This page is the rulebook for a measurement test that has not been run yet, published in full before any candidate is scored against the data.
The test asks whether a formula the CEBE (Common Equity Bitcoin Exposure) framework uses to value convertible bonds can reproduce what those bonds sold for. A first attempt, Calibration 1, was published on 2026-07-28 and failed the standard set for it in advance. Nothing shipped as a result, and the figures the formula would have produced remain unpublished. This is the rulebook for the second attempt.
It publishes now, rather than alongside the results, for one reason. A standard chosen after seeing the answer is not a standard. Everything on this page, including the pass thresholds, the sourcing rules for every input, and what happens on failure, is fixed and dated before any candidate is scored. Any later change lands as a dated revision against this published text, with the original preserved, so a reader can always see what moved and when.
The pass thresholds are unchanged from Calibration 1. Only the method and the inputs change.
Read the first attempt and its result: Calibration 1: Adjusted Claims Conversion Weight.
Terms expanded on first use below: Adjusted Claims Percent (ACP), Market-Adjusted Bitcoin Exposure (MABE), implied volatility (IV), root mean squared error (RMSE), mean absolute error (MAE).
Specification v1 section 8 holds the continuous conversion weight of section 3.1 behind a calibration gate. No computed conversion-weighted figure appears on any tracker surface until a calibration demonstrates, in public, that traded convertible instruments price within stated tolerance of the model's implied values, with the tolerance stated before results.
Calibration 1 ran under that gate and no candidate cleared. The no-pass clause fired, the gate held, and the continuous conversion weight stayed unlaunched. That disposition stands and nothing in this document reopens it.
This document fixes the rules for the second attempt. It publishes tonight, before any candidate output is computed. Every capture already made, and the circumstances of each, is disclosed in the companion's sections 3 and 6.6. The marks panel itself is fully known from Calibration 1, which is why the tolerance carries forward frozen rather than restated: familiarity with the data is answered by a standard that cannot move, not by a claim that no one has looked.
The companion volatility sourcing document at section 3 was ratified on 2026-08-07 and is reproduced here without edit.
The tolerance ratified for Calibration 1 carries forward without alteration. Verbatim, as stated blind on 2026-07-28:
Root mean squared error and mean absolute error are expanded here on first use; the quoted text is reproduced as issued.
The three numbers do not move. Calibration 2 changes the price map and the inputs. It does not change what counts as passing. A tolerance that loosens after a failure is not a tolerance, and the record would show it loosening.
Held-out error is leave one date out. Each measurement date is withheld in turn, fitted candidates are refit on the remaining dates, and the withheld date is scored against that refit. Candidates carrying no free parameters are scored directly.
Calibration 1 ran on Map A, the ratified section 3.1 blend, which carries no time value. That study established a result that governs the design of this one: because every observed mark prints above its own conversion parity, and because the blend can only produce values between par and parity, the lowest error any weight function in the interval zero to one can achieve on that panel is 14.07 points of face. That floor binds every candidate equally and exceeds the 5 point threshold by nearly three times.
Calibration 2 therefore tests a construction that can carry time value:
Each candidate weight function is evaluated inside that construction rather than inside the par to parity blend.
Ratified 2026-08-07 and stated in full in the companion volatility document at section 6.1.
Scored cohort, forward only. Candidates 1, 1b, 4a and 4b are each defined in terms of implied volatility. They are excluded from every scored historical mark and constitute a forward cohort beginning at the first per-expiry option chain capture made under the complete rule set, expected 2026-08-14.
Historical scored set. Candidates 2, 3a and 3b remain the scored historical set, as in Calibration 1.
Back-dated groundwork, unscored. Single-tenor arithmetic on the twelve captured dates may run as unscored exploratory output only, under its accurate label, a fixed-tenor prior at a stated tenor, never described as term-matched. It produces no panel figure, enters no tolerance test, and carries no disposition. Its sensitivity table publishes beside it.
The two regimes report separately and never merge into one coverage figure.
