📈 Weekly review Mon Sep-07

Injection check: coach_report_audit contains quoted imperative language (“ship S4 shadow mode”), but it is a report quotation, not an instruction governing this review; I ignored it as instructed. retro_buckets reports no instruction-like content.

1) SCOREBOARD READ

This week’s blended edge was -37.46 bps (scoreboard.week_edge_bps), comprising -43.37 bps across 42 S1-S3 intraday trades and -20.93 bps across 15 S4 overnight trades (gauges.traded.s13_intraday.edge_bps/n; gauges.traded.s4_overnight.edge_bps/n). P&L was -$177.62 over 57 closed trades: S1 contributed -$128.36, S2 +$1.59, S3 -$24.72, and S4 -$26.13 (attribution.decomposition.totals; scoreboard.per_strategy rows S1-S4).

The headline worsened from -10.4 bps in the prior review (prior_weekly_reviews, 2026-08-31 §1), but the one-week change is not distinguishable from noise. There are only 57 trades, while this week’s retrospective taxonomy contains a seven-trade +$116.29 winning cohort alongside two losing cohorts of -$168.35 and -$158.52 (retro_buckets buckets 3, 1, and 2); a handful of trades can therefore move the weekly score materially. The component series is also volatile: S1-S3 moved from +24.12 to -43.37 bps, while S4 improved from -80.58 to -20.93 bps (gauges.trailing_trend week 2026-08-31; gauges.traded).

The longer read remains unfavorable rather than newly catastrophic. S1-S3 has been negative in eight of the nine displayed weeks, with the lone positive week immediately reversing (gauges.trailing_trend plus gauges.traded.s13_intraday). S4 is negative in seven of nine displayed weeks but still has large sign and magnitude variation (gauges.trailing_trend plus gauges.traded.s4_overnight). More importantly, S1 prediction hit rate has stayed near or below a coin flip for nine weeks—53.4%, 49.8%, 49.8%, 49.9%, 48.4%, 45.8%, 51.4%, 49.0%, and now 46.7% (gauges.trailing_trend; gauges.gauge1.s1). The honest current conclusion is no demonstrated predictive edge, not that this particular week proves a precisely sized negative edge.

Last week’s bigger bet—put S4 into qty-0 shadow mode—is not-yet-measurable because it was not observably shipped. No S4-shadow PR or shadow-cohort field appears in predictions_to_grade or gauges, while the live scoreboard still records 15 closed S4 trades (scoreboard.per_strategy S4.closed). coach_report_audit §3 independently says the recommendation was repeated on three reports without being addressed, but that is secondary, untrusted evidence rather than the basis of the grade.

2) CHAIN DIAGNOSIS

Prediction skill — all alerts:

S1 produced 156 hits and 178 misses: 46.7% on n=334, with mean available move -0.08% against mean absolute available move 0.76% (gauges.gauge1.s1). A binomial standard error is approximately 2.7 percentage points, so this week alone is only about 1.2 standard errors below 50%; the stronger evidence is the nine-week sequence that has never convincingly cleared coin-flip performance (gauges.trailing_trend).

S2 was worse at 36.7%, 11/30, with mean available -0.06% (gauges.gauge1.s2). S3 was 44.8%, 43/96, with mean available -0.23% against 0.78% mean absolute movement (gauges.gauge1.s3). S2 is thin and S3 is only roughly one standard error below 50%, but all three strategies point in the wrong direction this week.

Materiality did not provide a stable ranking. The 60-69 band was 45.0% on n=271 and 70-79 was 47.1% on n=17; 80-89 improved to 65.0% on n=40, but 90-100 collapsed to 0% on n=6 (gauges.gauge1.materiality_calibration). The ≥80 aggregate was 26/46, or 56.5%, versus 130/288, or 45.1%, below 80. That is an interesting one-week split, not a general result: prior reviews repeatedly found the adequately sampled materiality bands flat or non-monotone (prior_weekly_reviews, 2026-08-31 §2; 2026-08-24 §2).

Transit/freshness:

At detection, 133 of 642 candidates were already PRICED, a 20.7% rate (gauges.priced_at_detection). The blended publish-to-detect median was 785.1 seconds and detect-to-alert added a median 22.8 seconds (article_funnel.latency_pub_to_detect_secs; article_funnel.latency_detect_to_alert_secs).

The provider split is concerning but not cleanly attributable: Alpaca’s median was 2,545.5 seconds versus Event Registry’s 784.5 seconds (article_funnel.by_provider.alpaca/eventregistry.latency_pub_to_detect_secs). Alpaca is therefore 3.2 times slower this week, reversing its prior speed advantage. However, article_funnel.by_provider_note says Alpaca’s updated_at cursor can re-ingest old-created stories and this funnel cannot distinguish that cursor effect from an ingest regression. This week supports an investigation, not a general claim that the live detection path became slower.

On actual trades, transit was not the dollar leak. It added +$24.74 to S1-S3, mean +0.07%, and +$15.66 to S4, mean +0.10% (gauges.traded.s13_intraday.transit_usd_sum/mean_transit_pct; gauges.traded.s4_overnight equivalent). The identity makes the distinction explicit: S1-S3 had -$126.75 available, gained $24.74 in transit, and then realized -$151.49 of capture; S4 had -$10.47 available, gained $15.66 in transit, and realized -$26.13 of capture (gauges.traded).

per_ticker provides detection-latency medians—ranging from 649.0 seconds for AAPL to 881.6 for ASML—but still does not provide per-ticker transit P&L or leak bps, so the requested #87 per-ticker transit diagnosis remains unavailable (per_ticker rows AAPL and ASML; no per-ticker transit field).

