempster-Shafer evidence fusion to produce analysis that acknowledges what it knows, what it doesn’t, and what might be.
The Archiviste V5 is not a simple document store. It is the memory and reference library of the intelligence engine. It ingests academic papers, central bank reports, and historical analyses, breaking them into indexed sections with extracted keywords, crisis types, asset classes, and Boyer mechanisms.
Four automated connectors continuously enrich the library: NBER (daily working papers), IMF (weekly Excel reports), Hérodote (Wednesday historical articles), and Persée (Friday academic archives). Each query to the Archiviste enriches its resilient database — the more it is used, the more knowledgeable it becomes.
Hand-deposited documents carry their provenance. Documents the server cannot reach
(for example a summit archive served only to visitors from some countries) are collected by hand and
deposited under a sub-folder of SOURCES/ with a PROVENANCE.json file: source, URL,
collection date and method; the category is the name of the folder holding the document unless the file
states it. Such a document is identified by its path within SOURCES/, so two files of the same
name in two categories remain two documents, and its SHA-256 fingerprint is recorded: a file identical to
one already deposited is reported as a duplicate and not ingested twice. One sub-folder can be ingested on
its own. Documents without a provenance file keep their file-name identifier.
One search, two stages. A query first reads the system’s own bases: press and social posts (whole word, accent- and case-insensitive, within the chosen news window), library documents by their text, causal chains by cause and effect (French and English equivalents included), and the semantic index. Semantic neighbours are kept only if they contain a query term or, for queries of more than two terms, reach a cosine floor of 0.5 (C): a rare name such as “Delos” has no reliable semantic neighbours. The external archives then answer source by source and are appended as they arrive; each runs in its own thread with its own time limit (D), so a slow source never holds the others.
Periods follow secular cycles. Periods are the secular cycles of Turchin & Nefedov, Secular Cycles (Princeton, 2009) — Rome, England, France, Russia, with their expansion, stagflation, crisis and depression phases — and, for the United States, the integrative and disintegrative phases of Turchin, Ages of Discord (2016). China follows its dynastic cycle, from the Xia (c. 2200 BCE) to the Qing (1912). Each source is queried only over the years it actually covers (Chronicling America 1770–1963, NBER from 1973…). Where a source dates the subject (Wikidata, newspapers), the period filters its results; where it dates the publication (Persée, Hérodote, the web), a recent study of 1315 remains relevant and is not discarded.
Every search is remembered. Query, period, the measured answer of each source (results, empty, time-out, error, duration) and the results themselves are stored. A repeated search shows its earlier results at once; the “sources” panel reports what each source actually returned rather than a configuration flag; the retained and discarded semantic scores are the data on which the 0.5 floor will be calibrated (C → B).
Before asking an LLM to analyze a situation, the system first retrieves relevant context from its own databases. This is not generic web search — it’s domain-specific semantic retrieval across 13 indexed knowledge domains, powered by BGE-M3 embeddings and FAISS vector search.
Three domains carry the system’s own work. mesures holds dated notes written from the
computed data: fertiliser and gas prices, the Russian and Belarusian share of fertiliser imports by bloc
(annual intervals, monthly measures, estimated months), EU acts, dated measures and rendezvous, Access2Markets
announcements and OMEGA directions. actes holds the official text of each EU act read from
Cellar, one document per article. documentation holds this methodology, the data audit and
the OMEGA dossier, one document per section. Every note opens with its class in brackets, in the text itself
since the language model reads the text, not the metadata: measured (source and date),
estimated (method and bounds), read automatically, not reviewed, official text,
official announcement, documentation. An absent value makes no note. Reports keep that class
when they use a figure: an estimate is written with its bounds, an unreviewed extraction is said to be
unreviewed.
History is kept: each chunk carries a fingerprint of its source and text, and a chunk already indexed is
never added again, while a new value is added with its date; the Sunday rebuild does not empty the domains
whose history lives only in the index (mesures, security, intelligence). The same fingerprints keep every other domain
free of duplicates, and existing indexes are cleaned once without recomputing vectors. Deposited PDFs carry
the collection date and the organisation of their provenance file.
An index is shared by every process that reads or writes it (the indexing worker, the web application, a
maintenance shell). A write takes a lock on its domain: a routine pass leaves a domain being written elsewhere
to its next cycle, and a rebuild waits for it. Files are replaced whole, never written in place, and a process
rereads an index rewritten elsewhere before it writes to it or searches it. The indexer's cursors are merged
with those on disk, so that one process never restores another's older values.
| Domain | Source | Max Docs | Reindex |
|---|---|---|---|
| rss | RSS articles (PostgreSQL) | 50,000 | Incremental: one domain every 2 hours in rotation; complete rebuild on Sunday at 04:00, one domain per task, smallest first |
| social | Social posts (multi-platform) | 20,000 | |
| economic | Latest market quotes | 10,000 | |
| conflict | Conflict articles | 5,000 | |
| intelligence | DS_Agent brief, Investo insight, Factor Z, energy and weather signals | 350 | |
| security | Sanctions, corruption, GDACS, ICC summaries | 500 | |
| archive | Archived articles and Archiviste documents (AI reports excluded) | 2,000 | |
| relations | State measures and bilateral events | 5,000 | |
| mesures | Dated, classed notes from computed data | 5,000 | |
| actes | EU acts, official text by article | 5,000 | |
| documentation | Methodology, data audit, OMEGA dossier | 5,000 | |
| demo | OECD, Eurostat, World Bank | 5,000 | Complete rebuild on Sunday only |
| sdr | SDR/HF radio detections | 1,000 |
Reports are generated asynchronously via Celery (queue: llm, concurrency=1)
using a multi-provider LLM architecture with automatic failover. The system never
depends on a single provider — it maintains a priority chain of 6 backends
and dynamically discovers free-tier models at startup.
| Use | Chain, in order | Daily limit of the first link |
|---|---|---|
| AI reports, Trade-Bot | Gemini 3.5 flash-lite → Gemini 3.1 flash-lite → OpenRouter (free models) → Ollama Cloud (gpt-oss:120b) → Cline | 500 requests |
| Frequent small calls (briefs, event validation, Oracle, editorial notes, digest) | Gemma 4 26B → Gemini 3.1 flash-lite (capped at 150 calls a day) → OpenRouter → Ollama Cloud → local model | 14,400 requests |
| Translations | Gemini 3.1 flash-lite → Gemini 3.5 flash-lite → OpenRouter | 500 requests |
| Reserve | Anthropic / OpenAI / DeepSeek (API key optional) | Last resort |
| Report Type | Focus | Sections |
|---|---|---|
| Geopolitique | Power dynamics, tension zones, scenarios | Executive summary, structural dynamics, tension zones, power relations, scenarios, weak signals, watch points |
| Economique | Macro indicators, systemic risks | Economic synthesis, indicators, risks, regional comparison, macro scenarios, weak signals, watch points |
| Securite | Threat assessment, hybrid/cyber | Threat assessment, conflict trajectory, non-state actors, hybrid/cyber threats, crisis scenarios, early warning, recommendations |
| Synthese | Cross-domain strategic dashboard | Strategic dashboard crossing all 3 dimensions, cross-domain feedback loops, integrated global scenarios |
| Eco-Strategique | Markets + supply chains | Markets & strategy, financial indicators, critical commodities, trade flows, strategic-economic scenarios |
| Geo-Securite | Geostrategy + security crossover | Geo-strategic analysis + security with cyber threats section |
<think>...</think>) are removed.[CAUSAL: cause → effect (confidence)] patterns are parsed into JSON-LD and cached in Redis.At report generation time, a market snapshot is collected from Redis — a point-in-time capture of every quantitative signal the system tracks. This snapshot is formatted into a text block and injected directly into the LLM prompt, giving the model access to live data it could never have in its training set.
| Category | Data | Source |
|---|---|---|
| Market Sentiment | Fear & Greed (market + crypto), VIX | FRED, CNN, Alternative.me |
| Equity Indices | S&P 500, CAC 40, DAX, Nikkei… | yFinance |
| Commodities | Oil, Gold, Metals, Agriculture | yFinance, AKShare |
| Quant Lab Signals | Carry Trade, Cascade, Correlations, Regime, LSI | Internal computation |
| Game Theory | Ormuz, Red Sea, Taiwan, Kashmir — DS confidence + Bel(escalation) | Dempster-Shafer fusion |
| Macro | IMF WEO (GDP, inflation), BIS credit gap, REER, property prices, BOJ TANKAN | IMF SDMX, BIS API, BOJ |
| Strategic | Supply risk scores per material:region, Factor Z | Flux Stratégiques |
| Energy | ESI per country, energy mix G7+Russia, France real-time MW | OWID, ODRE |
| Sovereignty | Maslow pyramid (5 levels: survival to influence) | Composite internal |
Classical probability forces a choice: either something is true or it isn’t. But intelligence analysis operates in a world of partial evidence and irreducible uncertainty. The platform uses Dempster-Shafer theory of evidence to model what the system believes (Belief), what it cannot exclude (Plausibility), and the gap between the two — the epistemic ignorance.
Analysis is not a one-shot event. The platform runs a continuous monitoring loop: the Oracle generates forward-looking scenarios, DS_Agent watches for emergent patterns, Game Theory models strategic actors’ decisions, and the alert system ensures critical signals reach the analyst in time.
The Oracle is a 3-phase pipeline that transforms domain signals into probabilistic scenarios with 1/4/12-week horizons. It combines signal extraction from the system’s own sources, Dempster-Shafer fusion with Shafer discounting (§4.5) over five domains, Game Theory among them, and LLM-assisted scenario generation. A dedicated 6-panel dashboard exposes the full epistemic state: tension trajectory, domain contributions, epistemic envelope, strategic scenarios, signal provenance with conflict K, and validation history.
A social post is stored at 0.0 / neutral / confidence 0 until its sentiment is analysed: such a post counts in the volume, not in the mean (same rule as the Investo-Meter). The weights of each domain are the author’s, set before any result (1 November 2026). Every row of the signal history carries its method; rows written before this method are kept as they were and are never mixed with the current ones (§5.1, auto-correction loop).
| Tension Score | Label | Scenario Calibration (static fallback) |
|---|---|---|
| ≥ 0.75 | Critique | popt=0.15 · pneu=0.35 · ppes=0.50 |
| ≥ 0.55 | Élevée | popt=0.22 · pneu=0.45 · ppes=0.33 |
| ≥ 0.40 | Modérée | popt=0.30 · pneu=0.50 · ppes=0.20 |
| ≥ 0.25 | Basse | popt=0.45 · pneu=0.40 · ppes=0.15 |
| < 0.25 | Très basse | popt=0.50 · pneu=0.35 · ppes=0.15 |
correct_tension() mechanism
adjusts the Dempster-Shafer tension based on observed hit rates per domain:
oracle:adjusted,
consumed by downstream systems (OMEGA graph, Trade-Bot, reports, Oracle dashboard).| Notion | Definition | Can increase while… |
|---|---|---|
| Confidence (Bel) | Support for a specific hypothesis, grounded in corroborated evidence. | …ignorance remains constant (new evidence confirms what was already believed). |
| Conflict (K) | Measured incompatibility between families of evidence during DS combination. | …confidence also increases (dominant view strengthens, but dissenting sources persist). This configuration is more concerning, not less. |
| Residual Ignorance (m(Θ)) | The share of hypothesis space that current data cannot discriminate. | …confidence and conflict both decrease (quiet period with insufficient signal). |
This independence is deliberate. A configuration where Bel↑ and K↑ simultaneously — strong dominant belief but rising contradiction — may be an early indicator of structural transition, the kind of signal OMEGA is designed to detect before it becomes visible in conventional indicators.
