Calibration D1: Pre-Registration

1. Standing and scope

This document pre-registers the calibration study for a weighting layer on share-delivery instruments, instruments that deliver shares against a cash strike or for nothing, confer no payment right, and never enter the CEBE numerator. The layer weights the denominator the way ACP weights the numerator: each instrument carries a stated prior probability of delivering its shares, published before the delivery record is scored against it.

Cal-D1 is a parallel study to Calibration 3, with zero contact with Cal-3's sealed holdout and its running sweep. Nothing here reads, cites, or depends on any Cal-3 artifact. The two studies share only the discipline: priors stated first, tolerance stated before results, disposition fixed in advance.

What this document is not. It does not apply any weight to any served surface. It does not alter ACP, Claims Grade, or any measurement figure anywhere.

2. The ratified priors, verbatim

Ratified 2026-08-16. Each weight is the stated prior probability that the instrument delivers its shares. Bands are read on moneyness, the stock price divided by the strike, both in the instrument's own currency at the observation date. Boundaries are left closed, right open, the same reading the ACP schedule carries.

Overlays, ratified with the table:

1. Strikeless and nil-cost instruments weight 0.95. Where vesting gates the delivery, the weight scales by the unvested fraction.
2. Contingency-unmet instruments weight 0.20, whatever the moneyness, until the stated gate is met.
3. Already-in-basic instruments are excluded entirely. Their shares sit inside shares_outstanding today; there is no delivery event left to weight.
4. One instrument, one probability. An instrument that carries both an ACP standing weight w and a denominator delivery weight takes both from a single prior: the delivery weight is 1 minus w. Two surfaces never state two independent opinions about one instrument's resolution.

3. Stated rationale

The shape is the textbook shape, stated as a prior and not fitted to anything. Delivery probability is monotone in moneyness. It passes near 0.5 at the money. It tops out below certainty, because the exercise-behavior record under ASC 718 shows holders of deeply in-the-money instruments routinely deliver less than the full pool inside any finite window.

The edges are volatility-adjusted, and asymmetrically. Fleet-level realized volatility is far above the equity baseline the textbook curve assumes, so the deep OTM floor sits at 0.20 rather than near zero, a deep OTM strike on one of these issuers is one squeeze away from the money, and the deep ITM ceiling sits at 0.85 rather than higher, the same volatility that rescues an OTM strike can strand an ITM one before the holder acts.

One open question is pre-registered rather than left to be discovered in the data: the tail-flattening hypothesis. At realized volatility of 80 percent and above, the true curve is flatter than these priors, and the deep bands move first. If the record supports it, the deep bands revise before the interior ones do. Stating this now is what makes that revision a finding rather than a fit.

The dataset also enforces an UNATTRIBUTED discipline: a delivery residual that cannot be attributed to an instrument by a filing enters no window, however well the arithmetic fits. The record already justifies it: a prior attribution consistent with the arithmetic was disproven by primary read, which is the standing argument for the discipline.

4. Exploratory and confirmatory split

The dataset splits into an exploratory set and a confirmatory set with its own sealed holdout. Seal mechanics match Calibration 3: the holdout membership list is committed as a hash before scoring, and the list discloses only at scoring. The bands may be studied freely on the exploratory set; the confirmatory set is touched once, at scoring, against the criteria in section 5.

Zero overlap with Cal-3's holdout is structural, not procedural. Cal-3's universe is claims-side convertible instruments; Cal-D1's universe is share-delivery instruments; and the census rule that excludes claims-side converts from this study is the same rule that keeps the two populations disjoint. No instrument can sit in both holdouts because no instrument can sit in both universes.

Seal timing, ruled 2026-08-16: the seal does not set until the gap-list returns land or are formally waived. At assembly the dataset held 31 resolved-outcome windows against the 40 the full-scoring requirement needs, and the MTPLF exercise-notice set was the largest single closure, so it carried gap-closure priority. A seal over a dataset known to be missing its largest outcome block would pre-register an underpowered study, and this document declined to do that.

Seal executed 2026-08-19, the gap returns having landed: the MTPLF outcomes file (SHA-256 48e52ba4, confirmed current) and the 18th and 19th Series terms rows (SHA-256 e6237c2f, VERIFIED from EDINET S100XTWY) brought the dataset to 54 resolved-outcome windows, 51 band-minimum eligible. The holdout membership list is held off tree by the operator; its SHA-256 hash, the split mechanics, and the gaps noted against the seal by the 2026-08-19 ruling, including the 27th Series terms-row withdrawal with its waived dangling references, are recorded in the seal record beside the dataset. The list discloses at scoring and verifies against the committed hash.