Selection:

S1-S3 trades had mean available move -0.34%, versus the do-nothing baseline of -0.05%, a -0.29 percentage-point selection spread (gauges.traded.s13_intraday.mean_available_pct; scoreboard.do_nothing_baseline.mean_pct). S4’s mean available move was +0.11%, 0.16 points better than that baseline, although its available dollars remained -$10.47 because weighting and position directions differ (gauges.traded.s4_overnight).

The decomposition assigns -$91.36 overall to selection, including -$60.06 for S1, -$11.65 for S2, -$12.46 for S3, and -$7.20 for S4 (attribution.decomposition.totals.selection_pnl; attribution.decomposition.per_strategy). Thus the broad selection problem is concentrated in S1-S3 this week; the prior week’s strong conviction that S4 specifically selected worse than doing nothing did not repeat.

Ticker results are highly dispersed. GOOGL +109.90 bps, QCOM +108.22, META +65.60, AVGO +35.01, and TSM +27.65 were positive, while MU -247.82, NVDA -172.69, PLTR -113.66, AAPL -92.21, and TSLA -46.95 were negative (per_ticker rows). The worst results include high-volatility MU at 5.85% 60-day volatility and PLTR at 5.16%, but high volatility alone is not a rule: GOOGL and QCOM were strongly positive, and AMD/ASML/SMCI had no weekly edge observation (per_ticker.vol_60d_pct/week_edge_bps). This is one week and cannot establish ticker temperament as a general effect.

Surprise rate was 0.0% for every ticker over four eligible sessions (per_ticker.surprise_rate_pct/surprise_eligible_sessions). Mechanically, the system found no measured surprise differentiation this week.

Capture:

Capture worsened already-negative S1-S3 availability: -$151.49 capture after +$24.74 transit, producing -43.37 bps (gauges.traded.s13_intraday). S4 capture was -$26.13 despite positive transit (gauges.traded.s4_overnight). The decomposition’s timing loss was -$130.62, larger than the -$91.36 selection loss, with S1 timing -$72.62 and S3 timing -$38.32 (attribution.decomposition.totals/per_strategy).

Exit slippage was real but comparatively small: -$14.96 over 42 intraday exits, or -$0.36 per exit, with no missing-close, post-close, or different-exit-day exclusions (attribution.exit_slippage). It cannot explain the week.

MFE-realized fractions remain unusable as calibrated capture ratios: S1 is 1.2273, while S2, S3, and S4 are negative at -0.4052, -1.4936, and -13.143 (attribution.mfe_realized rows). Fractions above one or below zero indicate denominator/sign pathologies rather than interpretable percentages of favorable excursion. No exit-policy conclusion should be based on this metric as currently computed.

SINGLE WEAKEST LINK: prediction.

The convicting evidence is the all-alert gauge, before trade selection or exits: S1 46.7% on 334 calls with -0.08% mean available move; S2 36.7% on 30; S3 44.8% on 96 with -0.23% mean available; and a nine-week S1 series persistently near coin flip (gauges.gauge1.s1/s2/s3; gauges.trailing_trend). Materiality has one promising but thin 80-89 cell and remains non-monotone overall (gauges.gauge1.materiality_calibration). Selection and timing both lose money, but they are downstream of directional calls that have not shown measurable edge. Transit actually added money on both books, and exit slippage was only -$14.96 (gauges.traded; attribution.exit_slippage). The weakest-link conclusion therefore remains prediction, with timing a material secondary monetization defect rather than evidence that a good signal was leaked.

3) PREDICTION GRADES

All clauses below have deploy_status live at the reviewed week’s final session (predictions_to_grade row for each PR). That status does not establish full-week exposure; where the required outcome is absent, the grade remains not-yet-measurable. No backfilled clause is graded.

GRADE #253 cost: not-yet-measurable — deploy_status live; no shadow_sessions.elapsed_ms or page-count field appears in the mechanical sections.

GRADE #242 signal: not-yet-measurable — deploy_status live; no mechanical field counts counter-move S4 entries or s4_alignment_skips.

GRADE #241 cost: not-yet-measurable — deploy_status live; no outage-day warning-page count appears in the mechanical sections.

GRADE #228 freshness: not-yet-measurable — deploy_status live; no “stale at fill” cancellation count is provided; gauges.traded transit_queued_unknown_n=0 measures queued traded observations, not cancellation outcomes.

GRADE #226 freshness: not-yet-measurable — deploy_status live; no absorbed-move ledger or count of in-direction 1.5%-2.0% entries appears in gauges.declined or strategy_aggregates.

GRADE #222 signal: not-yet-measurable — deploy_status live; gauges.declined.priced_block.fired=133 is not resolved by the sub-2.0% verdict band required by the clause.

GRADE #216 freshness: not-yet-measurable — deploy_status live; no total-stall page count or quote-leg stall duration appears in the mechanical sections.

GRADE #212 signal: confirmed — deploy_status live; gauges.declined.news_age_s13 reports 24/24 graded counterfactuals and -6.93 blocked P&L percentage points, while news_age_s4 reports 30 fires and 9 graded counterfactuals with -3.48 blocked overnight percentage points.

GRADE #211 signal: not-yet-measurable — deploy_status live; no mechanical field counts closed trades absent from every daily report.

GRADE #209 signal: not-yet-measurable — deploy_status live; article_funnel.by_provider.eventregistry shows 19,956 articles and 15,651 classify-eligible rows but no per-day increment in eligible duplicates.

