Detection spec-sheets
Bottom lineEvery completed study on one page: the hypothesis, the pass/fail gate written before the run, the recorded numbers, and the verdict — failures published on purpose. The evidence machine, not a highlight reel.
Why we publish thisWe publish failures and uncertified reads so descriptive context is never mistaken for alpha — the honest boundary IS the product.
Every claim on this site enters through a
pre-registered gate: the hypothesis and its pass/fail threshold are written before the run, and the result goes to the ledger — win or lose. These are the completed studies,
failures included on purpose: a page that only showed wins would be marketing, and the discipline is the product. Nothing here is a signal, a ranking, or a return promise. These sheets — with the coverage boundary and the live season record — are assembled into the print-ready
evidence pack.
⏱ Filing→surface MTTR (charter §9 KPI, live) — 1896 fresh results events since Jul-07: median 3d filing→board; 398 surfaced by the next evening. Day-level (real BSE date → first-seen date); the baseline seed run is excluded.
🎲 Placebo (M-02), PEAD SUE-Q5 x DELIV-T3 CAR60 — observed +5.95% vs null mean +2.98% / null p95 +4.54% over 200 shuffles (seed 42). Observed / null-p95 = 1.3×; empirical one-sided p = 0.005. Generated 2026-07-07 06:03:13Z — a real effect must sit OUTSIDE the shuffled-date band; this one does.
📏 Standing caveats (M-05) — printed beside every claim, not buried: 44.1% of the 22-year tape mass is unjoinable to fundamentals (the delistings live exactly there); 1,706 / 1,722 delisted names are fundamentally dark; the price archive is left-censored at 2004-07-23 (773 names). Every fundamentals-conditioned lift is therefore survivor-tilted — the survivor-vs-terminal re-cut (the attribution code) measures the direction of that bias, and the Deflated-Sharpe / PBO stages (M-03) are one import away in evlib for every study; factory auto-wiring lands with the next factory run.
PEAD on real BSE result dates, delivery-confirmeddescriptive lens CONFIRMEDbook FALSIFIED
pre-registered: 2026-07-05 (A-study) · within-season variant pre-registered 2026-07-05b · gate sha256 83a290a0ce12…
Hypothesis
A high Net-Profit surprise (SUE, no analysts) CONFIRMED by abnormal delivered value drifts upward over the next ~60 sessions — on REAL announcement dates (provenance_knowable), zero look-ahead.
Gate (before the run)
Descriptive: cohort t ≥ ~2 on leak-free dates. Book: net return/vol > 0.89 (Nifty-500 B&H) in BOTH walk-forward halves under tiered+ATR costs.
Recorded result
Descriptive: SUE-Q5 × DELIV-T3 mean CAR60 +7.62% (n=235, t_cohort 1.92); same surprise on thin delivery only +3.7% (n=200); population +3.5%. Bad news did not drift. Book: EVERY wrapper failed — trailing net return/vol 0.10 · no-delivery 0.02 · 1.5× cost −0.32 · hedged −0.58 · within-season (the pre-registered last cell) 0.06 — vs benchmark 0.85.
What ships
The results-reaction war room (who just reported + delivery confirmation + realized drift, base-rates labelled). No ranking, no trigger, no book — the tradeable version is recorded dead in the failure ledger.
Source of truth
the strategy ledger § Experiment 2026-07-05 · the pead research code
Accumulation-footprint detector (front-detect the insider from the tape)gate FAIL — published
pre-registered: 2026-07-05b — gate written in the module docstring before the run · gate sha256 0d86d11bd7c4…
Hypothesis
Insider/SAST accumulation episodes leave a detectable tape footprint BEFORE the filing becomes public (delivery %, trade size, volume, price character vs self- and cross-sectional controls).
Gate (before the run)
≥2 of 4 features clear Cliff's δ ≥ +0.20 vs BOTH control sets.
Recorded result
FAIL 1/4 — only avg-trade-size cleared (δ +0.329 / +0.250). Structural finding: 764 of 947 episodes had NO pre-public window at all — SEBI PIT T+2 disclosure means the tape you can trade is already public. Usable n=54. delivery-% showed ~no case elevation (δ ≈ +0.07), consistent with the accumulation tape's earlier alpha failure.
What ships
The trade-size ratio as a DESCRIPTIVE column (Ticket, on Screen+ / Positioning / dossier) — a characterization, never a detector. Detection pivots to post-public designs (disclosure-drift E-03, campaign arcs E-04), each requiring fresh pre-registration.
