New buyback authorization × published days-to-cover

Date: 2026-08-15

Verdict: REJECTED

2026 sealed screen: not opened

Capital/scheduling authority: none

What was tested

One frozen semantic candidate: source-reviewed new buyback authorization x lagged published days-to-cover in the grammar's exact high bucket (5 <= DTC < 10), long from the next-session official open through the fifth session's close. The estimand was the high-versus-complement difference in same-date, same-DTC-bucket matched event-minus-control returns.

The registration was committed before prices at 6c2a6b48; the tested runner was committed at 449d96c2; and the outcome-blind manifest receipt was committed at ae1a1698. Its manifest hash is 17cd0c7e134b5ef84d9037422dc2b29bd11bdbe2dbf7a5b258d8e6def958032b.

Frozen sample and result

• 19 high-DTC events across 19 dates and 18 issuers.
• 52 complement events across 45 dates and 49 issuers.
• Five controls per event, 287 unique controls, 100% exact target-session exit coverage.
• One otherwise eligible event failed the frozen match-distance caliper before outcomes.

The effect floor, cost stress, best-date removal, most-frequent-issuer removal, and date concentration gates passed. Four preregistered gates failed:

1. permutation p exceeded 0.05;
2. the date-cluster lower bound crossed zero;
3. the issuer-cluster lower bound crossed zero;
4. the prior-window interaction was +1.65%, far above the allowed 0.50% magnitude.

The full aggregate evaluation receipt is ec0a6b31d285906231f97e125a77eb4980b5f224f1a1550c358749af0d33d1f4; the ordered event-effect payload hashes to bbfcbc8085d491acd38890d3a8995f2d839e11885fc496fec6ba787b6a1dc467.

Interpretation

This is exactly the kind of attractive number the verifier is supposed to refuse. High-DTC buyback filers did outperform their controls after the filing, but much of the same relative move was already present before the filing. With only 19 high-DTC events, the apparent +2.53-point interaction is also too unstable across dates and issuers to clear the frozen inference gates.

The result does not support lowering the DTC threshold, changing the clock, conditioning on momentum, lengthening the horizon, or opening 2026. Those would be post-outcome rescues of the same lineage. Pilot 001 is retired. The reusable lesson for the generator is narrower: mechanisms tied to an initiation event need an explicit pre-event contrast, because a persistent participant state can make a stale event clock look causal.

Date: 2026-08-15
Status: preregistration; no decision-session or forward price for this semantic candidate has been queried
Phase: bounded discovery only
Capital authority: none

Prior-evidence disclosure

This is not a pristine first look at the broad idea of buyback demand. A different, human-designed 2026-08-14 setup tested an active authorization after a three-session markdown and a 15:30 VWAP reclaim; its development matched-control result was negative. Those outcomes are known. This pilot uses a different clock and estimand: the first session after a source-reviewed new-authorization filing, with lagged published days-to-cover as an interaction rather than a post-flush trigger. It may produce discovery evidence, never a confirmatory edge claim. A result cannot rehabilitate or reinterpret the retired 2026-08-14 setup.

The exact new authorization x high DTC semantic candidate has one outcome trial. There will be no alternative event class, DTC threshold, direction, clock, horizon, cost, eligibility rule, matcher, or outcome definition on this discovery epoch. A renamed or reordered version is the same trial.

Claim and estimand

Mechanism: a newly authorized corporate bid should matter more when covering the published short stock would require several normal-volume sessions. The directional claim is long.

• Initiator A: source-reviewed new_authorization, available at the first research session strictly after the SEC filed date.
• Constraint B: the most recent short-interest observation with settlement_date <= decision_date - 14 calendar days has 5.0 <= days_to_cover < 10.0 (the frozen grammar's high bucket).
• Decision clock: official open of that first research session after filing.
• Horizon: five sessions including the decision session; entry is decision-session open and exit is the close at decision_session_seq + 4.
• Primary estimand: the incremental matched interaction mean(event - controls | B) - mean(event - controls | not B).

not B contains the grammar's low, mid, and extreme buckets. Controls are matched within the exact four-level DTC bucket, so a small number of extreme observations cannot alter the complement merely through a different bucket mix.

Frozen epochs and trial budget

• Discovery: decision dates 2023-07-03 through 2025-12-31.
• Sealed historical screen: 2026-01-01 through 2026-07-31. It remains unopened unless every discovery gate passes. A failed or inconclusive discovery result does not consume it.
• Forward evidence: events after 2026-07-31; unavailable today.
• Semantic outcome-query allowance on discovery: exactly one. The prior-window placebo is part of that same predeclared query, not another candidate.
• Permutations: 10,000, seed 20260815. Date-cluster and issuer-cluster bootstraps: 10,000 each with the same seed namespace.