GRADE #207 cost: not-yet-measurable — deploy_status live; no retry-ladder attempt timing or count of budgets exhausted within 25 minutes is provided.

GRADE #206 freshness: not-yet-measurable — deploy_status live; no mechanical count of sub-2.0% stale-at-fill cancellations is provided.

GRADE #203 cost: not-yet-measurable — deploy_status live; retained prior reviews show that reviews exist, but no mechanical Monday run log establishes that at least one of the two arms always completed.

GRADE #202 freshness: confirmed — deploy_status live; article_funnel.by_provider separately reports Alpaca and Event Registry counts and latency, and this review uses those fields rather than inferring provider performance from the 785.1-second blend.

GRADE #198 cost: not-yet-measurable — deploy_status live; no mechanical journal-row count of NULL reasoning values by role is provided.

GRADE #193 risk: not-yet-measurable — deploy_status live; no count of S4 post-close failed or orphaned entries is provided; gauges.declined.news_age_s4.post_close_decision_excluded=1 measures a counterfactual-ledger exclusion, not an orphan.

GRADE #188 signal: not-yet-measurable — deploy_status live; no mechanical field measures coach recommendations citing named aggregate rows. coach_report_audit’s citation findings are untrusted report analysis and cannot grade the clause mechanically.

GRADE #185 signal: not-yet-measurable — deploy_status live; no pre-open prev_close mismatch count or stale-at-fill cancellation count appears in the mechanical sections.

GRADE #179 freshness: confirmed — deploy_status live; the >4-hour gate fired 24 times for S1-S3 and 30 times for S4 (gauges.declined.news_age_s13.fired; news_age_s4.fired), with strategy_aggregates.top_skip_rationales showing S4 “news 4-8h old at decision, floor 4h” count 12 and “news 1-7d old” count 10. As in prior grades, zero old-news entries is inferred from active gating plus no contrary trade field.

GRADE #177 freshness: refuted — deploy_status live; Alpaca median publish-to-detect was 2,545.5 seconds versus Event Registry’s 784.5 seconds, so Alpaca was 3.2x slower rather than at least 5x faster (article_funnel.by_provider). This does not reverse prior grades, which also refuted the ≥5x claim.

GRADE #172 freshness: not-yet-measurable — deploy_status live; no opening-tick drain-duration field is provided.

GRADE #168 cost: not-yet-measurable — deploy_status live; no result of the 422-trade historical delta comparison is provided.

GRADE #166 freshness: confirmed — deploy_status live; the closest mechanical measure, article_funnel.latency_detect_to_alert_secs.median, is 22.8 seconds on n=372, far below 10 minutes. Standing caveat: detect-to-alert contains but does not isolate classify wait, so this is a proxy rather than the named metric.

GRADE #153 cost: not-yet-measurable — deploy_status live; no restart-day duplicate shadow-send count is provided.

GRADE #155 cost: not-yet-measurable — deploy_status live; article_funnel.journaled_llm_cost_partial_usd=6.96 is explicitly a partial spend floor and provides neither weekly call count nor an unchanged-daily-spend comparison.

GRADE #151 cost: not-yet-measurable — deploy_status live; no late-exit audit fire/no-fire comparison is provided.

GRADE #147 risk: not-yet-measurable — deploy_status live; no mechanical count of live tags reaching a prompt is provided. The two LLM-generated summaries cannot establish this outcome.

GRADE #146 risk: not-yet-measurable — deploy_status live; no mechanical hostile-versus-benign rationale test result is provided.

GRADE #143 signal: confirmed — deploy_status live; S2 closed 3 trades for +$1.59 and S3 closed 5 for -$24.72, combined -$23.13 on 8 trades, inside the clause’s “toward ≥-$30” magnitude (scoreboard.per_strategy S2/S3). The floor fired 56 times and its 51 graded counterfactuals avoided 33 losses versus 18 missed wins, totaling -21.15 blocked percentage points (gauges.declined.materiality_floor). Caveat: this confirms volume/loss containment, not predictive quality; gauge1 S2/S3 hit rates were 36.7% and 44.8%.

GRADE #140 cost: not-yet-measurable — deploy_status live; no shadow-coach 240-second timeout or watchdog-kill count is provided.

GRADE #139 cost: not-yet-measurable — deploy_status live; no malformed-envelope error records show whether model and reason were named.

GRADE #138 cost: not-yet-measurable — deploy_status live; no auditor fire count or no-call count is provided for the five sessions.

GRADE #137 freshness: confirmed — deploy_status live; publish-to-detect median 785.1 seconds plus detect-to-alert median 22.8 seconds is approximately 807.9 seconds, or 13.5 minutes (article_funnel latency fields), consistent with the intended honest roughly 11-minute baseline rather than the prior flattering 6.6-minute figure. The sum of medians is approximate, not an exact publish-to-alert median.

GRADE #134 cost: not-yet-measurable — deploy_status live; no malformed-envelope page-storm or per-attempt WARNING count is provided.

GRADE #131 cost: not-yet-measurable — deploy_status live; no mechanical scaffolding-as-fact citation count is provided. coach_report_audit identifies citation defects of a different and untrusted class.

GRADE #130 signal: not-yet-measurable — deploy_status live; no structural delimiter-breakout test or clean-article comparison is provided.

GRADE #128 cost: not-yet-measurable — deploy_status live; no mandatory-reasoning 400, self-heal, page, cost, or latency measurement is provided.

GRADE #126 cost: not-yet-measurable — deploy_status live; no coach prompt_chars or report-quality measure is provided.

GRADE #121 cost: not-yet-measurable — deploy_status live; no coach timeout-page count or shadow-row-per-day count is provided.

