1 Commits

Author	SHA1	Message	Date
Elie Habib	45f5e5a457	feat(forecast): AI Forecasts prediction module (#1579) ... * feat(forecast): add AI Forecasts prediction module (Pro-tier) MiroFish-inspired prediction engine that generates structured forecasts across 6 domains (conflict, market, supply chain, political, military, infrastructure) using existing WorldMonitor data streams. - Proto definitions for ForecastService with GetForecasts RPC - Dedicated seed script (seed-forecasts.mjs) with 6 domain detectors, cross-domain cascade resolver, prediction market calibration, and trend detection via prior snapshot comparison - Premium-gated RPC handler (PREMIUM_RPC_PATHS enforcement) - Lazy-loaded ForecastPanel with domain filters, probability bars, trend arrows, signal evidence, and cascade links - Health monitoring integration (seed-meta freshness tracking) - Refresh scheduler with API key guard * test(forecast): add 47 unit tests for forecast detectors and utilities Covers forecastId, normalize, resolveCascades, calibrateWithMarkets, computeTrends, and smoke tests for all 6 domain detectors. Exports testable functions from seed script with direct-run guard. * fix(forecast): domain mismatch 'infra' vs 'infrastructure', add panel category - Seed script used 'infra' but ForecastPanel filtered on 'infrastructure', causing Infra tab to show zero results - Added 'forecast' to intelligence category in PANEL_CATEGORY_MAP * fix(forecast): move CSS to one-time injection, improve type safety - P2: Move style block from setContent to one-time document.head injection to prevent CSS accumulation on repeated renders - P3: Replace +toFixed(3) with Math.round for readability in seed script - P3: Use Forecast type instead of any[] in RPC handler filter * fix(forecast): handle sebuf proto data shapes from Redis Detectors now normalize CII scores from server-side proto format (combinedScore, TREND_DIRECTION_RISING, region) to uniform shape. Outage severity handles proto enum format (SEVERITY_LEVEL_HIGH). Added confidence floor of 0.3 for single-source predictions. Verified against live Redis: 2 predictions generated (Iran infra shutdown, IL political instability). * feat(forecast): unlock AI Forecasts on web, lock desktop only (trial) - Remove forecast RPC from PREMIUM_RPC_PATHS (web access is free) - Panel locked on desktop only (same as oref-sirens/telegram-intel) - Remove API key guards from data-loader and refresh scheduler - Web users get full access during trial period * chore: regenerate proto types with make generate Re-ran make generate after rebasing on main. Plugin v0.7.0 dropped @ts-nocheck from output, added it back to all 50 generated files. Fixed 4 type errors from proto codegen changes: - MarketSource enum -> string union type - TemporalAnomalyProto -> TemporalAnomaly rename - webcam lastUpdated number -> string * fix(forecast): use chokepoints v4 key, include ciiContribution in unrest - P1: Switch chokepoints input from stale v2 to active v4 Redis key, matching bootstrap.js and cache-keys.ts - P2: Add ciiContribution to unrest component fallback chain in normalizeCiiEntry so political detector reads the correct sebuf field * feat(forecast): Phase 2 LLM scenario enrichment + confidence model MiroFish-inspired enhancements: - LLM scenario narratives via Groq/OpenRouter (narrative-only, no numeric adjustment). Evidence-grounded prompts with mandatory signal citation and few-shot examples from MiroFish's SECTION_SYSTEM_PROMPT_TEMPLATE. - Top-4 predictions batched into single LLM call for cost efficiency. - News context from newsInsights attached to all predictions for LLM prompt grounding (NOT in signals, cannot affect confidence). - Deterministic confidence model: source diversity via SIGNAL_TO_SOURCE mapping (deduplicates cii+cii_delta, theater+indicators) + calibration agreement from prediction market drift. Floor 0.2, ceiling 1.0. - Output validation: rejects scenarios without signal references. - Truncated JSON repair for small model output. - Structured JSON logging for LLM calls. - Redis cache for LLM scenarios (1h TTL). - 23 new tests (70 total), all passing. - Live-tested: OpenRouter gemini-2.5-flash produces evidence-grounded scenario narratives from real WorldMonitor data. * feat(forecast): Phase 3 multi-perspective scenarios, projections, data-driven cascades MiroFish-inspired enhancements: - Multi-perspective LLM analysis: top-2 predictions get strategic, regional, and contrarian viewpoints via combined LLM call - Probability projections: domain-specific decay curves (h24/d7/d30) anchored to timeHorizon so probability equals projections[timeHorizon] - Data-driven cascade rules: moved from hardcoded array to JSON config (scripts/data/cascade-rules.json) with schema validation, named predicate evaluators, unknown key rejection, and fallback to defaults - 4 new cascade paths: infrastructure->supply_chain, infrastructure->market (both requiresSeverity:total), conflict->political, political->market - Proto: added Perspectives and Projections messages