Cal-num-5 and 6: the generalization ladder, sealed

The ladder as sealed, with its dated amendments, 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-v3/04-cal-num-5-6-ladder.md
sha256 e59d2e13ff1f1c71042959522267f64a6b60b35593d2bb121a32dc4ad6f9483d
bytes 6,973
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-5 and 6, the generalization ladder, sealed text

2026-09-03, amended 2026-09-04 on the census return (vol-source class field, tier-3 population) and 2026-09-05 on the tier-3 read (tier 3 named, unstated class, replacement rule) and again 2026-09-05 on the tier-3 class read (class carried per instrument, transaction-calibrated as a class). Ratified as written by Bobby 2026-09-03 with the thresholds set below; amendments ratified 2026-09-04. Both precede sealing. Seals the same day and under the same hash as Cal-num-4, after the census read returns. As v1's design with the amendments below. Three tiers, MSTR fixed at the top, incidental holders, zero digital assets as the bright line. Three issuers per tier by sealed structural rules. Per-issuer leave-one-date-out RMSE, no pooling. Independent credit as covariate. Interpretation rule stated in advance.

1. Subject

The ladder tests the Cal-num-4 champion, whichever rule wins, and says so in its registered question. If Cal-num-4 concludes "not resolved on this panel," the ladder tests the retained incumbent under its label from 03 section 3, and the ladder's pages carry that label. The ladder seals the same day as Cal-num-4 under the same external hash, so the crowning and the generalization test are committed together before any result exists.

2. Vol source per issuer, sealed with the selection rules

For each named issuer, the scored volatility source is stated before naming. The issuer's disclosed volatility assumption from its own fair-value notes is primary. Where it is absent, the issuer is ineligible rather than improvised. No captured series exists off MSTR and none is retrofitted.

Vol-source class, added 2026-09-04. The census of 2026-09-04 found that issuer-disclosed volatility assumptions are not one kind of input. Each scored instrument, not each issuer, carries a class beside its source. The unit is the instrument because the class read of 2026-09-05 found one issuer (NextNav) disclosing 104.50 percent for its warrants and 35.0 percent for its convertible conversion option on the same date in the same filing; a per-issuer class would pool two numbers three times apart. The class is one of six: issuer-observed (the issuer's own model input on its own share), third-party valuation (an independent valuer's input adopted by the issuer, as at Metaplanet), peer proxy (a volatility the filing states was derived from comparable companies because the share lacked history, as at Smarter Web), underlying asset (a volatility of Bitcoin rather than of the share, used to price an instrument struck on the share, as at Capital B), transaction calibrated (a volatility solved so the model reproduces the instrument's transaction price at inception, as NextNav labels its conversion-option input), and unstated. The class is read from the filing's own words and is stated with the accession. Where the filings read do not state a derivation for an instrument, its class is unstated; an unstated instrument scores in its own bucket and pools with no stated class. A derivation sentence found in the share-based compensation note does not class a fair-value instrument; the class read of 2026-09-05 found that trap set in both Holley's and NextNav's 10-Ks and did not fall into it. The interpretation rule in section 3 reads generalization within class; a result pooled across classes is not a result of this study.

Tier population, amended 2026-09-04. The census covered the roster, which contains no issuer with zero digital assets, so tier 3 cannot be populated from it. A tier-3 read is dispatched to the code desk for non-crypto issuers with a disclosed convertible or warrant volatility in a fair-value measurement on EDGAR, three named by the sealed structural rules. That read returned 2026-09-04 and named three issuers from three SIC codes, a selection rule adopted here so that a bottom tier of one industry cannot confuse generalization with sector: Holley Inc. (HLLY, CIK 0001822928, earn-out liability, five disclosed dates), Velo3D Inc. (VELO, CIK 0001825079, warrant classes and contingent earnout, four or more dates, class issuer-observed from the filing's own words), and NextNav Inc. (NN, CIK 0001865631, two instruments: warrants with two dates, and the 2028 Notes conversion option with three, running to 2028). Zero digital assets is tested by word-boundary scan on two filings per issuer, not assumed. Holley's and NextNav's fair-value instruments are unstated on a full-note read of each 10-K (every volatility occurrence read, 22 and 16), and they seal as unstated under the rule above; VELO's are issuer-observed from the filing's own words; NextNav's conversion option is transaction calibrated from its own label. Replacement rule: if a named tier-3 issuer's disclosed series ends before four scored quarters (NextNav's warrants run to expiry inside the window; its conversion option does not), its replacement is taken from the tier-3 candidate file by the same structural rules, dated, with the original retained in the record; the candidate file of 122 issuers is committed to the repo at sealing so the replacement pool exists before it is needed. Within VELO, one warrant class is chosen at harvest and held, never pooled across classes. Eligibility requires a minimum of two disclosed volatility-assumption series per issuer. If any tier still holds fewer than three eligible issuers after the tier-3 read, this section is rewritten before sealing to state what the ladder does with a thin tier, and that rewrite is part of the sealed text. The ladder does not seal on an unread census.

Amended 2026-09-17: MSTR's scored volatility source is the capture rail's listed-expiry implied volatility as governed by the Cal-num-4 champion, class captured-market, a seventh class; the ineligibility rule applies to issuers other than MSTR. The year-to-date grant-date assumption in Strategy's share-based compensation note is recorded and never used as an input.

Amended 2026-09-17: the ladder seals with tier-3 classes as read; an unstated class is an observation the study tests, not a gap to fill by substitution.

Recorded 2026-09-17: NextNav clears the two-series minimum exactly; the replacement rule above stands as written.

3. Absolute gates

The ladder inherits the same absolute gates as Cal-num-4, RMSE 5.0 and maximum deviation 12.0 points, applied per issuer. The interpretation rule reads generalization as passing the gates on named issuers within a vol-source class, not as ranking among them and not as pooling across classes.

4. Seal and owners

External seal per governance rule G-1, same day and same hash as Cal-num-4. Registration now. Harvest of the named issuers' quarterly fair-value marks after naming. Scores when the harvest completes. Thresholds set at ratification 2026-09-03. Two disclosed series minimum per issuer for eligibility. Fifty percent of total assets in digital assets for tier 1 membership.

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