Source of truth
the strategy ledger § Study 2026-07-05b · the footprint research code
Insider disclosure drift (E-03, post-public)gate FAIL — placebo caught it
pre-registered: 2026-07-07 (module docstring before the run) · gate sha256 b62429563b21…
Hypothesis
Conviction promoter-buy clusters drift up after their first public disclosure (the pre-public window is structurally dead — SEBI T+2).
Gate (before the run)
Top-value-quartile CAR60 t_cohort ≥ 2 AND observed clears the shuffled-date placebo p95 (n=200, seed 42).
Recorded result
The trap this page exists for: value-Q4 CAR60 +8.26% (n=66, plain t 2.87) — looks like a product. The placebo null's p95 is +9.52%: random windows of the same names drift just as hard (the 2025-26 tape). Inflation 0.87×, empirical p 0.085; cohort t is NaN because the feed is only ~10 months deep. NULL published.
What ships
Nothing. Re-attempt condition on record: ≥8 quarterly cohorts of feed depth and a placebo-clearing mean.
Source of truth
the strategy ledger § Studies 2026-07-08 · the insider drift research code
Filing-latency tell (late vs own norm)gate FAIL — null published
pre-registered: 2026-07-07 · gate sha256 06e7c085f500…
Hypothesis
Filing later than one's own historical norm is a tell — lateness predicts bad surprises and negative drift.
Gate (before the run)
|t_cohort(Q5−Q1 CAR60)| ≥ 2 AND |gap| > label-permutation p95 (500 perms).
Recorded result
Half the folk story is real: the latest filers carry the weakest surprises (mean SUE 0.77 vs 1.04 for the earliest). But it does not price: CAR60 gap −0.81% (t_cohort −0.84), inside the ±2.51% permutation band. Lateness predicts surprise MIX, not tradeable drift.
What ships
Nothing — no war-room flag. The surprise-mix fact may inform reading, never a signal.
Source of truth
the strategy ledger § Studies 2026-07-08 · the filing latency research code
Concall growth-intent walk-forward (real call dates)gate FAIL — covered-name beta
pre-registered: 2026-07-07; design correction recorded before the run (condition on statement PRESENCE at the call, not settlement) · gate sha256 c1c4cf9d5a1a…
Hypothesis
Calls pushing specific guidance content (debt reduction, capex, volume…) drift differently over the next 60 sessions — the old month-granular panel showed +2.8%/+2.3%/+1.5% tilts.
Gate (before the run)
Per type (n≥100): t_cohort ≥ 2 AND same-sign halves AND the largest passing type clears the date-shuffle placebo.
Recorded result
9,461 real-dated events. Six types pass t+halves (debt_reduction +3.52% t 2.43; capex +3.07% t 2.70) — and the placebo unmasks all of it: random windows of the SAME covered names drift +2.75% (p95 +3.66%); the observed means sit inside the null band (inflation 0.52×, emp-p 0.925). The old panel tilts are recorded NOT reproducible on real dates.
What ships
Nothing as an edge. Guidance remains a candor / kept-promise DESCRIPTIVE axis (Gate B fence unchanged).
Source of truth
the strategy ledger § Studies 2026-07-08 · the concall intent research code
Dividend-surprise drift (E-11, post-ex)gate FAIL — payer beta
pre-registered: 2026-07-07 · the first gate HASHED before its run (M-04 registry) · gate sha256 e9bd1a7ff792…
Hypothesis
A dividend far above a name's own norm marks strength that drifts over the following 60 sessions (post-ex, the conservative clock).
Gate (before the run)
Surprise-Q5 CAR60 t_cohort ≥ 2 AND the date-shuffle placebo clears (n=200, seed 42).
Recorded result
22 years, 9,166 usable events. Q5 drifts +1.61% (t 2.60) — and so does everything else: CUTS drift +1.99%, hikes +1.61%, the LOWEST-surprise quintile +2.28% (55 cohorts). The placebo seals it: post-ex windows (+2.13%) drift LESS than random windows of the same payers (null mean +2.48%, p95 +3.58%). The 'surprise' has no direction — it is dividend-payer beta.
What ships
Nothing. Class rule generalized: covered-name / payer-universe drift is the null every event claim must beat.
Source of truth
the strategy ledger § Studies 2026-07-08 (S83g) · the dividend drift research code
Rebrand pump (E-12, stitched rename series)gate FAIL — no pump exists
pre-registered: 2026-07-07 · gate hashed before the run · gate sha256 c3f48a42abb1…
Hypothesis
A symbol rename pumps on the new identity, then fades.
