Calibration 1 and Calibration 2 both failed. This is the third attempt, and what makes it different is that the rulebook now names a single construction with its parameters already fixed, rather than a family of candidates to be fitted at scoring time.
A search over 3,578 construction variants selected that specification. Everything the search touched is historical data, and historical data cannot both choose a specification and vouch for it. So the specification was written down, its parameters frozen, and a set of filed values that no part of the search has ever read was left sealed. That sealed set is the test.
This page is the rulebook. It carries no verdict, because the verdict does not exist yet.
Terms expanded on first use below: Adjusted Claims Percent (ACP), implied volatility (IV), root mean squared error (RMSE).
The specification cleared admission on the historical panel. That sentence is easy to misread, so here is the plain version.
The historical panel qualified the specification for scoring. It is not a result. Admission was computed on the same 41 filed marks the search consumed when it chose this construction, so it shows that the specification is well formed and prices that panel within tolerance on a live option leg. It shows nothing about data the specification has never seen. No out-of-sample result exists.
Publishing a registration that carries no result is not an omission. It is what a pre-registration looks like from outside: the criteria are fixed and public first, and the verdict arrives afterwards or the exercise is worthless.
Model value to face, in points of face, for one convertible series at one measurement date:
The bond floor discounts coupons and principal to the investor put date, semi-annually, at the US Treasury par yield for the nearest published CMT tenor plus a credit spread. The conversion option is a European call on conversion parity struck at par, expiring at that same investor put date, so both legs price the instrument to one terminal date.
The conversion weight w occupies the N(d1) slot of the option and that slot only. N(d2) is left as Black-Scholes computes it, which is what keeps the sourced volatility inside the price and leaves the option leg holding time value. Calibration 2's ruled coupling placed the weight in both slots, and those are the only two terms through which a European call's volatility enters, so that construction removed volatility from the price entirely.
The weight responds to standardised log-moneyness, not to raw moneyness:
Raw moneyness carries neither volatility nor tenor. A weight standing in for an exercise probability has to respond to both, and σ√T states them exactly.
Volatility is the sourced 30-day implied volatility at each mark, scaled to the option's tenor by the at-the-money term-structure ratio measured on the 2026-08-14 option chain. Only the ratio travels; the level at every mark stays pinned to that mark's own sourced figure.
| Parameter | Value | Role |
|---|---|---|
| A | 0.877938 | slope of the weight in standardised log-moneyness |
| B | 0.816968 | intercept of the weight |
| SPREAD_ADD | 8.803434 | percentage points added to the published B band spread |
Fitted once, by least squares in price space, on all 41 marks of the Calibration 1 historical panel. They do not move. The confirmation run applies them as given and does not refit for scoring.
Because the widening is fitted, the specification is exactly insensitive to the B band's level and uses only its variation through time.
The scored protocol is leave-one-date-out. Each measurement date is withheld in turn, the specification is refit on the remaining dates, and the withheld date is scored against that refit.
Those three numbers carry forward from Calibration 1 and Calibration 2 unchanged. This registration does not move them.
Admission is the scored protocol plus the degeneracy constraint of section 5, and nothing else. Every other protocol is computed, published, and gates nothing.
| Protocol | Role | Historical panel | Against 5.0 / 12.0 |
|---|---|---|---|
| Leave-one-date-out | admission, scored | 4.711 / 9.60 | PASS / PASS |
| Live option leg | admission, structural | 1.732 / 1.882 / 0.210 | PASS |
| Leave-one-series-out | disclosed diagnostic | 4.830 / 10.14 | PASS / PASS |
| Expanding-window forward | disclosed diagnostic | 4.746 / 11.77 | PASS / PASS |
| Leave-two-dates-out | disclosed diagnostic | 4.826 / 14.94 | PASS / FAIL |
| Half-sample split | disclosed, no tolerance | 7.308 / 13.59 | not applicable |
This structure was chosen with every one of those numbers already in hand, including the failure. Saying so is the point. Three things follow, stated rather than left to be inferred.
An earlier draft of this registration contradicted itself. It made admission the worst of four protocols, under which the leave-two-dates-out failure at 14.94 meant the rulebook rejected its own candidate on the data used to select it. That contradiction was found and raised before the seal was touched.
