Cal-num-4: champion selection, sealed registration

The registration text as sealed, reproduced verbatim from its source file. Published 2026-09-18. The hash is the SHA-256 of the source file's bytes as committed.

source verification/studies/calibration-v3/03-cal-num-4-registration.md
sha256 98c9af17f45552f7548de3f33a9e6aca0c0856a413def17c4561eed867582153
bytes 9,336
RESEARCH · NOT THE CORE STANDARD

Research extension. This page is a study, not a measurement. The core standard is CEBE, Senior Claims %, CEBE mNAV and Claims Grade, each published with its formula and date. Studies here test priors against outcomes and can be retired; retired studies stay on the site with their correction notes. Nothing on this page enters a tracker figure.

Cal-num-4, champion selection, sealed text

2026-09-03. Ratified as written by Bobby 2026-09-03 with the closures recorded in 00; candidate 6 withdrawn by ruling 2026-09-04 and replaced 2026-09-05 by a realized-volatility governor on Bobby's delegation. Design of record (v19 board) reconciled with the candidates-are-rules principle ratified 2026-09-01, amended for the external steelman of 2026-09-02. Supersedes v2 entirely. This is the text that seals; as of 2026-09-05 the MSTR gate is closed, the roster census and the tier-3 read have returned, and the seal waits on two items: the installation of the equity daily-close convention that candidate 6 cites, and the capture rail's commit to origin so the input series the study scores is a record before the hash is taken. The hash is taken on this file as ratified and any edit after ratification is a new study.

Family, Cal-num. Lineage, Cal-num-1, Cal-num-2, Cal-num-3 admitted with the exceptions recorded in the Cal-3 limitation note. Cal-num-4 is the program's first forward-scored selection under clean sourcing rules. That sentence is a claim about the protocol. It is not a claim about any candidate, including the incumbent.

1. Question

Which volatility-governor rule prices the panel best, by a margin the panel can resolve, while passing the absolute gates inherited from Cal-num-3? The field completes what phase one could not score. Cal-1's cap candidates were pending, not beaten, because no sourced implied-volatility series existed and the registration forbade fitting or assuming one. The capture rail running weekly since 2026-08-28 is that series.

2. Panel and duration

The scored panel is the quarterly confirmation slice, forward only. A scored mark is an issuer-disclosed fair value for a panel instrument in a filed 10-Q or 10-K dated after the seal, compared against each candidate's mark for the same instrument on the same date. Weekly captures under the capture protocol are inputs to the candidates and are never marks. No historical mark is scored, because R-VOL 6.6 forbids retrofitting captured values onto marks that predate the capture rules, and this study inherits that rule rather than excusing itself from it.

The study runs exactly four scored quarters. It does not conclude early, and no result is read as a selection before the fourth scorecard. The panel instruments are those of the Cal-num-3 registration, about six, so the study concludes on about 24 marks. That count is stated here so the reader knows the resolving power in advance. If at conclusion no challenger clears both the margin and the gates, the incumbent retains and the result publishes as "not resolved on this panel." That is a legitimate pre-registered outcome of this study, not a failure of it, and it is written here before any mark exists.

3. Candidate field

Six candidates. The original sixth, the issuer's disclosed volatility assumption, was registered on 2026-09-03 and withdrawn on 2026-09-04 before sealing, on the MSTR anchor's read of ten covering filings: Strategy's convertible notes are carried at Level 2 and no fair-value note discloses a volatility input, so the candidate's input has no referent in the filings. The only issuer-disclosed volatility is a year-to-date grant-date assumption in the share-based compensation note, which is a different object and is not substituted. The withdrawal precedes any scored mark. Its slot is taken on 2026-09-05, also before any scored mark, by a realized-volatility governor (candidate 6 below), selected by the copilot on Bobby's delegation so that the field carries a second independent challenger family and not only the cap gradient.

Six candidates. Every candidate is a rule with a stated reason. The gradient's numbers carry their reason as brackets.