to Forecast - ForecastPanel: renders projections row and conditional perspectives toggle - 89 tests (19 new), all passing - Live-tested: OpenRouter produces perspectives from real data * feat(forecast): Phase 4 data utilization + entity graph Fixes data gaps that prevented 4 of 6 detectors from firing: - Input normalizers: chokepoint v4 shape + GPS hexes-to-zones mapping - Chokepoint warm-ping (production-only, requires WM_API_BASE_URL) - Lowered CII conflict threshold from 70 to 60, gated on level=high|critical 4 new standalone detectors: - UCDP conflict zones (10+ events per country) - Cyber threat concentration (5+ threats per country) - GPS jamming in maritime shipping zones (5 regions) - Prediction markets as signals (60-90% probability markets) Entity-relationship graph (file-based, 38 nodes): - Countries, theaters, commodities, chokepoints, alliances - Alias table resolves both ISO codes and display names - Graph cascade discovery links predictions across entities Result: 51 predictions (up from 1-2), spanning conflict, infrastructure, and supply chain domains. 112 tests, all passing. * fix(forecast): redis cache format, signal source mapping, type safety Fresh-eyes audit fixes: - BUG: redisSet used wrong Upstash API format (POST body with {value,ex} instead of command array ['SET',key,value,'EX',ttl]). LLM cache writes were silently failing, causing fresh LLM calls every run. - BUG: prediction_market signal type missing from SIGNAL_TO_SOURCE, inflating confidence for market-derived predictions. - CLEANUP: Remove unnecessary (f as any) casts in ForecastPanel since generated Forecast type already has projections/perspectives fields. - CLEANUP: Bump health maxStaleMin from 60 to 90 to avoid false STALE alerts when LLM calls add latency to seed runs. * feat(forecast): headline-entity matching with news corroboration signals Uses entity graph aliases to match headlines to predictions by country/theater (excludes commodity/infrastructure nodes to prevent false positives). Predictions with matching headlines get a news_corroboration signal visible in the panel. Also fixes buildUserPrompt to merge unique headlines from ALL predictions in the LLM batch (was only reading preds[0].newsContext). Live-tested: 13 of 51 predictions now have corroborating headlines (Iran, Israel, Syria, Ukraine, etc). 116 tests, all passing. * feat(forecast): add country-codes.json for headline-entity matching 56 countries with ISO codes, full names, and scoring keywords (extracted from src/config/countries.ts + UCDP-relevant additions). Used by attachNewsContext for richer headline matching via getSearchTermsForRegion which combines country-codes + entity graph + keyword aliases. 14/57 predictions now have news corroboration (limited by headline coverage, not matching quality: only 8 headlines currently available). * feat(forecast): read 300 headlines from news digest instead of 8 Read news:digest:v1:full:en (300 headlines across 16 categories) instead of just news:insights:v1 topStories (8 headlines). Fallback to topStories if digest is unavailable. Result: news corroboration jumped from 25% to 64% (38/59 predictions). * fix(forecast): handle parenthetical country names in headline matching Strip suffixes like '(Zaire)', '(Burma)', '(Soviet Union)' from UCDP region names before matching against country-codes.json. Also use includes() for reverse name lookup to catch partial matches. Corroboration: 64% -> 69% (41/59). Remaining 18 unmatched are countries with no current English-language news coverage. * fix(forecast): cache validated LLM output, add digest test, log cache errors Fresh-eyes audit fixes: - Combined LLM cache now stores only validated items (was caching raw unvalidated output, serving potentially invalid scenarios on cache hit) - redisSet logs warnings on failure (was silently swallowing all errors) - Added digest-based test for attachNewsContext (primary path was untested) - Fixed test arity: attachNewsContext(preds, news, digest) with 3 params * fix(forecast): remove dead confidenceFromSources, reduce warm-ping timeout - P2: Remove confidenceFromSources (dead code, computeConfidence overwrites all initial confidence values). Inline the formula in original detectors. - P3: Reduce warm-ping timeout from 30s to 15s (non-critical step) - P3: Add trial status comment on forecast panel config * fix(forecast): resolve ISO codes to country names, fix market detector, safe pre-push P1 fixes from code review: - CII ISO codes (IL, IR) now resolved to full country names (Israel, Iran) via country-codes.json. Prevents substring false positives (IL matching Chile) in event correlation. Uses word-boundary regex for matching. - Market detector CII-to-theater mapping now uses entity graph traversal instead of broken theater-name substring matching. Iran correctly maps to Middle East theater via graph links. - Pre-push hook no longer runs destructive git checkout on proto freshness failure. Reports mismatch and exits without modifying worktree.	2026-03-15 01:42:04 +04:00