DS_Agent is a persistent background process that runs a 7-phase analytical cycle every 10 minutes. It acts as the system’s “always-on analyst” — collecting OMEGA data, querying the RAG pipeline, running cross-domain correlations, detecting emergent patterns, and producing bilingual (FR/EN) situational briefs.
omega:ds_agent_report (TTL 2h).
Real-world geopolitical actors are not passive data points — they are strategic agents making calculated decisions. The Game Theory module applies three layers of mathematical reasoning, each answering a distinct epistemic question:
A_shifted = A + shift (shift ≥ 0 to ensure positivity);
game value = 1/Σx − shift.Aα = α × Ahigh + (1−α) × Alow.
Solves three minimax games (on Alow, Ahigh, Aα)
and returns the robust strategy with uncertainty range on game value.| Scenario | Row Actor | Col Actor | Intelligence Sources |
|---|---|---|---|
| Strait of Hormuz | Iran | US Coalition | AIS maritime, RSS, UN votes |
| Red Sea / Houthis | Houthis | US Coalition | AIS, conflict events, RSS |
| Taiwan Strait | China | US-Taiwan | AIS, air traffic, RSS |
| Kashmir | Pakistan | India | Conflict events, RSS, sanctions |
Provenance. Each scenario input is measured, carried (older than its expected freshness, with its age) or absent (data catalogue, §1.10). An absent input produces no evidence: the ignorance rises, never a disguised calm. Inputs with no source in the system (nuclear posture, Iranian support to the Houthis, Suez revenue) are listed as unknown and submitted to the administrator instead of being invented. Each summary publishes its coverage: the share of its evidence sources actually observed (measured 1, carried 0.5). ACLED incidents only count when positive (the list holds at most 100 events worldwide over 7 days), so they stay out of the coverage. The Trade-Bot’s v2 engine weakens each scenario by (1 − ignorance) × coverage, and what is not observed counts in the ignorance that reduces position sizes (§6.6).
Triggers fed by existing sources. State measures against and by each actor (relations module, §1.5) for all scenarios, and for Hormuz’s diplomatic tension; armed incidents (ACLED/UCDP) in the Gulf, on Yemen’s coast and in Kashmir, US strikes in Yemen; TSMC falling 3 points or more below the semiconductor index; tech decoupling with the same thresholds as the China triple leverage; and a multilingual press detector (English, French, Chinese, Russian) that requires both the place and the event in the same article: PLA exercises, arms sales to Taiwan, Kashmir, the India–China border, India–Pakistan mediation. It does not feed the logistics stress index. The press signal of a passage is read under the name the RSS bridge publishes (“Ormuz” for Hormuz).
The Investo-Meter’s sector Factor Z (a sentiment gap rescaled by 20, not a standard deviation — see §3.4) has a single reader, used by every module, including the media-dissonance evidence of the scenarios.
Naval posture. In three sea zones (Persian Gulf and Arabian Sea; Red Sea and Gulf of Aden; South China Sea, Taiwan Strait and Philippine Sea), the press detector looks for two themes: a US aircraft carrier (carrier names, carrier strike groups, in four languages), which feeds the carrier modifier of the Hormuz and Taiwan scenarios, and a naval build-up of any navy (destroyers, frigates, submarines, amphibious groups, minesweepers), weaker evidence of escalation. Two articles in seven days attest a theme. The SDR corroborates without ever proving: each 10-minute scan keeps, per zone, the days on which a military, marine or aviation anomaly was triangulated there within 1,500 km; three days out of seven corroborate the press and raise the reliability of the evidence.
Container freight. No free and reliable container rate exists (the key named « harpex » holds a dry-bulk ETF), so the index is estimated from the software’s own data, in standard deviations from each component’s normal: AIS rerouting (share of cargo ships rounding the Cape of Good Hope rather than entering Bab-el-Mandeb, weight 0.40), container lines against the world market over five sessions (ZIM, Maersk, Hapag-Lloyd, COSCO Shipping, Evergreen, 0.35), and port congestion (Durban and Tanger Med against Djibouti and Piraeus, 0.25, a reduced weight until the port tracking audit). The index publishes its coverage; the Red Sea scenario weighs it accordingly.
Observed, not yet used. Marine insurers and reinsurers (Beazley, Hiscox, Lancashire, Munich Re, Swiss Re, Hannover Re) and fuel distributors (World Kinect, Rubis, Sunoco, Vopak) are archived and published with the index, and the Trade-Bot archives them at every scan, until the validation loop tells the sign and lead of their signal (a war-risk premium can serve an insurer or cost it). Per zone, persistent radio activity together with ports off their normal gives a weak hint of possible disruption, counted only when positive. The index also appears next to the logistics stress index in the Investo-Meter’s supply chain view.
The game theory page opens on the armed conflicts in progress, one tab per conflict, from two UCDP sources. The UCDP/PRIO Armed Conflict Dataset (yearly, consolidated) gives the sides, the secondary parties (states that send troops in support of a side), the type of conflict and its intensity in its latest year. The UCDP Candidate Events (monthly, preliminary) give the current year: each monthly file covers its own month (version 26.0.9: 1,984 of its 2,014 events in September 2026, read on 6 November 2026), so every month of the year is read and an event present in two files counts once. The page shows the year’s deaths (the UCDP “best” estimate; an event without an estimate is counted apart, never as zero), the number of events the UCDP flags for checking, and the events by country.
A state is a party only where the UCDP names its government (“Government of X”, X in the UCDP vocabulary, matched to its Gleditsch-Ward code); an armed group is never attached to a country by its name. A state is at war outside its territory when most of the year’s located events take place outside it (the share is published: Russia 96 % in Ukraine, Ukraine below half in Russia), or, without events of the year, when it is not a country of the conflict’s location. A secondary party is judged on the dataset’s year, never on another year’s events. The country of an interstate conflict’s location without events of the year establishes nothing.
Each state party carries the commitments of its alliances on the day (table of blocs and dated memberships): mutual defence (NATO art. 5 bounded by art. 6, EU art. 42.7, CSTO, Russia–North Korea, the US treaties with Japan, the Philippines, South Korea and ANZUS, the Alliance of Sahel States, the GCC), joint force (Peninsula Shield) and cooperation without clause (NORDEFCO). The Baghdad intelligence centre, an extinct coalition without an established end date, stays in the reference table and is not displayed. An alliance is not a belligerence: a commitment is shown as invoked only by a dated act. Conflict, country and party names are given in French, countries under the names of the French foreign ministry’s travel advice; a term outside the translated list stays as the UCDP writes it. A conflict gets a tab when it reaches war intensity (UCDP or 1,000 deaths in the current year), is interstate or internationalised, or has a state at war outside its territory; tabs are sorted by the year’s deaths, fifteen at most (configurable). A conflict the UCDP has not yet coded (“XXX475”) has no tab: no belligerent is identified. Refreshed daily; a failing source says so.
The table is published for reading: it feeds no scenario, no Oracle input and no Trade-Bot input.
Coalitions of the scenarios. Each scenario publishes, next to its game, the coalition of its two players: the support of the third parties it names (expert weights above 0.2, doubled for a nuclear power, each with what it rests on; a pair without a known link stays absent), nuclear backing, and the dated alliance commitments of each player, NATO and the EU at their members of the day. The coalition does not enter the payoffs: added uniformly to every strategy, it would change no equilibrium, only the value of the game and the robust interval read by the Trade-Bot.
The alert system translates analytical signals into actionable notifications across multiple delivery channels (dashboard, email, Telegram). It operates at two levels: system-level (OMEGA-driven) and user-configured.
AlertEvaluator classifies each OMEGA sweep result into one of three tiers:
| Tier | Cooldown | Max/h | Trigger |
|---|---|---|---|
| FLASH | 5 min | 6 | Nuclear/CBRN keywords, OR ≥2 critical + ESCALATING |
| PRIORITY | 30 min | 4 | ≥1 critical + ESCALATING, OR (critical + high ≥ 2) + ESCALATING |
| ROUTINE | 60 min | 2 | ≥1 critical, OR ≥3 high |
[STRONG RESONANCE] flag.
/brief, /status,
/sweep, /alerts, /signals, /portfolio,
/mute, /unmute, /help) with long-poll listener
on a background daemon thread.
The Investo Meter translates the entire analytical pipeline into a single question: “What is the geopolitical exposure of this asset right now?” It computes per-country risk scores, maps them onto a sector-exposure heatmap, and derives asset-level GeoScores.
| Component | Weight | Source |
|---|---|---|
| Security (corruption + sanctions) | 40% | Mean of two scores: corruption (World Bank CC.EST, or last cached WGI score) and sanctions — 20 points per authority maintaining a sanctions regime against the country, among US OFAC, EU, UK, UN Security Council and Swiss SECO (reference table checked 10 Oct 2026). Counted only when the corruption score is available. |
| Macro (credit gap + REER + GDP + inflation) | 25% | BIS + IMF WEO |
| Conflicts | 10% | Conflict index (below): ACLED political violence and demonstrations, UCDP/PRIO Armed Conflict Dataset, World Bank WGI Voice and Accountability. RSS hotspots (ConflictDetector) only as a fallback. |
| Energy Security Index | 10% | OWID + ODRE |
| Stress Index | 5% | Internal composite |
| Travel advisory | 5% | Consensus of 5 core government advisories — second-highest level, not the maximum (1→10, 2→35, 3→65, 4→95) |
| Fear & Greed (inverted) | 5% | CNN/Alternative.me |
/investo-meter/api/risk-formula serves the rules and every country’s
effective weights. Any change to the components or the weights is a new version. A score without any
country-specific component is a static fallback, flagged as an estimate. The base heatmap is
structural: hand-written reference values, weighted by the measured country risk when it
exists, and never a measure in itself.
The Ticker Registry is the single source of truth for all tracked financial instruments. Instead of hardcoded ticker lists, assets are registered dynamically via a search-and-register UI backed by a Redis-based registry with automatic classification.
The Maslow module models national sovereignty as a 5-level hierarchy of needs, from physical survival to geopolitical influence. Each level is computed from live data sources and aggregated into a composite sovereignty score per country.
| Level | Need | Indicators |
|---|---|---|
| 1 | Survival | Food security (FAO), water access, conflict exposure |
| 2 | Security | Military spending, sanctions burden, corruption index |
| 3 | Economy | GDP growth, credit gap (BIS), inflation, unemployment |
| 4 | Autonomy | Energy independence (OWID/ODRE), strategic dependencies, tech sovereignty |
| 5 | Influence | UN voting alignment, alliance networks, soft power indicators |
UN Comtrade figures distinguish three cases: a measured value; a flow with no declared line for a year the country did report, read as zero and flagged as such; and a failure (quota, key, network), which leaves the value absent instead of turning it into zero. A regional total names the countries it is missing and never counts them as zero; a country unknown to the official Comtrade reporters table is refused instead of being sent to the API. Each call goes through the shared quota governor, and answers are kept for seven days (cache and database), so that most days cost no call at all.