5. Success criteria, stated before data

Amended at criteria-review stage, pre-seal, in response to a sampling-noise analysis: fixed error tolerances alone would fail or pass small bands on noise rather than on calibration. The amendment replaces the fixed tolerance as the primary test with an interval test and retains the tolerances as the large-sample ceiling. The criteria below were fixed by the 2026-08-19 seal.

1. Per-band calibration. For each band, the observed delivery frequency is the pool-fraction-weighted rate of delivery across the band's confirmatory instrument windows. The band passes when the band's prior weight lies inside the 90 percent Wilson score interval around the observed delivery frequency, with n taken as the band's resolved confirmatory window count. The fixed tolerances of the original draft, 0.10 for the edge bands (Deep ITM, Deep OTM) and 0.15 for the interior bands (ITM, Near, OTM), are retained as the binding ceiling at large n: a prior within its fixed tolerance of the observed frequency passes even where the narrowed interval would exclude it, so as the interval tightens with sample size the acceptable error converges to the fixed tolerance and never below it.
2. Monotonicity. Observed delivery frequency, ordered Deep OTM to Deep ITM, is non-decreasing, ties allowed, on the pooled exploratory plus confirmatory record.
3. Minimum evidence. A band with fewer than 8 resolved-outcome instrument windows in the confirmatory set reports UNSCORED, never PASS or FAIL, and its prior stands as stated. Ruled 2026-08-17, pre-seal: windows resolved zero-to-date on open windows are retained in the dataset but are excluded from band minimums until the closing filing resolves them; EST-derived windows count toward band minimums.
4. EST sensitivity, ruled 2026-08-17, pre-seal. Every band score reports twice, once with and once without its EST-derived windows, so a PASS that depends on EST rows is visible on its face. Both readings publish together at scoring.
5. No-pass clause. Any scored band outside its acceptance region fails the schedule. Failure changes no served surface; the priors remain stated priors, and the failure publishes with the same prominence as a pass would have.

6. The dataset

Collection preceded scoring as its own ticket. The assembly is committed beside this page: input captures dated 2026-08-16 from the live instruments, snapshots, and deals tabs, the assembly script, the instrument-window dataset, a coverage report, and a gap list. One row per instrument per calendar quarter; moneyness computed from the snapshot stock price and the instrument's own terms at every use, never a stored ratio; outcomes sourced from the deal ledger, the instruments tab, and applied TSVs only.

Price sources, ruled 2026-08-16: the archival price spine capture (retrieved 2026-08-08, canonical listing venues quoted in listing currency) is approved as the moneyness anchor for windows the snapshot tab cannot price. Two conditions bind the approval and both are implemented in the assembly. First, every dataset row flags its price source (snapshot or spine); the snapshot remains primary and the spine fills only where the snapshot lookup gaps. Second, a robustness report is pre-registered for the scoring release: every computed window moneyness, both sources, is recomputed at 90 percent and at 110 percent of its window-start price, and every window whose band assignment changes under either perturbation is listed with its source flag. The report publishes with the scoring, whatever it shows.

No weight in section 2 has been applied to any row of the dataset. Collection and registration are complete; scoring is a separate, later act against the sealed criteria.

7. Publication order

This page published at /research/calibration-d1/ on 2026-08-19, after the read-back ratified the registration without changes, the same order Calibration 2 followed: document ratified, then the page, then the record.

1. The criteria, as fixed at the seal

Fixed in the pre-registration before the confirmatory arm was read. The band prior passes if it lies inside the 90 percent Wilson score interval around the observed delivery frequency, with n the resolved confirmatory window count, or within a fixed tolerance of 0.10 of the observed frequency, the large-n ceiling. A band scores only at or above 8 resolved confirmatory windows; below that it reports UNSCORED and its prior stands. Every band score reports twice, with and without EST-derived windows, published together. On failure the no-pass clause governs. No served surface changes, no weight applies anywhere, the priors stand as stated and revise only by dated revision to the registration.

2. The seal

The holdout membership list was held by the operator, off the tree, from sealing to scoring. First act of the scoring session, the list was re-hashed and matched the committed seal hash exactly, computed f9ede95906022a7079885f1ee4705a30b10ae4f2ba699cd9e78f5634346946ae against the identical expected value. The list then entered the record byte-exact beside the seal record and its noted gaps, including the 27th Series terms-row withdrawal with its waived dangling references. Twenty-eight instruments: thirteen confirmatory, fifteen exploratory.