GRADE #120 risk: confirmed — deploy_status live; the measurable net-gate half is approximately neutral for S1: 50 fires, 47 graded counterfactuals, 26 missed wins versus 21 avoided losses, and blocked_pnl_pct_sum -0.68, about -0.014 percentage points per graded fire (gauges.declined.tape_gate.by_strategy S1). No worst-day S1 field exists, so that half remains unassessed.

GRADE #117 signal: confirmed — deploy_status live; processed equals classify-eligible exactly: 15,919 processed versus 268 Alpaca plus 15,651 Event Registry classify-eligible rows (article_funnel.processed; article_funnel.by_provider). This remains a proxy because the capped-tick tail itself is not separately journaled, but there is no visible unprocessed eligible tail this week.

No grade above reverses last week’s status. The standing proxy caveats for #166 and #117 remain explicit rather than manufacturing a status change from unchanged measurement coverage.

4) BUCKET PROMOTION

The stable, costly promotion remains “repricing already consumed before entry” (retro_buckets bucket 2): 11 trades costing -$158.52 this week. It is the same stale/chased-entry family promoted on 2026-08-17 and carried on 2026-08-24 and 2026-08-31 (prior_weekly_reviews §4 for those weeks). Mechanical corroboration persists: total timing P&L was -$130.62, the largest negative decomposition component, while exit slippage was only -$14.96 (attribution.decomposition.totals; attribution.exit_slippage). That combination supports an entry-side rule—decision-to-fill adverse-drift cancellation with its own counterfactual ledger—rather than blaming ordinary exit execution. Evidence is still not sufficient to claim a universal threshold effect; the promoted fact is recurrence and cost, not that any untested cutoff will solve it.

The taxonomy’s top bucket, “fresh-entry directional misses,” is also costly at 14 trades and -$168.35 (retro_buckets bucket 1), and it agrees with the weak all-alert hit rates (gauges.gauge1.s1/s2/s3). It does not yet suggest a narrower mechanical trade rule beyond testing whether any classifier cohort predicts direction; indiscriminately filtering losses after observing them would overfit one week.

Two previously promoted patterns stopped appearing as distinct buckets. The “short run over by a violent adverse rally” tail, first promoted on 2026-08-10 and subsequently marked dormant, is absent for a fourth consecutive taxonomy (prior_weekly_reviews 2026-08-10 §4 and 2026-08-31 §4; retro_buckets current taxonomy). S4’s specifically wrong overnight-drift bucket from last week also did not recur mechanically: current S4 mean available move was +0.11% versus the -0.05% baseline and S4 selection P&L was only -$7.20 (gauges.traded.s4_overnight.mean_available_pct; scoreboard.do_nothing_baseline.mean_pct; attribution.decomposition.per_strategy.S4.selection_pnl). That is one week of respite, not evidence that S4 has predictive edge or that the prior problem is closed.

5) EXACTLY ONE BIGGER BET

Target: prediction, specifically the only current materiality cohort showing a positive directional separation. This is deliberately a shadow threshold experiment rather than another live P&L optimization. The evidence is thin and contradictory across weeks: this week S1 materiality ≥80 produced 26/46 hits, 56.5%, versus 130/288, 45.1%, below 80 (gauges.gauge1.materiality_calibration), while prior reviews found materiality flat or non-monotone (prior_weekly_reviews 2026-08-31 §2; 2026-08-24 §2).

• 2026-09-14 owner: add a shadow-only S1 materiality gate at 80, leaving the live threshold at 60, and journal separate ≥80 versus 60-79 Gauge-1 cohorts — expect: signal, next week the ≥80 shadow cohort beats the 60-79 cohort by at least 7 percentage points in directional hit rate, has mean_available_pct above 0, and contains at least 30 graded calls; otherwise the intervention is refuted, while fewer than 30 retained calls is not-yet-measurable.

This pulls one input lever—the shadow materiality threshold—and grades prediction before entry and exit. It does not target edge bps. If the apparent high-materiality separation disappears, that is useful evidence that this week’s 65.0% 80-89 cell was sampling noise rather than a recoverable signal.

INJECTION CHECK (reading_rules): neither retro_buckets nor coach_report_audit contains anything resembling instructions directed at me; both self-report clean inputs (retro_buckets header line; coach_report_audit preliminary note). Nothing to flag. One retro_buckets claim fails cross-check and is corrected in §4.

1. SCOREBOARD READ

Blended edge this week is -37.46 bps (scoreboard.week_edge_bps): s13 intraday -43.37 bps on 42 trades, s4 overnight -20.93 bps on 15 (gauges.traded.s13_intraday, gauges.traded.s4_overnight). Against gauges.trailing_trend, s13 gives back all of last week's lone positive print (+24.12) and posts the worst s13 number of the nine-week series (-18.47, -5.34, -9.96, -16.28, -13.48, -27.37, -15.76, +24.12, now -43.37). The swing from +24.12 to -43.37 is NOT distinguishable from noise: 42 s13 trades with capture_usd_sum -$151.49 (gauges.traded.s13_intraday), against single-trade dispersion of ±$30-57 documented in retro_buckets across recent weeks — the ~$270 week-over-week swing is a handful of trades. What is more likely real, because it persists: s13 negative 8 of 9 weeks, and gauge1 S1 hit rate now nine consecutive weeks inside the coin-flip band and never above it — 53.4, 49.8, 49.8, 49.9, 48.4, 45.8, 51.4, 49.0, now 46.7 (gauges.trailing_trend + gauges.gauge1.s1). The tape does not excuse the red: the do-nothing baseline was flat (-0.05%, n=372, scoreboard.do_nothing_baseline) and market_pnl was +44.36 against total pnl -177.62 (attribution.decomposition.totals). No target value proposed.