Trade alerts cover ten priority countries and four strategic products (rare earths, lithium, integrated circuits, crude oil), imports and exports. One request per country brings every year at once. The annual change is ct = ln(vt/vt−1), computed only between two consecutive measured years: a missing year is never filled with zero. The latest change is compared with the series’ own history, z = (c − mean) / standard deviation (floor 0.05): |z| ≥ 2 medium, ≥ 2.5 high, ≥ 3 critical, thresholds set before any result was seen. With fewer than five past changes the threshold is not calibrated: the alert then needs +50 % or −33 %, stays medium, and says so. The page shows how many series were measured, uncalibrated, late or absent; when none was measured it says detection is unavailable, never that trade is normal.
The Trade-Bot is the operational terminus of the intelligence pipeline. It consumes the output of every upstream system — OMEGA deltas, Oracle scenarios, Archiviste causal chains, Investo-Meter GeoScores, Flux Stratégiques supply risk — and synthesizes them into event-driven trading decisions. It does not predict prices. It detects regime changes and acts on causal mechanisms.
| Layer | Horizon | Signal Sources | Decision Weight |
|---|---|---|---|
| Tactical | 1–5 days | OMEGA critical deltas, Factor Z spikes, flash alerts, AIS anomalies | 30% |
| Strategic | 1–4 weeks | Oracle DS tension, Quant-Lab TA, causal chain triggers, Politico-Meter shifts | 40% |
| Structural | 1–6 months | Archiviste Boyer mechanisms, Maslow pyramid, real rates regime, strategic dependencies | 30% |
Each asset’s geopolitical score blends seven sub-sources: OMEGA, conflicts, sanctions, press, space weather, corruption and societal stress. A missing input takes its last known value while that value is no older than its maximum age, and that age is archived with the scan, as for prices: OMEGA risk level and resonance, carried separately, 2 h; space weather, 6 h (two Kp readings); sanctions, corruption, conflicts and societal stress, 24 h. The press is never carried: no article in three days is information, not a missing measure. Without this rule the weights are renormalised whenever an input misses a scan: on the archive of 10–15 October 2026 (207 scans, 90 assets) the score jumped by 5 points or more at 37% of scans (space weather present 28% of the time, one sanctions outage of 2 h 54, OMEGA’s two keys averaged even when one was absent); with it, about 25%.
A real change of value is never smoothed. The switching of OMEGA’s risk level from one cycle to the next is OMEGA’s own reading of the inertia of its deduced causal chains: it is kept as it is, and its two inputs are archived separately to be studied. The corruption layer reads the Investo-Meter’s corruption risk (World Bank WGI) by ISO code, the same value as in the published country risk formula (§5.5).
Every price timestamp is brought back to UTC; a timestamp without a time zone is the server’s local time and is converted as such, correctly across daylight-saving changes. A price is judged by its session (the date of the last trading day it comes from), not by how recently it was read: it is stale when more than one weekday is missing between that session and today (a holiday or the hours before the open are tolerated), or more than one calendar day for crypto-assets, which trade every day. Without a known session, the age of the reading decides: 26 hours on a weekday, 74 hours from Saturday to Monday, since Friday’s close remains the current price over the weekend.
A stale price never opens a position, never closes one, never serves as the reference of a decision and never grades it: the decision is recorded without a price rather than measured against a fiction. A position without a fresh price keeps its last mark, flagged on the dashboard. When more than half of the open positions are stale, the value of the book is neither recomputed nor added to the equity curve: a curve kept flat for lack of prices would read as stability.
A bet decided while its market is closed (night, weekend, after the close, or a weekday without a quote of the day, such as a holiday) is not filled at the last known price: it waits, and is filled at the opening price of the next quoted session, read from the price archive or, failing that, from the daily bar of the market data feed. Each asset’s session follows its exchange (time zone, hours, trading days: Riyadh trades Sunday to Thursday); CME futures and currencies trade almost continuously, crypto-assets continuously. Size and risk checks are computed again at the fill, on the book as it stands at the open; an opposite decision taken in the meantime replaces the waiting order, and an order without a readable opening price three days after its session expires, with its reason. An asset whose exchange is not known is filled at once, and the order says so. Both engines follow this rule, and each scan archives it with its observations.
The Trade-Bot is the financial role of OMEGA. OMEGA’s directional assessment of each asset weighs 25% of the v2 conviction, the engine’s hand-written tables the other 75%. The engine reads OMEGA’s intelligence direction only — causal chains, events, supply chains (§3.3) — without market orientation, rates shock or the precious-metals cascade: market orientation reads market stress, which the tables already read (risk regime), and would be counted twice. Measured on 15–21 October 2026, when the full direction was read, a single refuge-flight alert kept the orientation at +0.8: every risk asset received a full-strength “down”, −0.25 on each conviction, and 68 of 81 v2 decisions were SELL. The full direction stays archived and is graded separately in the report card (§6.8). The VIX, listed as an equity index, rises with stress in both OMEGA and the tables.
An assessment without a net direction counts as zero: OMEGA has looked and sees none. When OMEGA is missing, stale (older than 40 minutes) or silent on the asset, the conviction rests on the tables alone and the evaluation is flagged degraded. There is no veto either way: OMEGA opposing a net conviction makes the situation a conflict for the LLM arbiter, like an opposing layer; its absence blocks nothing.
Each asset’s inertia layer (fast, medium, slow) is measured on the asset’s own daily series archived since 2025, rather than declared. Its persistence profile is the correlation between the return of the past h sessions and that of the next h sessions, for h from 1 to 60 (time-series momentum, Moskowitz, Ooi & Pedersen, 2012). This profile is informed by the asset’s 2015–2024 history, whose weight is capped at that of the 2025–2026 base, and by its sector’s median. The own horizon h* is the strongest significant persistence (t ≥ 2 on the effective number of independent pairs). An asset without a significant horizon keeps its declared layer. With about 200 sessions of base, few assets reach that threshold, so momentum keeps the best-established positive persistence (the highest t), weighted by a confidence min(1, t/2); a significant horizon keeps full confidence; an asset without any positive persistence (mean reversion) has no momentum. A manual pin overrides the measure, and every pin is logged.
Changes of speed are the core of the engine. Four causes can make an asset one layer faster, and a 20-session volatility of 2.5 times its base can make it two:
A cause that stops keeps its effect for one half-life (h*). A cause lasting more than three half-lives is flagged as durable, a regime change.
Press of the exposure countries. An asset’s own press only holds the articles that name it; a national event is told through its actors, not through the companies it moves. Each new article is therefore counted once, in the hour and the day it enters the base, for every country its title names (country, city, US state, sub-national place or demonym, through OMEGA’s zone table, ambiguous names settled as in §5.1), or else for the place of action of its summary. A strait, a sea or a region names no country, and the outlet never decides the country: one press group runs editions for several countries. A surge is at least 20 articles over the last 24 hours and a share of the articles read at least three times the country’s own median daily share over the 28 previous days, with at least 14 counted days. A share is used rather than a count because the number of articles read varies sixfold from one day to the next (3,266 to 19,002 over the 28 days counted on 28 October): a general influx raises every country together, and is no surge. An hour or a day in which no article was counted stays unknown, never zero; a window is published with its number of counted hours, and below 12 counted hours it decides nothing. A surge gives no direction. Both engines read the same reading at each scan: it is one acceleration cause in the v2 engine, and in the main engine it makes the declared layer one step faster (§6.1). The reading of every scan, per country, is archived with the scan.
The financial system is read as an organisation (Mintzberg, contingency school). Its structure fits its environment without being determined by it; its regulation rests on the coherence of its sub-systems. Supranational market rules make interaction flows run mostly from the system to the groups, except at the end of a Schumpeterian cycle or after an anthropic break. Seven families are transposed:
Four measures are published:
Adhocracy, power and narratives have no daily series. Their exposure is declared and flagged as such until the archive of observations allows it to be measured.
MetaAnalyst is consulted only when the layers genuinely contradict
each other. It receives the table of signed contributions, the strongest argument on
each side and semantic precedents retrieved from the FAISS2 indexes (non-blocking, bounded delay),
but not the open positions. Its share of decisions is monitored against a target of
25% at most: a higher share would mean that the system’s signals are
either unclear or poorly exploited. The remote model is called first; a small local model is the
last resort. Every answer is validated (format, values, and a buy or sell must be supported by at
least one layer pointing that way); an invalid answer is rejected and the decision stays prudent.
The arbiter judges whether the opposed layer prevails at the asset’s horizon; it never reverses
the engine: the only bet it may confirm is the direction of the net conviction, beyond the action
threshold. Entries and exits thus obey the same conviction: a bet opened against it, or far below it,
would be closed within minutes by the signal-reversal exit.
TRADEBOT_CAPITAL
(100,000 by default) and stored by the book when it is created, so that returns stay right if the setting
later changes.The Trade-Bot evaluates its own discoveries (each with its causal chain), its seed universe, and the assets registered by the super-administrator. A ticker added from the Investo-Meter search — by any visitor, signed in or not — first spends seven days in quarantine, invisible to the bot. It then becomes a candidate if it has a price and a classification (sector, asset class, inertia layer), and enters the bot’s universe only when the super-administrator validates it; each validation or rejection is logged. A rejected candidate is still followed by the Investo-Meter. When one of the bot’s own discovery strategies proposes a ticker first registered by a search, the bot takes it over with its own causal chain: public curiosity never steers the bot, and it never hides an asset from it either.
Narrative emergence, one of the discovery strategies, reads the Investo-Meter’s sector Factor Z (press against social media, |z| ≥ 1.2, fresh readings only) and proposes the strategic producers of the diverging sector itself.
OMEGA’s weight in the conviction has to be earned by measurement. A report card, computed every Sunday, grades each archived assessment at the asset’s own horizon, under a protocol fixed on 14 October 2026 before any result:
Met, the criterion lets OMEGA’s weight follow its measured precision, in a new engine version. Missed in a domain, it is not a confession of failure but the refutation loop: the report card points to the distances, domains and source natures that err, so that the graph can be corrected — markets acting as a reality test for OMEGA’s causal claims, as historical series do for cliodynamic cycles. A counterfactual replays the whole archive with and without OMEGA in the conviction, counts the disagreements and grades both engines’ bets the same way.
The action threshold itself is measured in the same way, by a replay of the archive at 0.20, 0.25 and 0.30, under a protocol fixed on 27 October 2026 before any result. Bets are graded as above, net of the round-trip cost and, for a sale, of borrowing; a conflict the LLM would arbitrate is counted and never graded. The bets a lower threshold would add (|conviction| between it and 0.30) are judged on their own: a lower threshold is proposed only if that band holds at least 60 graded bets, a hit rate significantly above 50% (p < 0.05) and a positive mean net return. The study changes nothing by itself: a new threshold is the author’s decision, dated in the observation archive.
Every direct database access of the Trade-Bot goes through a per-process connection pool, with five seconds to connect and five seconds per query (longer only for the super-administrator’s exports): a slow database cannot freeze a worker indefinitely. The audit log is meant to be complete: a write that fails is queued, with its original date, and replayed in order at the next successful write, instead of being silently lost.
The public performance page is limited to twenty requests per minute per address. Its equity curve shows one value per day: a finer curve would make it possible to date each operation and reconstruct the positions, which the methodology forbids disclosing. When the data cannot be read, the page says so instead of displaying an invented capital and a zero return. The security scanner of the continuous integration (Bandit) blocks any new medium or high finding.