  1. Incumbent, fitted on ineligible inputs. Cal-num-3's convention as admitted. Its parameters were fitted on the volatility inputs the Cal-3 limitation note rules ineligible, and it enters this field under that label. Its degeneracy floors are carried as its own declared constraints, inherited unchanged so that no floor is set with this field in hand. If the incumbent retains at conclusion, the program's sentence is "retained against clean challengers on this panel." The word clean is never applied to the incumbent.
  2. Null. No governor. Implied volatility as captured.
  3. Cap 65. The 65 percent street cap convention, contributed 2026-07-28, attribution pending permission. Cal-1's candidate 4a, unscored there for lack of a sourced series. Until permission is granted the page carries the convention with no name attached, and the attribution note already published stands.
  4. Cap 50. Lower bracket of candidate 3. Not an independent proposal. In the field so a result at 65 is interpretable against a tighter governor.
  5. Cap 80. Upper bracket of candidate 3, same standing.
  6. Realized-vol governor. Sigma is the trailing 90-trading-day realized volatility of the common: daily log returns of the official daily close on the primary listing (Nasdaq NMS for the common), taken through the site's stock-price rail (Finnhub, Yahoo fallback) under the equity daily-close convention EQ-CLOSE-1, installed 2026-09-18, annualized by the square root of 252, computed at each mark date and frozen. No cap, no floor, no blend with the captured implied series. One window, registered here with no scored data in hand; no brackets, because a gradient around one idea is the problem the cap family already has. This candidate reads a different source from every other rule in the field, so a result against it is informative in a way the cap gradient cannot be, and it engages every quarter by construction, so it can never publish as indistinguishable from the null. Its known weakness is stated in advance: realized volatility lags, and in a quarter where implied moves before realized catches up this candidate will misprice. That is the test. v1 withdrew it as a duplicate of the incumbent's family; that reason was wrong, since the incumbent is a fitted convention and this is a rule that reads the tape, and the reinstatement records the reversal rather than hiding it.

The listed-IV-as-rule candidate from v1 remains withdrawn: it is the capture itself and cannot govern itself. That withdrawal preceded any run.

4. Statistics, gates, and selection

Per scored mark, deviation as in Cal-num-3. Reported per candidate at each quarter-end, RMSE on all scored marks to date under fully frozen parameters, and beside it, labeled, the leave-one-date-out refit statistic for continuity with Cal-3's published number. Gating and selection read only the frozen-parameter statistic.

Absolute gates, inherited from the Cal-num-3 registration exactly and not set with this field in hand. A candidate passes only with frozen-parameter RMSE at or below 5.0 points and maximum absolute deviation at or below 12.0 points on the full four-quarter panel. A candidate that fails either gate cannot be selected regardless of its rank in the field. Cal-num-3's repurchase-print MAE tolerance is deliberately not inherited, because the forward confirmation slice contains no transaction prints and a tolerance with no observations to bind is not a gate.

Selection. A challenger displaces the incumbent only if it passes both gates and its frozen-parameter RMSE is lower than the incumbent's by at least 1.0 point with maximum deviation no more than 1.0 point higher. Ties and sub-margin outcomes retain the incumbent under the label in section 3. Where more than one challenger clears, the lowest RMSE wins; a tie among challengers publishes both and selects neither. If the incumbent fails a gate and no challenger clears, the result publishes as "not resolved on this panel, incumbent failed gate," and the Cal-num-3 retirement rule in 02 governs what happens to the admission.

Succession. If a challenger wins, the ladder (Cal-num-5/6) tests the winner, Cal-num-3's page carries a dated superseded line, and nothing is deleted.

5. Engagement rule, stated before any run

The cap family is inactive wherever captured volatility sits at or below the cap. If, over the four scored quarters, a cap candidate's governor never engages, that candidate's result publishes as "indistinguishable from the null on this panel," not as a win or loss, and no selection is made on its account. Whether the governor engaged is a published fact per candidate per quarter. The DM to the contributor says scoreable, not settleable, for exactly this reason.

6. Seal

Sealed before any scoring. This document complete, the candidate list, the gates, the margins, the engagement rule, and the capture protocol reference. The seal is this file's hash, committed to the repo and published externally the day of sealing per G-1. Cal-num-5/6 seals the same day under the same external hash. Any change after the seal is a new study.

7. Owners

Registration, copilot. Ratification, Bobby. Captures, capture desk, weekly, standing. Issuer series read returned 2026-09-04 as a negative finding; candidate withdrawn; the MSTR gate on the seal is closed. Scoring, a fresh study desk booted on the sealed file at each quarter-end. Page, Code. DM, Bobby, after the Cal-3 limitation note is public and the sealed registration is on the site.

More studies, and the construction they test, are listed on the research page. The core standard's construction is documented on the methodology page.