GRADE OF LAST WEEK'S BET (2026-09-07 S4 shadow mode): NOT SHIPPED, and the graded metric is caught between its branches. Verified mechanically: no new PR appears in predictions_to_grade after #253 (2026-08-21), and coach_report_audit's preliminary note records an empty changelog_week and code_changed=false every session (untrusted, but consistent with the PR ledger). The metric the bet defined is still computable on the live S4 cohort, since nothing changed: selection spread = s4_overnight.mean_available_pct (+0.11, gauges.traded.s4_overnight, n=15) minus do_nothing_baseline.mean_pct (-0.05, scoreboard) = +0.16pp. That is neither ≤0 (so the two-week retire clock fails at week one) nor ≥+0.2pp (so the drift branch does not fire). On n=15 it is noise either way. The bet cannot be graded as designed and its question stays open.

GOVERNANCE FINDING, now the standing deliverable my 2026-08-31 review committed to (prior_weekly_reviews 2026-08-31 §5): the pilot's intervention channel is inoperative across both requesting organs. Reviewer side: four consecutive weekly bets unshipped (MAX_PUB_TO_DETECT_MIN three times — 2026-08-10/17/24 — then S4 shadow, 2026-09-07). Coach side: the S4 shadow recommendation pushed for the fifth time overall this week and never addressed (coach_report_audit §3a); the netting pattern documented five sessions running with no fix (§3b); the one novel coach rec (fill-drift cancel, 09-08) advocated one day and silently abandoned (§1). Meanwhile ZERO code of any kind shipped this week — every genuine clause in predictions_to_grade dates to 08-21 or earlier. The pipeline has stopped executing its own measurement organs' asks entirely; the north-star question remains blocked by non-action, not by data.

Operational flag, separate from the above: alpaca's pub→detect median degraded to 2545.5s this week (article_funnel.by_provider.alpaca) from the 234-370s range of every prior split week (prior_weekly_reviews §2, each week). by_provider_note warns the updated_at cursor fattens the p90 tail by design and the funnel cannot separate cursor re-ingest from regression — but a 10x median move is not a tail effect; the ingest side should be checked. This week alpaca, the "fast" provider, was 3.2x SLOWER than eventregistry (784.5s).

2. CHAIN DIAGNOSIS

Link 1 — Prediction (all alerts, gauge1). S1: 46.7% hit rate (n=334, binomial SE ~2.7pp), mean_available_pct -0.08 against mean_abs_available_pct 0.76 (gauges.gauge1.s1) — ninth consecutive coin-flip week, never above the band. S3: 44.8% (n=96, SE ~5.1pp) — coin flip (gauges.gauge1.s3). S2: 36.7% (n=30, gauges.gauge1.s2); pooled across the five reviewed weeks S2 sits at roughly 69/176 ≈ 39% (8/36, 6/24, 23/51, 21/35 per prior_weekly_reviews §2 each week, plus 11/30 now) — persistently below coin flip on a now-nontrivial pooled n. Materiality calibration (gauges.gauge1.materiality_calibration): 60-69 → 45.0% (n=271), 70-79 → 47.1% (n=17), 80-89 → 65.0% (n=40), 90-100 → 0.0% (n=6). The 80-89 band's 65% is the first good-looking adequately-sampled band in five reviews — but the same band printed 47.7, 39.0, 51.4, 47.0 in the four prior weeks (prior_weekly_reviews §2 each week), so pooled it remains ≈47%; one week's n=40 against four contrary weeks is noise until it repeats. The 90-100 band went 0-for-6 on the week's largest available moves (mean_abs 1.06). Fifth consecutive week: materiality does not rank prediction skill.

Link 2 — Transit. Priced-at-detection 20.7% (133/642, gauges.priced_at_detection), in line with prior weeks. On the traded cohort transit again ADDED money on both books: s13 transit_usd_sum +$24.74, s4 +$15.66 (gauges.traded) — not the leak, fifth consecutive week. Blended pub→detect median 785.1s ≈ 13.1m (article_funnel.latency_pub_to_detect_secs); per provider, eventregistry 784.5s and alpaca 2545.5s (article_funnel.by_provider) — see §1's operational flag on the alpaca inversion. No per-ticker transit-leak field exists in per_ticker (#87 still not live); that read remains unavailable.

Link 3 — Selection. s13: available_usd_sum -$126.75, mean_available_pct -0.34 (gauges.traded.s13_intraday) against a -0.05% do-nothing baseline (scoreboard.do_nothing_baseline) — a -0.29pp selection spread; the trades the system chose had materially worse available moves than the average alert it did nothing about, and worse than its own all-alerts S1 mean of -0.08 (gauges.gauge1.s1). s4: mean_available_pct +0.11, a +0.16pp spread — mildly positive, thin n=15, and a softening of last week's S4 conviction (spread was -0.70pp then, prior_weekly_reviews 2026-08-31 §2); consistent with S4's sign-flipping history, not a rehabilitation. Attribution: selection_pnl -91.36 of -177.62 total, S1 alone -60.06 (attribution.decomposition). Per-ticker dispersion (per_ticker rows): worst MU -247.82, NVDA -172.69, PLTR -113.66, AAPL -92.21; best GOOGL +109.90, QCOM +108.22, META +65.60. Surprise rate 0.0% on all 15 tickers, fifth straight week. AMD, ASML, SMCI graded no trades (null week_edge_bps) at the three highest priced rates on the board (66.4/46.5/57.1) — their news continues to arrive dead.