The Trade-Bot is an experimental speculative methodology designed to refine the system’s cliodynamic analytical framework. It is not an investment advisory service.
During its testing phase, the Trade-Bot will publish aggregate performance metrics (hit rate, Sharpe ratio, drawdown) to validate the causal mechanism detection approach. Individual positions, entry/exit points, and specific trade signals are never disclosed. All position data is archived internally for methodological audit but does not constitute, and must not be construed as, investment advice, a recommendation, or a solicitation to buy or sell any financial instrument.
This system is built for analytical research purposes. Any future integration with real capital markets will be conducted in strict compliance with applicable financial regulations (AMF, ESMA, MiFID II).
The intelligence engine does not only produce analyses — it evaluates its own predictions and feeds the results back into its models. This is the final phase of the intelligence cycle: perception → decision → evaluation → learning.
The sections above describe the system as it runs today. This journal dates every change of method, with the reason when the reason is a defect of the earlier method, so that any series, score or decision can be read against the method in force on its day. Changes to the inputs of the Trade-Bot during its shadow period also carry their exact deployment time in the data audit (“switch dates” table), which is what splits the shadow period. Newest first; “PR” is the change in the code repository, “audit” an addendum of the data audit of 16 October 2026.
| Date | Module | Change, and why | Reference |
|---|---|---|---|
| 2026-11-15 | Event detector | A keyword counts only in an article of its own measured language, and demonyms cover every country of the zone table in French and English. Defects of the earlier rules: in a sample of 20,000 articles over 14 days, every German “conflict” came from “war” (“was”: a lantern parade, a party funding leak); demonyms covered about fifty countries, so that “Serbian broadcaster” or “Kenyans” placed nothing (22 of 453 unplaced titles are placed by the new table). The same sample shows that a third of the press collected, in other languages, is not yet read for events. | audit 113 |
| 2026-11-15 | Observation archive | Each scan archives, for each asset, the paths of the three main contributions to OMEGA’s direction, its components and its dominant driver. Defect of the earlier archive: it kept only the direction, its strength, coherence, layer and the nature of its main source, so that the share of a source in a past direction could not be read back (the stories without a place, closed on 14 November, among them). What the engines read is unchanged. | audit 112 |
| 2026-11-15 | OMEGA resonance | The conflict-to-markets weight of the propagation matrix applies only to conflicts, and a conflict no longer sends resonance to volatility, hydrocarbons, safe havens or world equities. Defect of the earlier rule: a weight of unverified origin applied to every event (elections, unrest, sanctions included) and every financial node; after 32 interstate crises the expected channels did not move beyond the ordinary, and after 15 major wars only the Swiss franc passed the rule, the VIX and defence falling just short. | audit 111 |
| 2026-11-14 | Event detector | A concluded calm (ceasefire or truce agreed or in force, peace agreement signed, sanctions lifted) feeds no story of conflict, sanctions or diplomatic crisis, and the validating model rejects such a story. Defect of the earlier rule: the detector knew only escalation types, so that “Ukraine and Russia agree ceasefire” became a conflict in Ukraine and a lifting of sanctions a sanctions pressure, pushing assets in the direction of an escalation; over 30 days, 86 of 2,870 articles of such stories carried a word of calm, of which only about a third reported a concluded act. | audit 111 |
| 2026-11-14 | Event detector, Oracle, OMEGA | An article without a recognised place forms no story, and an event without place is linked to nothing by the causal chains of the Oracle and of the OMEGA graph. Defect of the earlier rule: every unplaced article of a type fell into one “Global” story named after its loudest title (a resignation in Paraguay gathering an article on the Pentagon; “conflict escalations” holding music awards and beer festivals), 118 such stories stayed active, and a place-less event counted as linked to every place, so that the Paraguayan political crisis was chained to elections in Haiti, South Sudan, Pakistan and Brazil. | audit 110 |
| 2026-11-14 | Event detector, OMEGA | The place table holds every country (French travel advice names, US travel advisory names, the 193 United Nations members), names that are also first names or common nouns needing confirmation by another place of the article. Defect of the earlier table: it held about half of the world’s countries, so that a resignation of the president of Paraguay or a copper strike in Chile was placed “Global” and reached no country, and it placed Guinea-Bissau in Guinea and the Democratic Republic of the Congo in the Republic of the Congo. | audit 109 |
| 2026-11-14 | OMEGA directions, event studies | A political crisis weighs on its country’s equities only in the countries where the multi-bloc event study supports a fall of the market (Turkey; India and Kazakhstan through their regime class); in China, France, Germany, Hong Kong, Japan, South Korea, Taiwan and the United Kingdom its evidence counts as ignorance. Defect of the earlier rule: it weighed every political crisis at −1 on the equities of its country, which the study does not support in eight of its eleven countries; across the whole matrix, 1 of 154 country-market cells passes the rule, where chance alone would give 5.6. In the study, an event measured on the two proxies of the Kazakh market counts once in its regime class, not once per proxy. The precursor test (defiance followed by a crisis) gets its own protocol, fixed before any computation: a crisis counts only from 20 sessions after the defiance, on samples not yet read. | audit 103 |
| 2026-11-12 | OMEGA directions | A story of conflict or of diplomatic crisis reaches the equities of its country only if that country is at war as measured by the hotspots (the rule already applied to the supply of producing countries). Defect of the earlier rule: an indictment of Iranian agents for a foiled attack plot, typed as a conflict and located in the United States, was the dominant driver of 36 of the 49 oriented assets, the S&P 500 and the Nasdaq included, at full strength. | audit 108 |
| 2026-11-11 | Trade-Bot | Validation of the wave detector on the price archive (protocol fixed beforehand): no direction qualifies for an automatic switch, for the market wave as for the semiconductor wave. A rising wave gave no edge 60 sessions ahead; a falling wave was followed by returns above the median (MSCI ACWI +3.2 points). The Offensive profile can therefore be declared only with a rising wave. Defect of the earlier setting: an Offensive profile in a falling wave gave more room to sales just when markets had tended to rebound. | audit 105 |
| 2026-11-11 | Trade-Bot | Reference of the semiconductor chain (transformers, end users, sector funds), each member with the share of its activity tied to the chain, its fiscal year and source. Defect of the earlier reading: the actors table ties a company to a material, never to a use, and a mechanical rule brought cable, solar and chemical companies into the chain while leaving RF chip makers out. No input of the engines changes. | audit 107 |
| 2026-11-11 | OMEGA zones, Trade-Bot | A sub-national place of the zone table (Hong Kong, Tigray, Amhara, Oromia, Darfur, Ituri, Goma) takes the zone of its country. Defect: these places had no zone although they named a country, so their stories were neither merged with the stories of the same type of that country nor grouped in its theatre. A sector wave (SOXX, semiconductor chain) is published as an observation beside the market wave, with its validation by the same protocol. | audit 106 |
| 2026-11-11 | Trade-Bot | Capacity profiles for both books (Prudent 8 positions and 35% invested, Balanced 12 and 50%, Offensive 16 and 75%), switched by the administrator and logged; the Offensive capacity serves only bets that ride the declared wave and are exploitable by their own trend (20-session average, support and resistance ratio). The technical summary of each asset publishes its 20-session average. One menu sets the profile, and a manual LLM analysis follows it; a decision cached under one LLM profile is no longer served under another (defect: the cache fingerprint ignored the profile). A market wave detector (S&P 500 and MSCI ACWI, 200-session average) publishes its reading as an observation; its validation protocol is fixed before any computation. No input of the engines changes until a profile is switched: the Balanced profile is the earlier setting. | audit 105 |
| 2026-11-11 | OMEGA directions, Trade-Bot | Aggregation of OMEGA evidence: one piece of evidence per theatre and direction, the strongest, after one per source; the strength becomes a noisy-OR by direction, bounded between −1 and +1, instead of a signed sum without bound; each assessment publishes its components and dominant driver. Defect of the earlier method: a hotspot and a press story of the same theatre counted twice (Yemen, −0.570 and −0.307 on semiconductors, 2 November), and evidence added up without limit. Both books may hold 12 positions instead of 8, at most two sharing one dominant driver (v1 layer caps 4/6/3 instead of 3/4/2), and an order refused for lack of room is examined again when a place frees up. Defect: on 9 November the v2 book refused all 13 of its pending orders at the cap of 8 and none came back when three positions closed, while both books held correlated short bets on semiconductors and metals from one OMEGA source, four of which closed at a loss within two hours that day. | audit 104 |
| 2026-11-11 | Event studies, OMEGA | The protocol of the multi-bloc event study of political crises is fixed before any return is computed (eleven countries, 188 dated events in three groups: domestic crises, popular or parliamentary defiance, interstate crises). It tests, country by country, the OMEGA rule that weighs a political crisis at −1 on the country’s equities, which the United States study did not support. The bloc reference gains dated economic blocs (EU-Turkey customs union, BRICS, SCO, RCEP, CPTPP, Eurasian Economic Union, NAFTA then USMCA); regime classes are read year by year from V-Dem, a year not yet published staying absent. OMEGA now records, each day and for each country, the events of its graph located in that country and their weights, an event without an established country counting for none, so that a later test can ask whether OMEGA saw a shock coming before it became public. | audit 103 |
| 2026-11-11 | Price archive | The Sensex, the BIST 100 and the TAIEX join the declared universe of the daily price archive (159 assets in all), with two London-listed certificates, Halyk Bank and Kazatomprom, archived under their own names as proxies of the Kazakh market, which has no index in the data provider. They serve the multi-bloc event study of political crises, whose protocol is drafted and fixed before any return is computed; the Nifty starts only in 2007 in the provider. No input of the Trade-Bot engines changes. | audit 102 |
| 2026-11-08 | OMEGA directions | The Hang Seng is read from its provider’s monthly factsheet (top 50 holdings, 90 % of the index, with their weight and Hang Seng sector) and published as an observation, outside the directions: weights a month old are not today’s composition, and the drift between two factsheets is measured first. The full composition with weights is not published by the provider. | PR Hang Seng, audit 101 |
| 2026-11-08 | Trade-Bot (v1 and v2) | Both engines read the ignorance of the OMEGA assessment: the v2 engine counts it as the unknown share of OMEGA’s weight in the position size, the v1 engine lowers its confidence by 0.2 × ignorance, as for the Oracle. A neutral OMEGA counted as fully known: on 8 November the S&P 500, ignorant at 99 % on its materials, weighed as a certainty of calm. Almonty has no country of production until its mines publish their output (it carried Canada, its seat). | PR ignorance d’OMEGA, audit 100 |
| 2026-11-08 | OMEGA directions | The S&P 500 is read through SPY (State Street) and the Dow Jones through DIA, with the GICS sectors of a dated reference (iShares file of 5 November 2026): iShares answers the server with HTTP 403, so on the first run only the Dow was read, by its companies and without any sector. FTSE 100, DAX and Hang Seng stay unavailable; an index weight weeks old is not today’s composition. | PR Nvidia, audit 99 |
