Question: Could our own engine have derived the best calls our Twitter/Discord follows made, from our own data, at or before the date they published? 10 diversity-selected cases, graded LEAD / MATCH / PARTIAL / MISS. Verdict: assign-follow-up — the tape and fundamentals halves of the calls were consistently in our own data at or before publish; the misses cluster in four nameable gaps, each now a filed task.

Method

As-of tape replay: for each case, per-symbol change7d/30d/3m, acceleration (7d pace vs 30d weekly pace, the insight-wire formula), distance from trailing-252d high, and relative volume were computed at the trading day before the source's publish date, from research/market-engine/data/stocks/<SYM>/ohlc/12m.json (252 daily bars, close/volume/sma20/sma50). Basket relative strength = basket median 30d minus SPY 30d from the same files. Detector verdicts apply the live wire thresholds (accel > 3, quiet-turn week >= 5 on a negative month, fresh-high within 3% of the 252d high, strength >= 15%/30d). Caveat named honestly: this proves derivability from data we carry, not that the wires ran on those dates — the wires were built 2026-07-02/03.

Scorecard

Totals: 3 LEAD-class, 4 PARTIAL, 3 MISS (2 by construction).

What the pattern says

1. The tape half of nearly every follow call was in our data first. Cases 1, 7, 8: our detectors (as now built) would have fired before the tweet — including the single most "found" call in the set (HLIT: a 6x-volume breakout at a fresh high with +10.5 acceleration is a textbook wire hit). The follows' edge on these was surfacing, not information — consistent with the standing front-run finding.
2. The story half is agent-read tier, and that's fine. Segment-mix re-rates (HLIT), founder positioning (OUST), layer books (crux) need a narrative read of filings/transcripts — this is exactly what the new deep-read queue routes: the wire flags the symbol, the agent does the read.
3. The real gaps are nameable and small:
   • Coverage/discovery (case 2): OUST screamed for 3 weeks before we covered it. Names outside watchlists are invisible to the wires; monster-discover exists but its cadence and hand-off didn't catch this.
   • Mega-cap fundamentals freshness + one missing detector (case 3): hyperscaler statements were stale on disk, and no detector watches FCF-collapse / capex-as-%-of-OCF — the Apollo claim was free public data.
   • Structural layer maps (cases 4, 9, 10): research-trace machinery exists but runs ad-hoc; the misses were un-run traces, not missing tooling.
   • Macro series (cases 5, 6): EIA/IRENA-class data has no intake path; sized as its own feasibility-gated decision.

Follow-ups filed (see ops/tasks/TASKS-ENGINE.md + research/tasks/)

• capex-ratio / FCF-trend red-flag detector for fundamentals-wire (#cross-wire lane)
• mega-cap statement freshness in the massive refresh path
• discovery hand-off hardening: monster-discover hits above an RS bar should open coverage rows
• research-trace refresh cadence for active perspective layers
• macro-series intake (EIA first) — feasibility-gated, decision row not build row