How the published record has actually done

1 · Forecast calibration

Realized adjusted-close return and alpha vs SPY from each frozen out-of-sample anchor. Two cohorts, never pooled. The proxy cohort scores direction and tier from the pre-registered anchors, whose probabilities were never stored. The distribution cohort scores the probabilities the engine actually published and begins accruing 2026-08-04. Over 480 anchored names · Scored as of 2026-09-09 (today).

Anchored is not the same set as published, and the difference matters here. 875 anchored = 865 that are published, plus 10 that were anchored but never published — there is no report to click through to for those (AVB, BRK.B, EA, EQR, GLD, MP, NPN, NSA, TMHC, WBS). A further 19 published names (AMAT, AMD, BRK-B, CBRS, CIEN, COHR, DELL, FIX, GLW, HONA, INTC, LITE, LRCX, MRVL, MU, SPCX, STX, TER, WDC) began coverage after the anchors froze and so score nothing yet, which reconciles to the 884 names the rest of the site covers.

Why this table sums to 479 where §2 reports 480. Both score the same 480 registered anchors. §2 marks each name to its own last recorded price in the append-only snapshot ledger, which holds a row for every name; this section marks to the latest adjusted close instead and skips what it cannot price that way — 1 had no usable adjusted-close series at scoring time (EQR, a HOLD). The two denominators are reconciled here rather than forced to agree.

Directional hit rate (predicted vs realized sign): 47.2%. Brier score on the target-implied probability: 0.264 (lower is better; 0.25 is a coin flip). Sources: frozen anchors output/performance/anchor_reg004.json, output/performance/anchor_regmid.json, output/performance/forecast_ledger_*.jsonl.

2 · Benchmark-relative scoreboard

A forecast is only good if it beat the alternative of buying the benchmark, so every return here is shown net of both SPY and the name’s GICS-sector ETF. Benchmark: SPY + GICS sector SPDR (alpha = stock return − benchmark, same window). Generated 2026-08-04 16:51.

Pre-registered out-of-sample anchor

480 matured forecasts across 480 names · window 2026-07-08 to 2026-08-03 · median horizon 26 days — the clean forward test: frozen before the outcome was known

Mean sector-relative alpha +1.1%, 95% CI +0.3% to +1.8%.

Legacy calibration set

No matured forecasts in this set yet. (assembled before pre-registration, so it is indicative only and cannot rule out selection effects)

Reading this against §1. Both sections now score the rating that was actually published, and share one definition of a tier “hit” — a BUY beats its sector, a SELL trails it, a HOLD tracks it within 5 points. They still differ on one thing: §1 marks every forecast to the latest close, while this section marks to each name’s own most recent recorded observation. Small differences in hit rate between the two are that, and not a disagreement about the record. The pre-registration-era set is empty by construction — every name under coverage is now pre-registered, so there is no separate legacy population left to report. The frozen anchor files were not rewritten.

3 · Options overlay

Settled outcomes only, from the pre-registered ledger's inception. No backfill: a structure never recorded at entry can never be claimed here. Ledger inception 2026-07-28; 60078 structures pre-registered, 55557 still open.

Open book (mark-to-market)

The latest mark run priced 55473 of the 55557 open structures at 2026-09-09: 54831 off listed strikes, 642 carrying a leg whose strike is no longer quoted (modelled, and excluded from the means below). Mean open P&L on the listed subset 1.47 against a mean underlying move of -1.6%.

The two open counts reconcile by name. Of the 55557 unsettled registrations, 55473 were marked and remain open; the rest are 5 past expiry awaiting settlement (deliberately unmarked — settlement is the resolver's job); 79 deferred this run (chain unavailable, spot missing, or a leg unreadable — retried next run).

Mark-to-market of structures still open, from the same chain the overlay prices from. DIAGNOSTIC ONLY — open marks are not results and are excluded from every settled statistic above. Legs whose strike is no longer quoted are modelled and counted separately; the means shown use listed legs only.

Attribution caveat. Every settled structure is shown beside the same-name stock-only return over the identical window. Our own backtest attributes this sleeve's historical return to the volatility risk premium rather than to the ranking we publish, so the honest read is whether the structures beat simply holding the shares — not whether they made money.

Settlement assumption: held to expiry; early exercise/assignment not modelled.

4 · MCH Alpha Score (unvalidated diagnostic)

The §8 pass conditions cannot be evaluated today: no forward window has closed on any date carrying the full input set (earliest 2026-07-30; first 21-day close ~2026-08-31, 63d ~late Oct 2026, 126d ~Jan 2027). MAS ships as an unvalidated diagnostic column that drives no rating, stance or size; its accuracy record is published here under the same pre-registered anchor methodology as the rating tiers when windows close.

MAS is shown on the Alpha Lab table as MAS-1Y / MAS-5Y and drives no rating, stance or size — the engine fails its own QA if anything consumes it. This section is its accountability record, held to the same pre-registered anchor methodology as the rating tiers above: forward sector-relative alpha on adjusted closes, per snapshot date, never pooled, at horizons registered before any window closed (AM-057). λ = 0.3 · config ros-1.20.0 · 95 eligible of 893 names · as of 2026-09-09.

Status: preliminary — 0 non-overlapping monthly cross-section(s) at the primary horizon; verdicts need ≥12

An interval that crosses zero means the data cannot yet distinguish the score from noise at that horizon. Weights, λ and gates never move on these results without a new pre-registration amendment.

Input diagnostics this run — computed today, no forward returns involved: max |ρ| between blocks 0.97 (flagged > 0.85: A_signal~C_value, A_signal~D_asymmetry); corr(expected alpha, alpha/σ) 0.77 — expected to be high by construction, recorded, never re-weighted on.

Signal diagnostic — short-horizon IC

This section measures whether the published ranking (expected_alpha_pct across the fresh universe) carries short-horizon information. It does not measure forecast accuracy — the book’s forecasts are 12-month views — and it feeds no rating, stance, size, or scoreboard figure. Definition (pre-registered, AM-060): daily cross-sectional Spearman rank correlation (midrank ties) against forward GICS-sector-ETF-relative adjusted returns at 1, 5 and 21 trading days; block-bootstrap 95% CIs with block length ≥ horizon; labels flip by observation count, never by result. The live series is never backfilled.

Pre-registered null-result commitment: if, after the minimum history, every horizon’s CI includes zero, that finding is reported here with the same prominence as any positive result. The three horizons form a term structure — a deliberately slow valuation signal should show information rising with horizon; a null 1-day IC alone is not evidence of no information.