| 2026-11-08 | OMEGA directions | NVIDIA enters the actors table as an end user of silicon and hafnium, with no country of production (its wafers are made by foundries outside the United States, in Asia, without a named country): it reads their scarcity through the world price only. It weighed 6.9 % of the S&P 500 with no material at all. TSMC and GlobalFoundries, foundries, are relabelled transformers of silicon; lighting and concentrated solar power get their GICS sectors. | PR Nvidia, audit 98 |
| 2026-11-08 | OMEGA directions | Five equity indices (S&P 500, Dow Jones, FTSE 100, DAX, Hang Seng) are read through the holdings file of an ETF that replicates them: their companies of the actors table carry a direction at their weight (5 November 2026: 11.9 % of the S&P 500, 1.7 % of the Hang Seng), the rest only a sector width, published as ignorance. They read no material before, for lack of sourced weights; reading an index as its sector had pulled the Nasdaq down at full force through gallium. | PR indices par leurs positions, audit 97 |
| 2026-11-07 | OMEGA directions | Eleven companies held by REMX enter the actors table, from a sourced list validated by the author (coverage of the fund 85.5 %, from 63.1 %); a company with mines in several countries splits its weight by the output of each site (EQ Resources: Spain 66.5 %, Australia 33.5 %, WO3 of the quarter to 30 September 2026). A single country per company would have placed all its output in one of them, and two entries would have counted its weight twice. | PR acteurs de REMX, audit 96 |
| 2026-11-07 | OMEGA directions | REMX is read as a weighted basket of its real holdings (issuer file, daily): lithium at about 27 % of the fund through its producers, rare earths, tungsten, each at its weight. It received every material of the “rare” sector at equal weight: gallium (one holding, 7.6 %) weighed as much as lithium, and tungsten (about 13 %) was not read. | PR composition des fonds, audit 95 |
| 2026-11-07 | OMEGA fragility | A third, social dimension of the fragility base, its amplitude set by history from a protocol written before any result: the United States enter at F = 1 (stressed down months 2.46 times deeper, interval 1.21–3.91); France and Germany stay observations. The base knew market rates and structure, not the households that carry consumption. | PR fragilité sociale, audit 93 et 94 |
| 2026-11-07 | Real Cost Index | France and Germany read the last complete month of the HICP, no longer the last month published: from the flash estimate of a month (total only) to the release of its sub-indices, about two weeks each month, both indices had no component (none from the end of October to 7 November). The FAO Food Price Index is read at its current address; the former one answered 404, and food lost its world prices. | PR coalitions et ICR, audit 92 |
| 2026-11-07 | Game theory | The coalition of the two players leaves the payoff matrices of the five scenarios and is published next to each game, with the dated alliance commitments of the players (NATO and the EU at their members of the day). Added uniformly to every strategy, it changed no equilibrium, deviation, belief or risk level, only the value of the game and the robust interval handed to the Trade-Bot’s analyst (measured on 7 November: 5 outputs of 87 to 97 per scenario; the value of the China triple-leverage game changed sign). Its graph called the United States–Israel link a defence pact, which no treaty establishes, and knew NATO and the EU by thirteen members each. | PR coalitions, audit 91 |
| 2026-11-07 | Trade-Bot orchestration | The Trade-Bot’s timed tasks (scan, decisions and position management of both engines, mark-to-market, evaluation, daily snapshot, calibration) run on a worker of their own. They waited in the batch queue behind long tasks: on 7 November at 00:25 UTC its two processes ran the hourly flux-risk precomputation (up to 20 minutes) and the local-conflict extraction (up to 25 minutes), with 26 tasks waiting, among them the Trade-Bot’s mark-to-market and position management, which never retry. Both engines move together; their code and loaded modules are unchanged. | PR worker Trade-Bot, audit 90 |
| 2026-11-06 | Game theory | Armed conflicts in progress, one tab per conflict: belligerents from the UCDP (yearly dataset and the current year’s monthly events), states at war outside their territory with the share of events, and the commitments of their alliances (thirteen, of three natures, dated memberships). The alliance graph of the scenarios knew NATO and the EU by thirteen members each, without dates or the nature of the commitment. Display only: no engine input changes. | PR conflits UCDP, audit 89 |
| 2026-11-06 | OMEGA graph and directions | Export controls reach the supply of the regions they target, the other regions through the world price; their deadlines are dated rendezvous without direction. A story such as “China reinstates its ban on gallium exports to the United States” reached no material (a sanction in China carries no supply risk outside a measured war), but would have lifted crude oil, the United States being an oil producer; China’s suspension of that ban ends on 27 November 2026. | PR contrôles à l’exportation, audit 84 |
| 2026-11-06 | OMEGA graph and directions | Materials have one supply node per region: a cut route reaches the supply of its region of destination, by that region’s import dependence, and the other regions through the world price (0.5); companies of the strategic-materials actors table read their materials by their role (cost for users, substitution for producers outside the disrupted origin); an equity index reads no sector until its sector weights are sourced. On 5 November, the Yemen escalation, through the China–Europe route, pulled the Nasdaq down by 1.52 through gallium, at full force, although the route serves Europe. | PR matériaux régionaux, audit 83 |
| 2026-11-05 | OMEGA directions | Crude oil is wired to the straits: a conflict in a zone that borders one reaches Brent and WTI, weighted by the strait’s crude flow; the defence sector counts a conflict by its demand only. On 5 November, the Yemen escalation reached gallium, terbium and dysprosium through the China–Europe corridor but had no path to crude oil (Brent and WTI without any signed evidence), and it pulled defence down through the cost of terbium (−0.79) harder than the demand it creates (+0.5). | PR détroits et défense, audit 82 |
| 2026-11-05 | OMEGA hotspots | A de-escalation pushes only when it holds two complete months in a row; an escalation still pushes from the first month. On 5 November, Burkina Faso read 30 deaths for September against a median of 338.5, a fall more likely due to under-reporting under a media blackout than to an easing. | PR désescalade, audit 81 |
| 2026-11-05 | Conflict index, OMEGA hotspots | The escalation is measured on the last month the data fully cover (7 days after its end, from the version date of the ACLED file); each month sums all its rows; HAPI is queried one request at a time with retries. On 5 November, a file versioned 30 October gave Nigeria 702 deaths for October against 1,048 in the next version, so that Myanmar and Mali read as de-escalations; the most aggregated level alone gave Ukraine zero deaths; and 44 countries out of 76 fell back on old files after rate-limit refusals. | PR complétude, audit 80 |
| 2026-11-05 | OMEGA hotspots, conflict index | A hotspot pushes by its escalation (deaths of the last month against its own 24-month normal, from two robust deviations, both ways), no longer by its level, which remains a state and sustains defence demand alone; ACLED series through the HDX Humanitarian API, the files as fallback. On 4 November, eleven of the twelve upward commodity directions came from the levels of stable wars, written as daily changes (Nigeria at 83 lifted Brent by 0.83 every day, more than the news of the day). | PR escalade, audit 79 |
| 2026-11-05 | Conflict index, OMEGA hotspots | Each country carries the date of its ACLED data; a download refused by the source falls back on the last downloaded copy, dated; data older than 45 days, or a carried-over value without a date, makes no hotspot. Since 2 or 3 November, the data host has answered the server’s downloads with an anti-bot challenge: the index carried over the previous value for all 76 countries, without any age limit and under a fresh computation date. | PR reprise datée, audit 78 |
| 2026-11-05 | Oracle, event detection | Forty candidates a pass instead of twenty, validated in batches of ten, at most two batches a minute. With ten titles per candidate, four batches a pass exceeded the first model’s token budget and the last batch went to the fallback model, whose daily quota is shared by all small calls. | PR reprise datée, audit 78 |
| 2026-11-05 | Oracle, event detection | Corroboration spans 24 hours instead of one pass: a story reported by one newsroom waits for a second one, and an active story is fed by a single article; the confidence of a story is capped at 0.75 for its whole life. The first pass of the two-newsroom rule (5 November, 00:11) set aside 19 single-article stories, Hormuz and Iran among them, because a pass reads only 45 minutes of articles; and stories born under the earlier rule kept the model’s 0.9 or 0.95 as long as they received articles. | PR corroboration sur 24 h, audit 77 |
| 2026-11-04 | Oracle, event detection | A story becomes a candidate only when two distinct newsrooms report it, and its confidence follows the number of newsrooms, which the language model can lower but not raise; articles published more than 24 hours before their reading are left out; the model reads ten titles with their newsroom and rejects questions, hypotheses, opinion and projections, and its label states only what the titles establish. Found on 4 November: ten articles from a single opinion magazine, published before the US midterm vote and read in one block the next morning, made a “contestation of the election results by Trump” (confidence 0.7) that the press of the day contradicted; it turned 33 US assets downward in OMEGA. | PR sources indépendantes, audit 76 |
| 2026-11-03 | OMEGA directions | Coherence is measured on signed evidence, and unknown-sign exposure is published as ignorance; a strait, a sea or a precise region counts as a common place for triggering a chain. Traced on 3 November: the Oracle detected the attacks on ships in the Strait of Hormuz, OMEGA gathered evidence that was all upward for Brent (+0.99), yet published no direction, because unknown-sign paths counted against the direction (coherence 0.32) and a strait, naming no country, could not trigger the Hormuz chain. | PR cohérence et lieux, audit 74-75 |
| 2026-11-03 | Trade-Bot (comparison engine, v2.5) | The structural levels (real rates, ICR, LSI, HHI) leave the conviction and are published as observations, until they return as deviations from their own history. They were read against fixed “neutral” values that are not measures: replaying the v2.3 shadow period, the real rate (2.63% against 1.25%) gave gold −1 at every scan for ten days, and 90% of the convictions leaned down. | PR v2.5, audit 72-73 |
| 2026-11-03 | Port occupancy | The water mask of a port extends its window from 60 to 120 days when the cloudy season leaves too few clear views. Santos had two clear views per pixel over 60 days and no value. | PR masque de Santos, audit 71 |
| 2026-11-02 | AIS transits | Ship types are kept in a lasting register, and a passage has no variation while fewer than 80% of its ships have a known type. Static data are rare for each ship and every restart forgot all types: half an hour after a restart, 103 ships out of 156 in the counting zones had none (Malacca: 61 unknown for 1 merchant ship). | PR registre des types, audit 70 |
| 2026-11-02 | AIS military watch | Two Baltic watch zones (Kaliningrad, Gulf of Finland) are added to the military watch; they count no transit. The AIS connector forgets a ship silent for six hours; it kept every ship it had ever received. | PR veille Baltique, audit 69 |
| 2026-11-02 | AIS transits | The transit volume counts merchant ships only (cargo ships and tankers), and military ships are counted apart in a watch that enters no index. The volume counted every ship under way, ferries, fishing boats and pleasure craft included, and the AIS military type was filed as towing. | PR transits marchands, audit 68 |
| 2026-11-02 | Port occupancy | Occupancy is measured on real berths, where ships have been seen in sixty days, and the published value is the radar alone; the optical measurement (near infrared, against the port’s permanent water) is published in validation; Haifa is added. The earlier ratio counted every shoreline of the box as empty berths and left out berths occupied most of the time (Ningbo at 1.2%), and it averaged optical and radar measurements up to 40 points apart. | PR ports, audit 66-67 |
| 2026-11-02 | OMEGA directions | Only a fund whose materials are declared is an alternative producer, of those materials; a company is not one by its sector alone, pending a sourced table of what each company produces. The sector rule gave every company of a producing sector all of that sector’s materials: measured on 2 November, Air Liquide and Linde, classed as rare-earth companies from the “Specialty Chemicals” label of their data provider, rose as alternative producers of dysprosium. | PR précision du dossier, audit 65 |