A strict worst-of-four gate admits nothing sound. Across the 3,578 variants, six clear such a gate on both tolerances and all six fail the degeneracy constraint. Zero clear it while keeping a live option leg. A gate passable only by a construction whose option leg has stopped being an option is not a quality bar.
The operative test is therefore the sealed set, not this panel. Every number in the table above was computed on data the search consumed. That panel selected this specification and cannot also vouch for it.
The leave-two-dates-out failure is named here as the known risk and is not hidden behind a relaxed threshold. Its status is fixed in both directions before the seal breaks: it cannot be restored to gating if it passes, and its 12.0 threshold cannot be relaxed if it fails.
Clearing the numbers is necessary and not sufficient. The specification must also keep an option leg that behaves like an option.
This matters because a convertible construction can stop pricing optionality by two routes that look nothing alike and are identical in effect. Calibration 2 reached it through the coupling, by placing the weight in both N slots. This calibration's search reached the same place from the opposite direction: given a free multiplier on the sourced volatility, the optimiser drove it to whatever ceiling it was given, and at a high enough level the option leg saturates onto conversion parity and volatility stops reaching the price. Those constructions score well and price nothing.
| Measured case | Volatility-price sensitivity | Mean σ√T | Mean N(d2) |
|---|---|---|---|
| Calibration 2 coupling, volatility absent by construction | 0.000 | 1.39 | not applicable |
| A saturated search leader, volatility multiplier 2.56 | 1.148 | 3.56 | 0.075 |
| This specification | 1.732 | 1.882 | 0.210 |
| Plain Black-Scholes at the sourced volatility | 2.660 | 1.39 | 0.325 |
Registered floors: volatility-price sensitivity at or above 1.00 points of face, mean σ√T at or below 2.00, mean N(d2) at or above 0.05. A specification clearing the scored protocol on a construction that fails any of these is reported distinctly from a pass, and is not admitted.
Of the three parameters, the credit widening is much the loosest. Across the nine leave-one-date-out folds its standard deviation is 0.919 points, against 0.071 and 0.051 for the other two.
Derived as the nine-fold mean plus and minus three fold standard deviations, rounded outward. Twenty-five independent refits span 7.084 to 11.144 and sit strictly inside the band. In total-spread terms this asserts, in advance, that this issuer's convertibles price wider than the CCC index across the whole plausible range.
If a diagnostic refit on the confirmation panel falls outside that band, it is a failure of this specification and publishes as one. It is not a data surprise and not an occasion to widen the band. The direction is registered too, so an out-of-range value can be read: the widening moves inversely with the sourced volatility, so a value above the band is the signature of an understated conditioning volatility, while a value below it has no ready explanation and is the more serious case.
Stated here because a registration that publishes only its strengths is not a registration.
The confirmation set holds a tenth measurement date of filed convertible fair values that no exploratory run has read. Its state is committed and auditable, and its provenance is recorded in full, including the part that is less than perfect: the 2029 repurchase print inside it has already been scored twice, under Calibration 1 and Calibration 2, so it is carried as a demoted secondary observation and is not a virgin cell. The Q2 2026 fair values are untouched.
When the seal breaks, the confirmation run applies the frozen specification without refitting, scores the scored protocol at 5.0 and 12.0, publishes the disclosed diagnostics against their stated tolerances, applies the degeneracy floors, and tests the widening against its registered range.
The result publishes either way, pass or fail, following the Calibration 1 and Calibration 2 precedent. If the specification fails, that finding publishes and the section 8 gate holds. No threshold, protocol role or parameter on this page may be moved after the confirmation numbers are seen.
Full registered text, including every date rule, tenor convention and mechanism citation: docs/specs/cal-3-registration-2026-08-19.md in the CEBE Tracker repository, ratified text at blob 0b5e95f23249f54d158c57eb7c58b7e2dacf2d69.
CEBE framework, cebetracker.io. Part of the calibration record. Preceded by Calibration 1 and Calibration 2, both of which failed. The gate this registration answers is Specification v1 section 8.