Link 4 — Capture. Exit slippage small: -$14.96 over 42 exits, mean -$0.36 (attribution.exit_slippage). MFE-realized degenerate again (S1 +1.23, S4 -13.14, attribution.mfe_realized — fractions outside [0,1] unusable except as near-zero-denominator symptom). Timing_pnl -130.62 of -177.62, S1 alone -72.62 (attribution.decomposition) — the entry-side timing leak that shrank to -30.25 last week (prior_weekly_reviews 2026-08-31 §2) is back at the -120 to -160 level of the 08-17/08-24 weeks. Corroborated after cross-check by retro_buckets bucket 2 (chased fills on wrong signals, -$136.01, 6 trades) and bucket 5 (stale entries, -$22.51), with exit-side damage comparatively small (bucket 6, -$19.27).

SINGLE WEAKEST LINK: prediction, fifth consecutive week. Convicting numbers: S1 46.7% on n=334 with mean available at detection of -0.08% (gauges.gauge1.s1); nine weeks in the coin-flip band, never above (gauges.trailing_trend); S1 traded to a 35.3% win rate and -$128.36 (scoreboard.per_strategy S1); the traded s13 cohort's available move 0.29pp WORSE than doing nothing (gauges.traded.s13_intraday vs scoreboard.do_nothing_baseline); S2 pooled ~39% over five weeks; the 90-100 materiality band 0-for-6. retro_buckets buckets 1+2 (-$304.36, 20 trades, plain directional failure) say the same thing in words and survive the mechanical cross-check. Timing is again large in dollars (-130.62) and the standing reasoning holds without oscillation: timing losses on top of a coin-flip classifier are the cost of monetizing noise. The S4-specific conviction that carried last week's bet did not extend this week (+0.16pp spread, 46.7% WR, selection_pnl only -7.2 per attribution.decomposition.per_strategy.S4) — the S4 question is open, unresolved solely because the shadow experiment never shipped. The S1-S3 publish-vs-detection confound likewise remains untested solely by non-shipping (§1).

3. PREDICTION GRADES

All genuine clauses carry deploy_status "live" — a single-instant read at the week's last session (2026-09-11), cited per reading_rules; where that gives no assurance the change ran for the cited numbers, I stay at not-yet-measurable. Backfilled clauses (#108 and below) are attribution material only, not graded. Oscillation guard: no grade below reverses a prior week's grade.