| 2026-11-02 | Trade-Bot (main engine dossier) | The dossier states the Oracle’s precision by domain, under its current method, or says it is not available. It read domain keys that the Oracle never published, so the line was always empty. | PR précision du dossier, audit 64 |
| 2026-11-02 | Lightning stream | The lightning stream follows the same single owner as the AIS stream, renewing its lock and publishing from its own thread; its lock had the same flaw (renewed only when the scheduled task landed on its owner), so strikes received by a former owner were lost to the map. | PR précision du dossier, audit 64 |
| 2026-11-01 | OMEGA, ticker registry | Social unrest moves nothing before 14 days of the same story, then acts through the country’s flows (its production), not its equities. Every automatic exposure classified before the unknown-country rule and never checked is unverified, checked daily against the declared country and read by neither engine until then (only the old “US” default was checked). Measured on 1 November: Gen Z protests in Germany pulled ten equities down at full strength, seven of them British, Norwegian, Dutch or Irish companies that the registry exposed to Germany alone. | PR troubles et expositions, audit 63 |
| 2026-11-01 | OMEGA | An event triggers a written chain only where the chain places its cause, and a sanctions story acts through energy prices only, when it targets a producer. Measured on 1 November: a UN vote on the US embargo against Cuba triggered the 1973 OPEC embargo chain and pulled ten miners down; a military deployment in Colombia triggered the Taiwan escalation chain and pulled five semiconductor stocks down; US sanctions against Rwanda raised crude oil; UK sanctions against Sudanese commanders pulled the FTSE down. | PR OMEGA indications, audit 62 |
| 2026-11-01 | AIS stream | The process that holds the AIS connection renews its lock and publishes from its own thread, and a connector that has received nothing publishes its health only. Measured on 1 November: the published state said “not connected, 0 message” and all nine passages were uncovered, while another process of the worker was receiving thousands of positions. The lock was renewed only when the scheduled task happened to land on its owner; it expired, and the next process published a connector just started. | PR AIS, audit 61 |
| 2026-11-01 | Trade-Bot (both engines) | Engine v2.4: the analytical engine applies to its smoothed conviction the Oracle precision factor that the main engine applies to its confidence, so that the two engines stay comparable; both read only precisions validated under the Oracle’s current method, over all its domains. The main engine read the precisions of three domains, validated on composites worth 0.5 by default. | PR fraîcheur et facteur, audit 60 |
| 2026-11-01 | Data catalogue | The expected freshness of every key follows the cadence of the task that writes it. Measured on 1 November: 18 keys were expected more often than written (the market regime every hour, written every 4 hours and 8 at night; energy stress and the Maslow pyramids every 6 hours, written once a day), so that up-to-date values were declared carried and withheld from the main engine’s dossier. | PR fraîcheur et facteur, audit 60 |
| 2026-11-01 | Oracle | Energy stress and port activity are read in the shape their writers publish, and the market regime and Fear & Greed learn from values younger than their writing cadence. Measured on the server after the deployment of the components: both sources were measured but their components were absent (wrong field names), and the two others would never have left their warm-up, being always older than the catalogue’s one-hour freshness. | PR Oracle formes, audit 59 |
| 2026-11-01 | Oracle | Twenty-two components read the system’s own sources through the catalogue, each on its own scale (deviation from its own history), with the author’s weights; flows and supply (freight, LSI, transits, ports, GDELT tone) become a fifth domain; an absent component or domain is absent and enters the fusion as ignorance, a partly covered domain is discounted by its coverage; the signal history accepts absence and carries its method, and validation, weights and calibration use the current method only. Measured on 31 October: of nine components, three read tables nothing fills any more, GDELT a column that does not exist, the RSS counts failed on their LIKE patterns, ACLED saturated, and every failure was worth 0.5, which the fusion took for evidence of stability; game theory counted a missing risk level as “moderate” and the absence of any scenario as 0.5. In OMEGA, an Oracle event node without a published precision is no longer weighted by an invented reliability of 0.5. | PR Oracle composantes, audit 58 |
| 2026-11-01 | OMEGA | The Archiviste precedents and the lagged-correlation anticipations follow the direction of the moves, as the level itself does: only a move toward risk counts, and only a positive correlation announces a follower. A sharp fall of the VIX with a precedent raised the level; the branch for five anticipations or more came after the branch for three and was never reached; the cache of external precedents was never read. | PR OMEGA sens, audit 57 |
| 2026-10-31 | Trade-Bot (both engines) | The press tone of an asset is read against the ordinary tone of the news, then against the asset’s own history. The v2 engine took twice the raw mean sentiment, the main engine the raw share of negative articles (about 62 out of 100 for an ordinary asset): a press negative by nature pushed both engines to sell, and since mid-October they had decided 159 SELL for 10 BUY. A missing score counted as a neutral sentiment of 0. | PR ton de presse, audit 56 |
| 2026-10-31 | Trade-Bot (both engines) | Technical analysis is computed on the price archive for every asset of the registry, under each asset’s declared ticker, around the clock, with an unmeasurable trend absent and a silence published as such. Measured on 31 October: none of the 107 assets had a technical reading. The registry looked an asset up under its own name (GOLD) instead of its ticker (GC=F), the task covered 35 fixed symbols from 8:00 to 20:00 UTC in a key that expired at night, and an unmeasured trend counted as a measured sideways trend. The tactical layer was left to momentum and press sentiment. The main engine no longer falls back on a global score for an asset without its own analysis. | PR analyse technique, audit 55 |
| 2026-10-29 | RSS feeds, Politico-Meter | The feeds of transnational and supranational institutions (EU, UN) carry their institution’s bloc instead of a country, and are national press for the personalities of that bloc only. An institution has no country of edition: the validated list gave several UN feeds the country of the region they cover. | PR flux institutionnels, audit 54 |
| 2026-10-29 | Politico-Meter, RSS feeds | Transnational blocs become a reference with dated memberships, and a bloc’s national press follows from its member countries. 72 feeds carried the code EU in place of their country, because the Politico-Meter stored the national press of its European personalities in the feed’s country: correcting those countries would have stripped those personalities of all national press. | PR blocs, audit 53 |
| 2026-10-29 | RSS feeds, Oracle events, country press | Each feed carries the country of its edition, declared in its configuration, and the emitter reads it before the domain’s extension. The extension says nothing for generic domains (.com), which carry most large outlets. | PR pays des flux, audit 52 |
| 2026-10-28 | Oracle events, country press | A place name the article leaves ambiguous takes the emitter’s country when the domain’s national extension and the measured language agree; otherwise it stays absent. Local press omits the precision that international press gives (“Georgia” on a Georgian site). The feeds’ declared language is not read: 809 of 829 active feeds carry the column’s default value. | PR émetteur, audit 51 |
| 2026-10-28 | OMEGA directions (both engines) | An event located in Israel, Lebanon or Syria reaches the assets exposed to that country. Their names are also places of a theatre (Israel-Palestine, Levant) and were read as regions without country: a conflict located in Israel reached no Israeli asset, while one located in Gaza did. | PR pays de théâtre, audit 50 |
| 2026-10-28 | Oracle events, country press | An ambiguous place name is settled by the article itself (a US state after a city, the other places named, capitals, compound names), and a located story comes before a story without place. An “election” in “Georgia”, the US state, was placed in the Caucasus, and three of the twenty candidates were stories without place. Israel, Lebanon and Syria, also theatre names, were never counted in the country press. | PR homonymes, audit 49 |
| 2026-10-28 | Trade-Bot (both engines) | A country press surge is measured on the country’s share of the articles read, not on its count, and a window under 12 counted hours decides nothing. The count made a general influx a surge for every country: the articles read per day ranged from 3,266 to 19,002 over the 28 days counted. Corrected before any measured reading reached the engines. | PR part des pays, audit 48 |
| 2026-10-28 | Trade-Bot (both engines) | A press surge on one of an asset’s exposure countries, counted by the content of the articles and never by their outlet, makes the asset one layer faster in both engines, without direction. An asset’s press only held the articles naming it: Petrobras had none for four days (reading of 27 October) while Brazil went from 32–39 articles a day to 269 on 26 October, the day it opened 17% lower. | PR presse des pays, audit 48 |
| 2026-10-28 | Oracle events | Up to twenty candidates per pass, the best of each type first. The ten largest, all types together, were kept: over six hours on 28 October the political crisis in Brazil, all three diplomatic crises and six of the seven sanctions stories fell below the cap, and the last energy crisis dated from 20 October. | PR candidats par type, audit 47 |
| 2026-10-28 | Oracle events, OMEGA | The event detector knows political crises (contested results, coups, martial law, impeachments) and elections. A political crisis is a risk for its country’s markets, an election a dated rendezvous without direction. With no political type, the Brazilian presidential election of 25 October (269 articles the next day, a result contested by the losing side, Petrobras opening 17% lower) produced no event, nothing in OMEGA and nothing in the Oracle. | PR événements politiques, audit 47 |
| 2026-10-27 | Trade-Bot | A bet decided on a closed market waits for the opening price of the next quoted session, for both engines. It was filled at the previous close: positions opened on a Saturday or after a European close (24 and 23 October) skipped the opening gap that a stress session puts to the test. | PR exécution à l’ouverture, audit 46 |
| 2026-10-26 | Price archive | The 22 individual stocks shown in the Console’s sector tabs join the daily price archive (154 assets in all); they showed n/a in the trend column. The archive pass now spaces its calls to the data provider by one second (at most 60 calls per minute) instead of 0.4 second, a conservative ceiling since the provider publishes no formal quota. A contract test fails when a symbol of the sector tabs is not archived. | dashboard proposal §13 |
| 2026-10-26 | Strategic flows, GDELT media risk | The tone term of the per-material media risk score is the tone of the producing countries weighted by their share of world production, read from the open GDELT event files, instead of the tone of the articles (which the article list does not carry, so the score was absent). Producers outside the 66 key countries, or without an event, do not weigh; the covered share is published and a coverage below 50 % leaves the tone absent (the threshold is a parameter of the method). The coefficients of the score are unchanged. | audit addendum 44 |
| 2026-10-26 | Stress Index, GDELT tone | The GDELT tone component of the composite flow stress (weight 25 %) reads the mean tone of the open GDELT event files collected by the relations module (events between key countries over the last seven valid days, weighted by events), published twice a day by the Maslow bridge. It read the tone of the conflict articles of the map, which the document API does not provide and the code read as 0.0, so the component was absent. The normalisation is unchanged (−5 → 100, +5 → 0). This is not a conflict tone: the four event classes are aggregated together, so the component measures the tone of relations between key countries. | audit addendum 45 |