3. Method

A window is an instrument-quarter, one instrument observed across one calendar quarter, so the twenty-eight instruments of section 2 resolve into more windows than instruments. Observed delivery frequency is the mean, over the band's eligible confirmatory windows, of each window's delivered fraction of its pool at window start: the pool-fraction-weighted rate of delivery. Moneyness was recomputed at use as price at window start over strike; the stored ratio column was not read. Zero-to-date windows were excluded per the 2026-08-17 ruling. Window prices carry their source flag, snapshot or spine, and the pre-registered price-source sensitivity publishes in section 5.

4. Results

Every figure in this report is from the scoring run of 2026-08-21, against the holdout sealed 2026-08-19 and the criteria fixed at that seal.

Three of the seventeen Deep ITM confirmatory windows delivered anything at all. The share-weighted rate, total delivered over total pool, reads 0.0045; it is disclosed here and is not the registered statistic. No verdict depends on an EST row. The confirmatory Deep ITM set contains none, and both readings are identical by construction at this record.

The other bands did not reach the scoring minimum and report UNSCORED with priors standing: Deep OTM 0 eligible confirmatory windows against a minimum of 8, prior 0.20; OTM 3, prior 0.25; Near 3, prior 0.50; ITM 0, prior 0.75.

Monotonicity on the pooled exploratory plus confirmatory record: OTM 7 windows at 0.0000, Near 7 at 0.0000, Deep ITM 27 at 0.0223, non-decreasing with ties allowed across the bands that hold windows. HOLDS. This is a schedule-level check and was not a gate on the Deep ITM verdict.

5. Robustness

Under the pre-registered price-source condition, every eligible priced window recomputed at 90 and 110 percent of its window-start price. Six of 41 windows change band, all six on snapshot-sourced prices, 6 of 32 snapshot against 0 of 9 spine. Zero confirmatory windows enter or leave Deep ITM under perturbation; the verdict does not move on price wiggle. The row-level table publishes in the companion robustness report.

6. The decision

The Deep ITM prior of 0.85 sits 0.8332 from the observed frequency, against an interval whose upper bound is 0.1648 and a fixed tolerance of 0.10. The fail is not close, and per the no-pass clause it changes nothing that serves. The priors stand as stated in every band, no weight applies anywhere, and revision happens only by dated revision to the registration. This page is the publication the clause requires, at the prominence a pass would have received.

7. The structural finding, and what D2 changes

The magnitude of the miss is the diagnosis. Seventeen deeply in-the-money windows produced deliveries in three, an observed per-window rate near 2 percent. That is not noise around a miscalibrated prior; it is a mismatch between the object the prior describes and the object the test measured. The 0.85 was stated as the probability that a deeply in-the-money instrument delivers its shares, a claim shaped like a lifetime, and it was scored against per-quarter delivery frequency. A holder whose lifetime delivery probability is high, spread across a multi-year instrument life, implies a per-quarter hazard in the low double digits at most, and the observed record is consistent with holders sitting on deep in-the-money positions rather than exercising into any given quarter. The prior and the test were talking past each other, and the registered test failed as registered.

The units mismatch does not explain the whole miss, and this page states the residual rather than leaving it to be computed. An 0.85 lifetime probability spread across roughly twelve remaining quarters implies a per-quarter hazard near 14 percent. The observed rate is 1.68 percent. Restating units therefore closes about a factor of six of a gap that spans about a factor of fifty, and a factor of roughly eight remains unexplained by the restatement alone. The hypothesis covering that residual is the one the record already suggests, holders sitting on deep in-the-money positions rather than exercising into any quarter, possibly joined by resolution paths that end without share delivery at all. Which part of the residual belongs to a wrong prior and which to a wrongly specified observable cannot be split until a lifetime observable exists. D2's design assumption is accordingly not that restating the object produces a pass; it is that restating the object produces a test whose result means something either way.

Calibration D1's holdout is now disclosed, which makes the entire historical record exploratory from this date forward. D2 is designed from that record: the test object restated so the prior and the statistic describe the same thing, per-window hazards or lifetime outcomes, priors fitted on the disclosed record and labeled as fitted, and a confirmatory arm built from future quarters, which seal themselves because they have not happened. The D2 pre-registration publishes as its own dated document before any confirmatory window resolves against it.

8. Reproduction

Pre-registration and seal record: /research/calibration-d1/, ratified 2026-08-19. Disclosed holdout list: cal_d1_holdout_2026-08-19.tsv beside the seal record. Scoring artifacts of record: SCORING-RECEIPT-2026-08-21.md, ROBUSTNESS-REPORT-2026-08-21.md, and the scoring runner, merged to main as #C-111. Hash of the disclosed list matches the committed seal hash; the match is quoted in the receipt.

Calibration record by @chcbearsfan | cebetracker.io

Revision log