Gate (before the run)
CAR22 t_cohort ≥ 2 AND placebo clears (pump); mean(CAR60−CAR22) < 0 with |t| ≥ 2 (fade). Wolfe power rule: 111 usable events = full-power claim.
Recorded result
The folk story is simply dead: CAR22 −0.41% (below even the placebo null of +0.47%), CAR60 +0.12%. No pump, and the fade leg earns no claim (pooled t 0.23; the cohort means run negative — outlier-carried, recorded as nuance). Method dividend: renames are unmeasurable on naive per-symbol series — this study shipped the reusable STITCHED old→new loader.
What ships
Nothing tradeable; the stitched loader joins the harness for every future boundary-crossing study.
Source of truth
the strategy ledger § Studies 2026-07-08 (S83g) · the rebrand pump research code
Wolfe waves — geometry as a selection lensdescriptive selection edge (BULL only)trade book FALSIFIED
pre-registered: §A locked 2026-06-24 (rules), §C trade-mechanics tested 2026-06-25
Hypothesis
Completed Wolfe geometries mark exhaustion points whose resolutions beat matched baselines (entry pt-5 zone, SL zone-edge ±0.3%, T1/T2).
Gate (before the run)
Per-trade net edge vs matched non-pattern baselines; side-split honesty.
Recorded result
The EDGE IS SELECTION, not mechanics: BULL completions carry +1.37% abnormal (median +2.14%) — but the mechanical trade book is dead: median −2%/trade net, the top 1% of trades carries 58% of profit. BEAR side is tail-only: regime-stripped −0.19%, decaying to −0.94% in 2021-26.
What ships
The Wolfe scanner + chart overlay as a descriptive lens (BULL-weighted, side split shown); never a mechanical trigger. Harmonic (XABCD) ships under the same fence.
Source of truth
the strategy ledger Tier-2/§C · the Wolfe rule book · the wolfe code
Rigorous calendar seasonality on the PIT idiosyncratic residualdescriptive estate LIVEtradeable calendar edge FALSIFIED
pre-registered: 2026-07-12 — 3 hypothesis families sha256-hashed BEFORE any compute (2882ccbc · cb32d1b9 · e566904c), frozen in research.db.prereg_registry; the engine docstring is part of the hash, so the gate cannot move.
Hypothesis
Do Indian indices / sectors / stocks run systematically hot or cold on the calendar (month · ISO-week · weekday) AFTER the market move is stripped out — a tradeable idiosyncratic-residual seasonality?
Gate (before the run)
A cell CERTIFIES only if it clears BOTH placebo nulls (circular-block + cyclic-rotation, p<0.05), family-wide BH-Yekutieli FDR, ≥15 scored years, out-of-sample sign-stability, AND a pre-pledged India mechanism — the year-label shuffle is banned (zero-width). Nothing is graded on the same data it was found in.
Recorded result
0 cells certified — across index + sector, all ~2,427 EQ stocks, and the broad Nifty 50/100/200/500 deepened to 2004 (~20 scored years, N-gate PASSES). The single strongest broad cell (Nifty 500 February) clears one placebo (p=0.0498) but dies under FDR. Individual names look overwhelming — MARUTI September up 89% of 18 years — yet clear the single-calendar placebo only to die under the multiple-months correction. Deeper data did not rescue it: the greying is on merits, not thin history.
What ships
The Seasonal Tape estate — /dash/seasonal-tape + screen + divergence + the event-cadence lens + the /dash/stock embed — kept strictly DESCRIPTIVE. Every drilled cell shows its year-by-year dispersion and a placebo ‘why-grey’ read; nothing is ranked, triggered, or tradeable. The greying IS the finding.
Source of truth
PROJECT_STATE §Session 120/130 · the strategies folder · the seasonal tape code (frozen families in research.db.prereg_registry)
Numbers are hand-carried from the strategy ledger (the canonical record) and the machine ledger in research.db.strategy_runs; the standing corollary those failures prove: price strength is the only gross forward-return engine we have found, value/quality/credibility/accumulation are context layers not rankers, and no factor here is a fundable net-of-cost alpha vs the index at AUM. Every return/vol on this page is mean return ÷ volatility, annualised — NOT a Sharpe ratio: no risk-free rate is subtracted, so it reads higher than a textbook Sharpe would. Descriptive research record, not investment advice.