| 2026-10-26 | Price archive | Five sector ETFs shown in the Console’s sector tabs (aerospace and defence, technology, biotechnology, agribusiness) join the declared universe of the daily price archive; they showed n/a in the trend column. A contract test now names the individual stocks of those tabs that are not archived. | dashboard proposal §13 |
| 2026-10-26 | Dashboard, price history | Each line of the Console’s Markets tab carries a 20-session trend line read from the daily price archive (one grouped request per tab, never a live call to the market-data provider). The line is drawn from raw closes only; the current session is marked as provisional with a hollow end point; a symbol that is not archived shows n/a and an unreachable archive leaves the cell empty, with no filled-in or interpolated values. The tab had no history column; the candlestick charts remained one click away. | dashboard proposal §13 |
| 2026-10-26 | Strategic flows, alerts | An alert’s creation time is written in UTC and published as an ISO 8601 time with a Z suffix. The database column stored the server’s local time (Paris), which the API labelled GMT: an alert looked one to two hours younger than it was, and the 24-hour window used by the Trade-Bot and OMEGA actually covered 25 to 26 hours. The duplicate check uses the same clock. | audit addendum 42 |
| 2026-10-26 | Maslow, GDELT tone | The macro tone and the per-country tone read by the Maslow pyramid come from the open GDELT event files already collected by the relations module, as a mean weighted by events over the valid days of the last week. They were the mean of the per-material article tones, which the document API does not provide and the code read as 0.0: the macro tone was 0.0, no country had its own tone, the French input read the global tone before its own, and the other countries fell back on the global tone. A country without a measured tone is now absent. | audit addendum 43 |
| 2026-10-26 | Strategic flows, GDELT media risk | The per-material media risk score from the GDELT document API keeps its coefficients but exists only when each of its three inputs is measured. The article list carries no tone, so the tone term was read as 0; articles with an unresolved country were counted as one country holding 100 % of the coverage, adding 20 points; and a failed request returned an empty list scored 0. Every material with 25 or more articles therefore scored 70 (“critical”): five such alerts reached the alert table and the Trade-Bot, and the score weighed half of each material’s media signal and fed the Maslow macro tone. The score is now absent (n/a) with the missing components named; an unmeasured tone gives an article no severity; a failed request is a failure; the five stored alerts are kept but no longer served or counted. The Trade-Bot inputs listed in the data audit change for both engines at the same instant. | audit addendum 40 |
| 2026-10-26 | Oracle events | The events summary (count, mean and top events that feed the political boost) now counts the last 24 hours exactly. It compared a local-time string with a UTC string, so every event of the cutoff day was counted and the window was shifted by one to two hours. The event list is ranked on persistence before being cut: it kept the first three times the limit by raw confidence, which let old radio anomalies push out fresh events (19 radio events out of 20 listed). A sensor event’s label follows its latest reading instead of the one it was created with, and a confidence of 0.0 is no longer read as 0.5. The per-level confidences and the boost formula are unchanged. | audit addendum 41 |
| 2026-10-26 | Dashboard, strategic flow alerts | The Console’s strategic-flows tab lists the last seven days of alerts from the alert table, with the state of the generator that produces them: time of its last run, whether it is late or failed, which sources were silent and how many materials GDELT gave no measured score. An empty list is shown beside a warning when the generator has not run, since an empty table is otherwise indistinguishable from a stopped generator. A missing score or severity shows n/a, and an unreadable time shows n/a rather than a computed age. The tab had no alerts. | dashboard proposal §14 |
| 2026-10-26 | Stress Index, country risk | The composite flow stress publishes no score when none of its four sources is available: it published 0 (read as “calm”), and each missing component was published at 50. A Fear & Greed without a value is not a source, a spread or a tone of exactly 0 is read instead of dropped, and the market-sentiment component is labelled for what it reads, the crypto Fear & Greed. The Investo-meter country risk no longer replaces a missing stress or Fear & Greed by 50: the component is left out, the weights of the others are renormalised, and a measured 0 stays 0. The scheduled task publishes a failed computation instead of leaving the previous value served for up to seven hours. Not an input of the Trade-Bot (it reads the corruption component only). | dashboard proposal §11–§12 |
| 2026-10-26 | Dashboard, quotes and events | Quotes are cached 60 seconds for all visitors, carry the UTC time of the reading and can be fetched up to 20 at a time. The Console separates news events from HF radio anomalies (19 of 20 listed events), shows the radio level and ratio from the latest reading (the event label is frozen at creation) and gives the zone and the radius exactly as the SDR module publishes them (the radius runs from about 1,000 to 8,000 km in the readings observed), with no threshold of its own. The confidence of a news event is presented as a language-model estimate, not a probability; the radio “confidence”, a constant per level, is not shown. Statistical deviations are listed beside events, never merged with them, with partial days and derived indices flagged. | dashboard proposal §10, §12 |
| 2026-10-26 | Dashboard | A one-line pulse strip and a tabbed Console replace the five-column bottom drawer. The strip ranks the two best and two worst daily moves among the nine headline equity indices that have both a price and a change (a missing change is never read as 0; a reading older than 36 hours is flagged), gives the OMEGA level, the composite flow stress (n/a when no source feeds it: the published score is then 0, which is not a measure) and the number of active weak signals, each with the age of its reading. The drawer repeated the OMEGA brief, showed 8 of 8 radio events, hid behind the ticker and ran past the screen on a 1366 px laptop (two-line header). The quote API no longer turns an unknown change into 0. | dashboard proposal §4, §11 |
| 2026-10-25 | Dashboard, strategic flows | Values shown in the dashboard keep their class. A strategic-flow score is marked computed (with its date) or static (the catalogue reference), and the list no longer lets the static value replace a computed zero, nor a 0 stand for a missing value. An absent score, confidence, quote change, sentiment or country risk reads n/d and stays out of the shares and the average, instead of counting as 0 or as “neutral”. The macro pulse publishes no score while its cache warms up. The old fallbacks produced figures no source had given. | dashboard proposal §2, defects 1–6 |
| 2026-10-25 | Collection | Fertiliser imports of three blocs by supplier (UN Comtrade), with the Russian and Belarusian share as an interval bounded by the re-export hubs. The input channel had prices and import dependence but not who supplies whom, so the redirection of sanctioned sales could not be seen. | PR flux par bloc, audit 30 |
| 2026-10-25 | Collection | Fertiliser blocs as non-overlapping cores (the 27 EU members read as one, the United States, the BRICS) and circles of partners (the 84 official EU trade partners; the United States’ free-trade partners, Japan and the EU), with each EU member’s take of Russian imports. The blocs were ten, six and nine large importers: an arbitrary sample, with trade between EU members counted twice. | PR noyaux et cercles, audit 31 |
| 2026-10-25 | Collection | Monthly reading of the Russian and Belarusian fertiliser share (three months against the same months a year earlier) for the bloc cores and the large EU importers. The annual average hid turning points such as the return of Russian potash to France at the end of 2025. | PR lecture mensuelle, audit 32 |
| 2026-10-25 | OMEGA graph, directions | Agricultural input channel: measured fertiliser and gas price shocks (2 σ), measured drops of Russian supply in the EU and the United States, and dated trade measures reach the crops by their agronomic needs and the fertiliser producers outside Russia, in the slow layer; the schedule holds the four steps of Regulation (EU) 2025/1227, each with a declared weight; a daily watch of Access2Markets flags new EU trade measures and reads the cited regulations from Cellar (dates, duties, chapters, countries, each with its article). The graph knew wars only through the producing country: a war waged abroad, sanctions or a gas shock did not reach farming. | PR canal d'intrants, audit 34 |
| 2026-10-25 | OMEGA graph, Trade-Bot dossiers | Dated measures read from EU acts enter the input channel by themselves after 48 hours of review, weighted by the duty change in ad valorem points over 25 (no measured price, no weight), signed by origin; the author confirms, corrects or rejects with a logged reason; deadlines and checkpoints are dated rendezvous without direction, shown in both dossiers. Measures waited on a manual entry into the schedule: one that was forgotten counted for nothing, and the author's own reading could not be compared with a fixed rule. | PR mesures datées des actes de l'UE, audit 37 |
| 2026-10-26 | RAG | Indexes are locked per domain while written, replaced whole, and reread by a process when another has rewritten them; cursors are merged on disk. Each process kept the copy it loaded at start-up and wrote it back at its next pass: a cleaning done from a shell would have been undone by the worker, and two processes could write the same domain at once. | PR index partagés du RAG, audit 39 |
| 2026-10-26 | RAG | Three domains carry the system's own work (dated, classed notes from computed data; EU acts by article; methodology and audit by section), each note opening with its data class, which reports keep. Chunks carry a fingerprint and are never indexed twice; the Sunday rebuild runs one domain per task, smallest first. The computed data were absent from the RAG; 90 % of the security domain and half of the archive domain were duplicates, and the Sunday rebuild stopped before reaching its last domains. | PR RAG enrichi, audit 38 |
| 2026-10-25 | Collection | A missing month is named, and an estimated three-month share is published beside the measure as an interval bounded by the month’s last twelve declared months. Filling the gap with the mean of the neighbours was checked and rejected: on June 2026 it gave 4.6 % against a real 6.3 %. The BRICS monthly window ends at the members’ median last month: it ended at the latest member’s, and China, whose monthly data stop in December 2024, erased the bloc. | PR mois manquant, audit 33 |
| 2026-10-25 | Archiviste | Hand-deposited documents carry a recorded provenance, are identified by their path and deduplicated by fingerprint, and one sub-folder can be ingested alone. Documents were identified by file name only, so two homonymous files overwrote each other, and nothing recorded where a deposited document came from. | PR provenance Archiviste, audit 29 |
| 2026-10-25 | OMEGA directions | A press story carries a supply risk on its country’s commodities only where the measured violence reaches the war threshold. An isolated attack near Paris and contaminated stories in India and the United States lifted wheat, maize and soybeans. | PR intensité des histoires, audit 28 |
| 2026-10-25 | Collection | Agricultural inputs (fertilisers, gas, fuels) collected from the World Bank, INSEE and the European Commission, read against each series’ own history, with intervals for annual values. A war waged abroad reaches agriculture through its inputs (sanctioned gas, nitrogen fertilisers, diesel), a channel the graph did not have. | PR intrants agricoles, audit 26 |
| 2026-10-25 | OMEGA graph | The producer supply channel of an event requires a war on the territory (UCDP war threshold, 60) for a hotspot, and excludes social unrest. After the first hotspots, the United States, a party to the conflict in Yemen, lifted soybeans, and the unrest in France lifted wheat. | PR canal direct, audit 25 |
| 2026-10-24 | OMEGA graph | A hotspot requires an armed conflict in the sense of the UCDP (state party, or territory at war in the theatre of a state party), weighs its lethal violence and floor only, and all hotspots are one background source per asset. On the real index after the first deployment, thirty hotspots gave defence demand a strength of 11 against 0.45 for a press story, and Brazil and Mexico (organised crime, no UCDP conflict) lifted iron ore and silver. | PR correctif points chauds, audit 24 |