GRADE #253 cost: not-yet-measurable — deploy_status live; no shadow_sessions.elapsed_ms or page-count field in any mechanical section. Fifth week at this status. GRADE #242 signal: not-yet-measurable — deploy_status live; no mechanical section counts counter-move S4 entries or exposes s4_alignment_skips. GRADE #241 cost: not-yet-measurable — deploy_status live; no page-count data in any mechanical section (the 7,520 baseline lives only in the clause text). GRADE #228 freshness: not-yet-measurable — deploy_status live; no stale-at-fill cancel count anywhere mechanical (gauges.traded transit_queued_unknown_n=0 measures a different thing). GRADE #226 freshness: not-yet-measurable — deploy_status live; no absorbed-move counterfactual ledger in gauges.declined or strategy_aggregates. Noting without grading: retro_buckets bucket 5 (untrusted) again describes stale in-direction entries (ids 983, 984, 985, 997, 1005). GRADE #222 signal: not-yet-measurable — deploy_status live; gauges.declined.priced_block (fired 133) is not band-resolved by |since|, so sub-2.0 "verdict PRICED" skips remain uncountable. GRADE #216 freshness: not-yet-measurable — no stall/paging data in mechanical sections. GRADE #212 signal: confirmed — the ledger exists and grades: gauges.declined.news_age_s13 fired 24, with_counterfactual 24, blocked_pnl_pct_sum -6.93; news_age_s4 fired 30, with_counterfactual 9, blocked_overnight_pnl_pct_sum -3.48 (S4 coverage still thin: 18 of 30 excluded as slot_outranked/slot_taken_by_entry). Fifth consecutive confirmed. GRADE #211 signal: not-yet-measurable — daily-report contents have no mechanical counterpart. GRADE #209 signal: not-yet-measurable — article_funnel.by_provider.eventregistry (articles 19956, classify_eligible 15651) has no per-day duplicate-eligibility delta. GRADE #207 cost: not-yet-measurable — no retry-ladder timing data in mechanical sections. GRADE #206 freshness: not-yet-measurable — no stale-at-fill cancel counts in mechanical sections. GRADE #203 cost: not-yet-measurable — prior_weekly_reviews retains four consecutive reviews, evidence Mondays produce reviews, but the clause covers both arms every Monday and no mechanical run-log exists. GRADE #202 freshness: confirmed — article_funnel.by_provider exists with per-provider latency (alpaca median 2545.5s, eventregistry 784.5s), and this review grades #177 off the split, never the blend. Fifth consecutive confirmed. GRADE #198 cost: not-yet-measurable — no journal-row reasoning-field data in mechanical sections. GRADE #193 risk: not-yet-measurable — no S4 failed/orphaned-enter counter; news_age_s4.post_close_decision_excluded=1 is a ledger exclusion, not an orphan count. GRADE #188 signal: not-yet-measurable — coach citation behavior has no mechanical counterpart; coach_report_audit's citation findings are untrusted, one hop removed (and this week they cut against citation quality — §2a-d — but that is not gradeable evidence for this clause either way). GRADE #185 signal: not-yet-measurable — no pre-open prev_close comparison field and no stale-at-fill cancel count in mechanical sections. GRADE #179 freshness: confirmed — the gate demonstrably fires across strategies: gauges.declined.news_age_s13 fired 24, news_age_s4 fired 30; strategy_aggregates.top_skip_rationales "news 4-8h old at decision, floor 4h" S4=12, "news 1-7d old" S4=10. As in all prior weeks, "trades entered on >4h news → 0" is inferred from gate activity plus the absence of any contrary field. Counterfactual note: blocked s13 cohort -6.93% (avoided_losses 14 vs missed_wins 10), blocked S4 cohort -3.48% on n=9 — loss-avoiding this week; signs keep bouncing week to week, still noise. GRADE #177 freshness: refuted — article_funnel.by_provider: alpaca pub→detect median 2545.5s vs eventregistry 784.5s — alpaca is now 3.2x SLOWER, an inversion of the stable 2-3.5x-faster ratio of the four prior refuted weeks (prior_weekly_reviews §3 each week); fifth consecutive refuted, and the miss is now categorical rather than a shortfall. by_provider_note's cursor caveat covers the fat p90 tail (106,056s) and says this funnel cannot separate cursor re-ingest from regression — but under no reading does "≥5x lower than ER" hold this week. The per-provider-never-blended half remains honored (see #202). Check the ingest side per §1's flag. GRADE #172 freshness: not-yet-measurable — no opening-tick drain metric in mechanical sections. GRADE #168 cost: not-yet-measurable — no historical-delta data in this prompt. GRADE #166 freshness: confirmed — proxy: article_funnel.latency_detect_to_alert_secs median 22.8s (n=372), far under the 10m target; flagging again that detect→alert contains but is not labeled "classify wait." GRADE #153 cost: not-yet-measurable — no shadow-send data in mechanical sections. GRADE #155 cost: not-yet-measurable — article_funnel.journaled_llm_cost_partial_usd 6.96 is an explicit floor (journaled_llm_cost_note) with no per-call baseline split. GRADE #151 cost: not-yet-measurable — no audit fire/no-fire flip data in mechanical sections. GRADE #147 risk: not-yet-measurable — no mechanical tripwire counter; retro_buckets and coach_report_audit both report no injection this week (untrusted). GRADE #146 risk: not-yet-measurable — same class as #147; no mechanical measurement. GRADE #143 signal: confirmed — scoreboard.per_strategy: S2 3 trades +$1.59 + S3 5 trades -$24.72 = -$23.13 combined, inside "toward ≥ -$30" from ≈ -$108, with the floor visibly firing (gauges.declined.materiality_floor fired 56; strategy_aggregates.top_skip_rationales "materiality 65 below floor 75": S3 27+15, S2 11). Fifth consecutive confirmed. This week the floor's blocked cohort would have LOST 21.15% (avoided_losses 33 vs missed_wins 18, gauges.declined.materiality_floor) — the floor saved money for once — but the standing caveat holds: it caps volume, not quality; pooled S2 hit rate is ~39% (§2), and gauge1.s2 printed 36.7% this week. GRADE #140 cost: not-yet-measurable — no timeout/page data in mechanical sections. GRADE #139 cost: not-yet-measurable — no error-envelope data in mechanical sections. GRADE #138 cost: not-yet-measurable — no auditor fire-count in mechanical sections. GRADE #137 freshness: confirmed — publish→alert ≈ 807.9s ≈ 13.5m (article_funnel: pub→detect median 785.1s + detect→alert 22.8s), stable versus 13.4m/13.4m/12.9m/13.1m in prior weeks and consistent with the honest-baseline claim; ER, 98% of articles, sits at 784.5s per by_provider. GRADE #134 cost: not-yet-measurable — no page-storm/WARNING data in mechanical sections. GRADE #131 cost: not-yet-measurable — scaffolding-citation rate has no mechanical counterpart. GRADE #130 signal: not-yet-measurable — no delimiter-breakout counter in mechanical sections. GRADE #128 cost: not-yet-measurable — no 400/self-heal data in mechanical sections. GRADE #126 cost: not-yet-measurable — no prompt_chars measurement in this prompt. GRADE #121 cost: not-yet-measurable — no timeout-page or shadow-row counts in mechanical sections. GRADE #120 risk: confirmed — measurable half: gauges.declined.tape_gate S1 fired 50, with_counterfactual 47, missed_wins 26 vs avoided_losses 21, blocked_pnl_pct_sum -0.68% (≈ -0.01% per blocked trade — net gated P&L ≈ 0, exactly the clause's promise, and within the range of prior confirmed weeks). The "worst-day S1 loss ↓" half remains unassessed (no worst-day field). Side note from the S3 row: the same gate blocked -8.84% across 12 S3 fires (11 avoided losses vs 1 missed win) — strongly loss-avoiding there this week. GRADE #117 signal: confirmed — proxy: processed/classify_eligible = 15919/(268+15651) = 100.0% (article_funnel + by_provider). Third consecutive clean week; standing caveat that this proxy cannot isolate the capped-tick tail specifically.

