review

The point is not to record what happened. It is to find the thing that would have put today's winners on this morning's list — and to check whether the screen beat the pool it drew from. Output is DATA: funnel rows and, rarely, a FINDINGS entry. Lever opinions are forbidden here — one day's winners are noise; proposals go through the ledger with pre-registered hypotheses and operator sign-off.

All data via the gammarips MCP only.

1. Outcomes on every pool name, not just yours

1. One batched get_liquidity returns each pool contract's day bar. Rows whose day.last_updated is not today are excluded AND counted — never dropped silently.
2. For names that traded meaningfully, replay_contract (minute, today): its anchor (close of the first bar at/after 10:00 ET) is the honest entry reference. Compute post-anchor MFE%, MAE%, close%, and target/stop touches from the bars — never from the day bar alone (the day high can predate 10:00). Honesty caveats carried verbatim: delayed tape, session_complete, thin-tape (no bar = no prints that minute; touches are evidence, not ticks).
3. Take session volume from get_liquidity's day_volume, never by summing bars (bars can duplicate on some contracts).
4. Sub-$0.50 contracts get flagged — a lowball print can fake +200%; never let one top the table unflagged. Contracts with a handful of prints all session get status: unlabeled_illiquid — there is no honest anchor for a tape that never printed, and they are never counted as missed winners.
5. Upsert outcome_session onto every row of today's funnel via python scripts/funnel_log.py upsert. A log of only your own picks cannot answer the question this skill exists for — the regret data lives in the passed-over rows.

2. Did the screen earn its keep today?

Rank the pool by the 10:00-anchored move, then answer plainly:

• Did the ranked list contain today's winners? If not, what separated them from what was brought forward?
• Did the tradeability screen throw away a winner? Check every dropped name. If a "winner" won on a tape of single-lot prints, the screen was right and the winner was never enterable — say that explicitly instead of counting it as a miss. Record both sides: winners the gates dropped AND losers they avoided.

Then run python scripts/edge.py — it asks the same question across every accumulated day and reports when N is too small to conclude anything. Believe it over the day's narrative, and read its day-demeaned column before its pooled one.

3. Backfill — T+3 is where the money is

Winners peak overwhelmingly AFTER the entry session ([[three-day-harvest-curve]], [[path-calibrated-giveback]]), so outcome_3d is the column that matches a real hold.

• Cheap path: query_outcomes(view="surface", scan_date=...) — it fills on a lag (the window's last session must be strictly before today, so the newest filled scan_date is ~4 trading days back; WINDOW_OPEN inside that frontier is normal, not a fault).
• Fallback when the surface can't serve a day: replay_contract(granularity="day", from_date=entry_day, to_date=entry_day+2 sessions) — ~3 bars per contract, a whole pool is ~50 light calls. Compute 3-day MFE/MAE against the anchor already stored on outcome_session, never against the day open. Bar highs/lows are touch ceilings, not fills.
• Rows the surface can't serve and whose anchor is missing get status: unlabeled; sub-floor tapes get unlabeled_illiquid. Both tails are non-random — keep them visible, never fake-score them.

Then python scripts/skill_eval.py — the per-rule scorecard: does each exclusion_reason earn its keep, weighting losers-avoided equally with winners-dropped? Proposals go to the operator through your skill ledger; this skill never edits a rule.

4. Findings

Append to eval/FINDINGS.md only when the accumulated evidence moves, or when a process failure cost a real decision. Each entry: date, observation with numbers, what it implies for /trade, and the current N. Most days add nothing, and that is the correct output.

Never

• Change, propose inline, or "just this once" adjust any lever or doctrine rule.
• Score against a fixed bracket as though it were your exit ([[fixed-exit-composites-negative]]).
• Count a winner that traded a handful of contracts as an opportunity missed.
• Fetch any non-MCP market data or external pick.