epic(forecast): generator signal integrity & correctness — dead signals, market-calibration polarity, silent no-fire

Why this is its own epic (upstream of #4930)

#4930 built the verification loop (resolver + Brier scorecard) and it is now live and accumulating. But a full-code + production sweep on 2026-07-09 found the generator feeding it has real garbage-in: two entire signal layers are silently dead in prod, a market-calibration bug pushes probabilities the wrong direction, and a systemic silent-no-fire pattern hid all of it. Verification is not moot — a bad Brier will expose these — but the track record is being built on broken inputs right now, so fixing the generator is the top priority and is time-sensitive (every hour of forecasts accumulates on the broken inputs).

This epic tracks the generator-side fixes; #4930 tracks scoring them.

Sweep method

4-agent adversarial audit of scripts/seed-forecasts.mjs (signal wiring → writers · detector/blend correctness · missing-signals gap · MiroFish sim completeness+accuracy) + live prod grounding (feed freshness, key existence, stored-shape checks). Every finding below was cross-checked against code and prod.

Findings → sub-issues

🔴 P1 — silent dead signals & directional bugs (drop-everything):

• fix(forecast): FRED macro/rates signals silently dead — shape mismatch kills all 10 (P1, #5097) #5098 — FRED macro/rates family dead — 10 signals, shape mismatch
• fix(forecast): ACLED conflict signal not produced in prod + health blindspot — conflict EMA blind to near-real-time (P1, #5097) #5099 — ACLED conflict signal not produced in prod + health blindspot
• fix(forecast): market calibration — outcome-polarity (ceasefire inflates escalation) + no liquidity floor + cap bypass (P1, #5097) #5100 — Market calibration: outcome-polarity + liquidity-floor + cap-bypass

🟠 P2 — consistency / calibration / latent + observability:

• feat(forecast): signal observability — kill silent feed no-fire (per-feed logging + health coverage + temporal keep-warm) (P2, #5097) #5101 — Forecast signal observability — silent-no-fire logging + health coverage + temporal:anomalies keep-warm
• fix(forecast): MiroFish sim — demotion-gate threshold + completeness gaps (geo-dedup, archetypes, round-2 bias, timingMarkers) (P2, #5097) #5102 — MiroFish correctness + completeness
• fix(forecast): detector calibration — state-derived cap escape, UCDP gate under-forecast, projection de-anchoring (P2, #5097) #5103 — Detector calibration — state-derived cap, UCDP gate, projection de-anchoring

De-risking (verified during the sweep)

• MiroFish does NOT rewrite published probabilities — decorations are display + deep-path acceptance + next-run seeding only (the sole pred.probability= write is the market blend). Sim bugs can't directly corrupt Brier → lower urgency than the P1s.
• Prior hardening intact — fix(forecast): pre-#4930 hardening — substring country-match probability corruption, NaN guards, forecast-id churn, sim outcome/lock CAS (adversarial audit) #4933/fix(forecast): non-finite probability guards — NaN can no longer enter or anchor the probability spine (#4933) #4969/fix(forecast): stable semantic-slot ids for military and state-derived forecasts (#4933) #4973 (word-boundary matching, NaN guards, stable ids) all still in place, no regressions.
• Key-string wiring is clean — every input key has an exact-match writer; the dead signals are a shape mismatch (FRED) and an operational gap (ACLED), not typos.
• iran-events is correctly flag-gated off (iranEventsEnabled() default false) — the 33-day-stale feed is not consumed.

KPI

• Every declared signal-type actually fires (no silently-empty input family reaching a forecast).
• Zero silent feed no-fire — each forecast input has presence logging + health coverage.
• Known directional / cap / polarity bugs fixed → the epic(forecast): verified probabilistic forecasting — resolution engine, public Brier track record, learned calibration, LLM ensemble (MiroFish++ roadmap) #4930 Brier reflects genuine model skill, not wiring artifacts.

Related

• epic(forecast): verified probabilistic forecasting — resolution engine, public Brier track record, learned calibration, LLM ensemble (MiroFish++ roadmap) #4930 (verification epic this unblocks) · fix(forecast): resolvability coverage — reach the ≥80%-hard KPI (57% live; #5066 follow-up, Lane A #2) #5091 (Lane-A resolvability coverage) · Fix forecast hard resolution coverage #5072 / feat(forecast): add resolution ledger and scorecard #5024 (resolution engine)