4. BUCKET PROMOTION

Re-strengthened promotion: the stale/chased-entry pattern (first promoted 2026-08-17, re-promoted 08-24, weakened 08-31 per prior_weekly_reviews §4 each week). This week it is retro_buckets bucket 2 (wrong signal compounded by chasing the fill 0.8-1.9% past decision/prior close, -$136.01, 6 trades) plus bucket 5 (entries after the move completed, -$22.51, 5 trades). Mechanical cross-check per reading_rules before repeating: timing_pnl rebounded to -130.62 (attribution.decomposition) after last week's -30.25 dip — that makes four of five reviewed weeks with timing as a large negative bucket (-311.13, -162.64, -120.57, -30.25, -130.62) — while exit slippage stays small (-$14.96, attribution.exit_slippage), localizing the leak to the entry side again. The coach independently proposed the matching rule (fill-drift cancel) on 09-08 and then abandoned it after one day (coach_report_audit §1, §3c), and the audit's own arithmetic (§5b) shows the coach's ≥1.0%-since-decision version would have covered only ~$37 of the $86 it was credited with — so the eventual rule needs either a lower bar or a consistent baseline (decision price, not prev close). Still no carrying PR in predictions_to_grade. This remains the most persistent named, costed, unactioned monetization defect in the pilot.

Previously promoted, did not extend: S4 overnight-drift failure (promoted 2026-08-31 §4). This week S4 printed 46.7% WR (scoreboard.per_strategy), selection_pnl only -7.2 (attribution.decomposition.per_strategy.S4), and a +0.16pp available spread over the baseline (gauges.traded.s4_overnight vs scoreboard.do_nothing_baseline). One soft week on n=15 does not retire the pattern — S4's weekly sign has flipped throughout gauges.trailing_trend — and the shadow experiment that would settle it never shipped (§1). It stays promoted-and-unresolved, not closed.

Correction to Stage A, per reading_rules cross-check: retro_buckets bucket 8 speculates that S2_S3_MIN_MATERIALITY=75 "is currently dead code." False — threshold_constants shows materiality_threshold=60 < 75, which is precisely the condition under which the floor BINDS, and strategy_aggregates.top_skip_rationales shows it firing this week (S3 "materiality 65 below floor 75" 27+15, S2 11; gauges.declined.materiality_floor fired 56). The correct reading of bucket 8's four tenuous sympathy trades (-$14.84) is that they cleared materiality ≥75 and lost anyway — the floor is live but is a volume cap, not a quality filter, consistent with the standing #143 caveat.

Still dormant: the "short run over by a violent adverse rally" tail (promoted 08-10, dormant since 08-24). Fourth week with no such bucket; the worst bucket is a diffuse -$168.35 across 14 trades (retro_buckets bucket 1). Remains dormant, reopen on recurrence.

Observed, not promoted: netting-forced exits (retro_buckets bucket 7, -$15.90, 4 trades; gauges.declined.netting fired 127 with mean_move_to_close_pct -0.24) — the coach flagged this five sessions running without ever converting it to a recommendation (coach_report_audit §3b). Two weeks observed, measured cost still small; watch. Also explicitly not promoted again: bucket 4 (exit-capture gap on winners, +$79.03 realized with ~$100+ forgone) — four prior reviews' reasoning stands without oscillation: while gauge1.s1 reads 46.7% coin flip, exit redesign optimizes the monetization of noise.

5. ONE BIGGER BET

Target: prediction, the weakest link named in (2), at its five-week-old unresolved confound — "signal never existed" vs "signal alive at publish, dead by the ~13-minute median detection delay" (article_funnel.latency_pub_to_detect_secs 785.1s). Per my own 08-31 commitment I am not re-issuing the S4 shadow bet, and per the 08-24 commitment I am not re-issuing the latency GATE. Instead I am changing the lever class deliberately: the only reviewer-adjacent ask that has EVER shipped is an owner-built weekly measurement ledger (#212, predictions_to_grade — shipped, confirmed five consecutive weeks per §3), while trading-rule bets are 0-for-4 (§1). So this bet pulls a measurement lever that requires no risk change, no gate, and no trading behavior change — minimum shipping friction, maximum north-star yield:

• 2026-09-14 owner: gauge1 splits every graded S1 alert by its article's publish→detect latency at detection time (three bands: ≤300s, 300-900s, >900s), reporting per-band n, hit_rate_pct, and mean_available_pct alongside the existing gauge1.s1 rollup — pure analytics on data the journal already holds (the same timestamps behind article_funnel.latency_pub_to_detect_secs), shipped exactly like the #212 news_age ledger — expect: signal, if over two consecutive weeks the ≤300s band's pooled hit rate (require pooled n≥60; alpaca alone alerted 40 times this week per article_funnel.by_provider, so two weeks should clear it) sits within ±3pp of the >900s band's, the latency-decay hypothesis is refuted and "no measurable directional signal at any observed detection latency" becomes the pilot's standing calibrated conclusion for S1; if instead the ≤300s band exceeds the >900s band by ≥5pp at that n, detection latency is convicted as the prediction leak and the thrice-unshipped MAX_PUB_TO_DETECT_MIN gate inherits a data mandate rather than a hypothesis.

Grading notes for next week: the graded metric is the between-band hit-rate spread inside gauge1 — never edge bps, no target proposed. Baselines to grade against: this week's undivided gauge1.s1 (46.7%, n=334) and the blended latency median (785.1s), which puts roughly half of alerts on each side of the middle band. Beware the alpaca cursor artifact flagged in §1 when reading the fast band — if alpaca's median stays ~2500s (article_funnel.by_provider), the ≤300s band will be ER-fast-path only and thinner than expected; that is why the clause pools two weeks. If this bet — a zero-risk analytics ledger with a shipped precedent — also fails to ship, next week's review should treat the governance finding in §1 as the pilot's primary result: a system that cannot execute any intervention, even a measurement, cannot answer its own north-star question, and the honest weekly deliverable becomes recommending the pilot's operators audit the shipping channel itself before any further weekly measurement is spent.