Outcome-blind event eligibility

The roster adapter's semantic deduplication, SEC filing clock, stable case identity, and 14-day DTC lag are mandatory. The event must also have, at the decision open:

• a research_daily_baselines row with at least 20 prior sessions;
• no prior_20_has_source_gap and no has_split_after_date measurement flag;
• prior close at least $5 and trailing prior-20 median dollar volume at least $25 million;
• prior-session realized_vol_20d and momentum_20d;
• current security metadata classified as common stock or ADR common, with missing type treated as common stock, and no ETP flag.

The current security-type map is a disclosed survivorship limitation. It may exclude a row but may not become a feature. Event repeats from the same issuer remain eligible after semantic action deduplication; inference and concentration gates explicitly cluster them.

Frozen risk set and matcher

For every event, controls must be on the same decision date and pass the identical history, source gap, split, price, liquidity, feature, and security-type gates. A control must have a lag-safe DTC observation in the same exact low/mid/high/extreme bucket and no source-reviewed buyback event of any supported class with a decision clock within plus or minus five research sessions.

The only matching features are values known before the decision open:

1. log trailing prior-20 median dollar volume;
2. prior-session realized_vol_20d;
3. prior-session momentum_20d.

For each event, standardize each feature by the eligible same-date, same-bucket control pool, rank controls by Euclidean RMS z-distance, break ties by symbol, and freeze the nearest five. Matching is with replacement across different events. Every selected control must have RMS distance at most 2.5; otherwise the event is unmatched. No sector, current market cap, same-day volume, event-day return, later security state, or outcome availability can enter matching.

The full ordered match manifest, input query hash, source snapshot, exclusion counts, and manifest SHA-256 are frozen before any outcome query. An outcome-missing control is not replaced by a better ranked control after returns are visible.

Input and data-coverage gates (checked before effect computation)

• at least 18 matched B=true events on at least 15 dates and from at least 15 issuers;
• at least 45 matched B=false events on at least 30 dates and from at least 40 issuers;
• exactly five frozen controls per included event;
• at least 98% exact target-session exit coverage in each event/constraint cell and its controls.

If a target close is absent, use the last observed close from the decision session through target session and flag the fallback. If exact-exit coverage fails, return INCONCLUSIVE_DATA; never choose another control or enlarge the date range.

Costs, statistics, and counterfactuals

Gross return is adjusted open-to-adjusted close. Primary net return subtracts 30 bps round trip; 50 bps is the required cost stress. Identical control costs cancel algebraically from each matched effect, but all four absolute cells and the high-DTC event book are reported net of both costs.

All of the following must pass:

1. the primary interaction is at least +1.00 percentage point;
2. a one-sided DTC-label permutation within decision calendar year has p <= 0.05;
3. the one-sided 95% lower bounds from both decision-date-cluster and event-issuer-cluster bootstraps are above zero;
4. the high-DTC event-minus-control effect is positive, and the high-DTC event book's mean return remains positive after 50 bps;
5. the interaction remains positive after excluding the best decision date and separately after excluding the most frequent event issuer;
6. the prior-window placebo interaction, measured from the open at decision_seq - 5 through the close at decision_seq - 1 on the same frozen symbols, has absolute magnitude below 0.50 percentage point and one-sided permutation p > 0.10;
7. no single decision date contributes more than 20% of the sum of absolute event-level matched effects.

The four raw cells, matched residuals, bucket counts, independent dates and issuers, match-distance distribution, control reuse, exact/fallback exit counts, bootstrap bounds, permutation values, and leave-one-cluster diagnostics are mandatory output. A favorable main effect with no incremental interaction is a rejection, not a near miss.

Mechanical verdict

• PASS_DISCOVERY: every input, coverage, effect, inference, placebo, cost, and concentration gate passes. This only authorizes running the unchanged sealed screen.
• REJECTED: the powered and measurable sample reaches outcomes but any economic, statistical, placebo, robustness, or concentration gate fails. The semantic lineage is retired.
• INCONCLUSIVE_INPUT: an input or match-count gate fails before outcomes; no outcome may be read.
• INCONCLUSIVE_DATA: the frozen roster is powered but required outcome coverage fails. The roster may not be repaired after inspection.

No historical verdict authorizes paper orders, live orders, scheduling, or risk allocation.