| 2026-10-24 | OMEGA graph | Conflict hotspots: every country whose conflict index reaches the armed-conflict threshold is a background node of the graph, feeding defence demand, bordered corridors and producer materials, never the country’s equities (slow layer). The graph only knew the wars the press still covered: an old war fades with its stories (half-life of 7 days), and the war between the United States and Iran, open since February, had left the graph. The conflict index covers fourteen more countries in open conflict (Yemen, Sudan, Syria, Lebanon, Palestine, the Sahel…), which had none. | PR points chauds, audit 23 |
| 2026-10-24 | Oracle | Nothing enters without validation. When the model did not answer, or answered too long a list for its 500-token reply (ten candidates at once), every candidate was recorded unvalidated and silently: the pass of 23 October at 22:45 recorded ten of ten. Candidates are now validated by five, an incomplete answer counts as none, and the articles wait for the next pass. Figurative wars and military exercises are not conflicts. | PR validation obligatoire, audit 22 |
| 2026-10-23 | Oracle | A conflict needs a strong term or a weak term with an armed actor, and the validating model works from a definition of each type and checks the place. The first stories of the corrected detector showed “conflicts” in France, India and the Philippines made of street shootings, a cyberattack, a video-game metaphor and a war mentioned in passing, all validated by the model. | PR précision des conflits, audit 21 |
| 2026-10-23 | OMEGA | The past pace of a story is measured on the comparable series of signals only. The day the detector began to link every article (instead of at most 5 per pass), the last 24 hours were compared with a capped past: Israel-Palestine, Russia, China, “Europe” and “Global” all showed a full “surge” that was a change of unit, not an escalation. | PR #122 bis, audit 20 |
| 2026-10-23 | Oracle, OMEGA | Event detection by whole words and by the place of the action. Before: “war” matched software and trade wars, “strike” matched airstrikes, and the place was the first name of a fixed list that did not include France (60 articles on the French riots, 3 kept, all “Global”) and found “Mali” in Somalia; the language model rewrote the place without seeing any title; a story became a new event every day; the same article was linked up to three times (28,072 duplicate links). | PR Oracle (b), audit 19 |
| 2026-10-23 | OMEGA | Freshness is measured on the whole story, not on its latest Oracle event. The Oracle opens a new event for the same war every day: measured on the latest one, a war of several months looked one day old (the half-life never applied) and its past pace looked nil, so any signal counted as a surge (Ukraine, Russia and Iran all at once). Oracle times, written in local time, were compared with UTC (2 hours too old in summer). | PR O2-correctif 2, audit 18 |
| 2026-10-22 | OMEGA | An event reaches the countries its location names (Iran: Iran), not every country of its theatre: a conflict in Iran oriented the Saudi, Qatari and Kuwaiti markets downwards, and Hormuz all eight Gulf countries. A strait, a sea or a region names no country; a location naming two countries reaches both (only the first was read). | PR O2-correctif, audit 16 |
| 2026-10-22 | OMEGA | The wiring of the graph. Oracle events enter by family (15 geopolitical stories, 5 sensor events): 13 of the 15 events kept were radio anomalies. Events of one zone and type are one source (Gaza and Israel were two). A geopolitical event fades with a half-life of 7 days unless its signals surge (it weighed its confidence for 30 days). Conflicts reach trade corridors through the straits their own zone borders, and corridors the materials of their countries of origin, by signed links: every conflict joined the Logistics Stress Index wherever it took place, and the index joined every material, by association links that carried no direction. An active conflict raises the demand of the defence sector (no such link existed: a war reached defence groups only as consumers of materials, downwards). Weight floors (0.45 to 0.60) and default values (0.3 to 0.6) are removed. One source weighs once per direction on an asset (it weighed once per path). Radiation and space-weather events are shown but are no longer evidence. | PR #119, audit 16 |
| 2026-10-22 | OMEGA | Evidence of the propagation must measure a change: risk counters are signed by their variation and neutral counters are no evidence (the press volume counted as “risk rising” for a 1.1% fall); a radio anomaly weighs the detector’s level times its confidence (13 anomalies all weighed 0.88, whatever their size); the TANKAN counts by its quarterly change around its publication (its level, +24, counted at every sweep). Nodes built on a stale catalogue key are neither sources nor relays (all the graph’s exposure passed through a stale logistics stress index). | PR #117, audit 15 |
| 2026-10-22 | OMEGA | The sanctions metric counts new tightening State measures over 7 days (it read a table that does not exist); healthy sources are counted from the data catalogue (the method it called did not exist). Collectors are rewired to their real sources (air traffic, WebSDR receivers, weak signals); a collector that fails publishes its reason. | PR #114–#116, audit 12–14 |
| 2026-10-22 | OMEGA | Count severity by empirical rank among the metric’s own changes over 7 days. Before: an absolute threshold
of 5 articles on a base of about 3,900 made a 1.1% fall of the press volume “critical” in 98% of the
sweeps; a first correction by robust z-score (thresholds 3/4/5) still made 40% of the sweeps “critical”
for conflict fatalities, a count that moves by steps. The risk level counts only moves toward risk, neutral
counts no longer vote on the direction, and the provisional level is None (pending) instead of a
“STABLE” written by default. A query translation defect that silenced the conflict-events count is
fixed. |
PR #113, #114, audit 11–12 |
| 2026-10-21 | Trade-Bot | Engine v2.3. OMEGA counts by its intelligence direction only: between 15 and 21 October a single refuge-flight alert gave −0.25 to every risk asset and 68 of 81 v2 decisions were SELL. Momentum keeps the best-established persistence, weighted by min(1, t/2): with about 200 sessions of base almost no asset reached t ≥ 2, and momentum had vanished from the engine since 16 October. The VIX rises with stress. | PR #112, audit 10 |
| 2026-10-21 | Prices | Quotes are rounded according to their magnitude: currencies to 4 decimals instead of 2 (EUR/USD stood frozen at 1.12). | PR #111 |
| 2026-10-17 | Pattern engine | Historical market analogues since 1998 and their walk-forward evaluation: no skill shown (Brier skill from −0.004 at 7 days to −0.076 at 90). Kept as descriptive context in the Quant-Lab, outside every calculation. | PR #104, #109, audit 4, 8 |
| 2026-10-17 | Pattern engine | Engine version 2. The first engine read keys that no task writes and fields at the wrong place (most snapshots held no Quant-Lab value); a missing value became a neutral 0.5 and a missing OMEGA “STABLE”, 0 and 0.0; outcomes were measured with the prices of the day of the check (a late check gave the same move at T+7, T+14 and T+30); Oracle forecasts were never read in their published format and a neutral scenario counted as a success; dimensions 19–22 were bounded price levels frozen at 0 or 1; the cosine ignored levels; classification rules fired on absent values read as 0; the Investo-Meter accepted a concordance of 0.50 without the injection gate. First-engine data are kept as legacy, injection suspended. | PR #100–#102, audit 2 |
| 2026-10-17 | OMEGA | The market regime modulation runs: the detector’s regime names are mapped. OMEGA expected the names risk-off, risk-on and transition, which the detector does not publish, so the modulation had never run. | PR #99 |
| 2026-10-17 | Data truth | No fabricated value left in Strategic Flows, strategic materials, the Quant-Lab, the Investo-Meter and OMEGA:
absent inputs are None with their provenance instead of 0, 50 or a default constant. |
PR #98, audit 1 |
| 2026-10-17 | Real Cost Index | France and Germany on Eurostat ECOICOP v2 and the household budget survey; effort rate by income quintile as an interval; the index is absent instead of 100 when no component is measured; the United States effort rate comes from the BLS Consumer Expenditure Survey revalued by CPI-U. | PR #102, #103, #105–#108, audit 3, 5–8 |
| 2026-10-16 | Trade data | Comtrade distinguishes a measured value, a flow with no declared line and a failure (failures were turned into
zeros); calibrated trade alerts. The dossier’s “Comtrade volumes” signal, which had never fired
(it read a key written for other materials and a growth rate nobody computed), reads EU import growth. The global
supply concentration is removed from both engines: the world total counted as the only supplier gave a
fabricated HHI of 1 (dated in the observation archive as supply_hhi). |
PR #97 |
| 2026-10-16 | Trade-Bot | Prices dated in UTC and judged by their session; a stale price never opens, closes or grades a position. Bounded database pool, audit log with a replay queue, public page limited and reduced to a daily curve, blocking security scanner. | PR #95, #96 |
| 2026-10-16 | Trade-Bot | Investo-Meter searches spend seven days in quarantine, then wait for the super-administrator’s validation; configurable starting capital; narrative emergence reads the sector Factor Z (it read a global social index and proposed the producers of the most tense sectors of the scan, not necessarily the diverging ones); cache of the v2 LLM arbitrations. | PR #94 |
| 2026-10-15 | Data catalogue, game theory | Common catalogue of sources (seven keys read without any writer, removed); provenance and coverage of the game theory evidence. Before: an absent AIS transit counted as 0% and gave a 0.5 mass to the status quo; the Hormuz press signal was looked up as “Hormuz” while the bridge publishes “Ormuz”; the logistics stress index was declared available when absent; 17 triggers were constants in the code. Investo-Meter → Trade-Bot contract. Naval posture and home-made container freight index. | PR #90, #91 |
| 2026-10-15 | Trade-Bot | Missing geopolitical inputs carried with their age. The corruption layer reads the Investo-Meter: it had never been filled (its summary named countries in French, the Trade-Bot looked for ISO codes). | PR #89 |
| 2026-10-15 | Investo-Meter | Country risk formula r1 published with effective weights and coverage; an unknown exposure country stays unknown (it was “US” by default). | PR #88 |
| 2026-10-15 | Trade-Bot | Engine v2.2 decisions: OMEGA weighs 25% of the conviction, decisions and stops at the asset’s pace (they were 3/5/8% stops and 3/7/14-day horizons per declared layer), report card protocol (fixed on 14 October). Cost model c1, net results; one sizing rule for both books (main engine v1.3); blocking continuous integration. | PR #87 |
| 2026-10-14 | Trade-Bot | Measured inertia and the organisational model of the financial system, published without changing decisions (layers had been declared). | PR #86 |
| 2026-10-10 | Trade-Bot | Engine v2.1: the LLM arbiter only bets in the direction of the conviction, beyond the action threshold. In the first week, bets opened against it or far below it were closed within minutes by the signal-reversal exit. | PR #82 |
| 2026-10-10 | Investo-Meter | Sanctions counted as regimes (20 points per authority, reference table checked that day); conflict index (ACLED, UCDP/PRIO, WGI). Game theory: China triple leverage scenario. | PR #75, #76, #85 |
| 2026-10-10 | LLM | Remote calls routed by use (Gemma 4 for frequent small calls; the flash-lite models reserved for reports and the Trade-Bot). | PR #77 |
| 2026-10-08 | Weak signals | Measurement foundation: long-term archive of expiring series, robust deviations from a 28-day reference; then (9 October) press measures by theme and the pace of State measures. | PR #55, #57, #70, #71 |
| 2026-10-07 | Collection | Travel advisories from nine issuers, consensus basis kept on the five core issuers; State measures; bilateral GDELT events and the shift detector; UN Security Council resolutions and votes by State. | PR #45–#48, #51 |
| 2026-10-02 | OMEGA, Trade-Bot | The Trade-Bot’s directions come from the OMEGA intelligence engine (omega:directions), with
the fragility base, the currency anchors and the A-share references. The graph loses its development-phase node
and link ceilings, gains parent nodes and one vocabulary for French and English. |
PR #12–#17 |
| 2026-10-02 | Archiviste | Unified search in two stages, Turchin periods, memory of every search; each external source answers in its own thread with its own time limit (a slow source held the others). | PR #18 |
| 2026-10-01 | Trade-Bot | Directional engine v2 in shadow mode on a separate book; reliable measurement and input hygiene of the main engine. | PR #9, #10 |
| before 2026-09-28 | Oracle, orchestration | Shafer discounting replaces the replication of masses in the Oracle fusion, which crushed ignorance exponentially. Task deduplication locks adopted after an incident of more than 4,600 queued tasks. | first version of the repository |