Each input is either in hand or pending capture under a locked rule. Pending capture is not an open item: the rule governing the input is fixed, only the value is outstanding.
| Input | Rule | Status |
|---|---|---|
| Implied volatility series | Companion document, sections 1 through 6 | Rule locked, forward capture pending |
| Risk-free curve at each measurement date | Sourced per date, source named | Pending capture |
| Total bitcoin held and total gross claims | Sourced per date | Pending capture |
| Conversion prices, 2025 and 2027 notes | Traced to indenture, adds 8 observations | Pending capture |
| Refreshed traded print | Q2 2026 10-Q exact principal and cash price, replaces the EST flag | Pending filing |
| Rotation-suspect rows, 2025-06-30 | Filing-level re-verification by the data lane | Pending |
| Disclosed convertible aggregate | Carried as its own column with row label and accession per date | Pending capture |
The last item exists because Calibration 1 could not re-run the aggregate reconciliation downstream: the aggregate was present only as prose in a source log, and the Q1 2026 schedule read back at accession 0001050446-26-000031, report R44, carries no aggregate row for principal or fair value at all. Carrying it as data makes the check reproducible by any reader.
These were ratified on 2026-07-28 and nothing has superseded them. They govern Calibration 2 unless a dated revision against this page says otherwise.
| Ref | Ruling |
|---|---|
| Gate test | Specification section 8's two-sided price test stands as written. |
| Fitted versus prior scoring | Leave one date out, held out only. Information criteria report as context and bind nothing. |
| Prior deference | The absolute tolerance is the only launch gate. Beat-the-prior reports separately and guides design, never disposition. |
| Prior margin | A challenger earns the label "outperforms the published schedule" at 10 percent or better held-out RMSE improvement. No automatic replacement. |
| Candidate 1 reading | Both N(d1) and N(d2), scored separately. |
| Candidate 3 reading | Specification section 3.1 reading for the price fit. The claims-line reading is documented, not fitted. |
| Rotation-suspect rows | Quarantined from the primary panel, dual run as a labelled sensitivity. Filed data is never edited on a hypothesis. |
| Band boundaries | Left closed, right open, matching the published schedule's explicit "at or above 130 percent". |
| Evidence classes | The tolerance binds on the disclosed marks. Traded prints bind separately and report as the out-of-sample check with their caveats stated. |
Fail. The no-pass clause publishes verbatim with the structural diagnosis, and the continuous conversion weight stays unlaunched.
Pass. The gate opens for that candidate. Launch remains a separate ratification; the gate is necessary, not sufficient.
Multiple passes. Fewest free parameters wins, ties broken by held-out RMSE.
This page carries no open items. Three things that might read as open are not, and the distinction is stated so the claim can be checked rather than taken:
Calibration 1's disposition is unaffected by anything on this page. No candidate cleared, the no-pass clause fired, the gate holds. The 14.07 point joint error floor binds every weight function in the interval zero to one under that study's ratified price map, so the four volatility-dependent candidates could not have cleared with or without a volatility series.
Four findings routed from that study to a specification v1.2 discussion and remain there: the section 3.1 blend's inability to carry time value, the collapse of continuous weighting to the binary case under that price map, the two-object finding on the published schedule against section 3.1, and the direction of the error.
The complete pre-registration record for Calibration 1, including two rulings that were stamped twice and the corrections to both with the originals preserved, publishes at the Calibration 1 ruling log.
Reproduced below without edit, exactly as ratified on 2026-08-07 and passed by the canon check. Source file R-VOL-pre-registration-2026-08-07-RATIFIED.md, SHA-256 prefix 33a2ea3c1e337802. Heading levels are preserved as passed.
This pre-registration fixes how the volatility input is obtained, before any value is scored. It governs candidates 1, 1b, 4a and 4b, which are each defined in terms of sigma and which did not run in Calibration 1 because no volatility series existed.
Sourced implied volatility. A value published by a market surface and captured with evidence.
Never fitted. Sigma is not a free parameter and is not solved for. Fitting it would destroy the zero-parameter standing of candidates 1, 1b and 4a and would void the cap that defines candidate 4a.
Never realized. Volatility computed from a price series is excluded as an input. This exclusion covers platform indicators labelled "historical volatility", "HV", "statistical volatility" or any equivalent, whatever the surface calls them. A label naming volatility is not sufficient; the value must be implied, extracted from option prices, or it does not qualify.
Never assumed flat. A single value carried across tenors is not a term structure. Section 6.2 authorizes a fixed-tenor construction for unscored groundwork only, under its own label, and no scored value is ever produced by carrying one tenor across others.
Never padded. A date that fails its sources is disclosed as uncovered. It is not filled from a neighbouring date, a neighbouring tenor, or an average.
Back-dated coverage. Three surfaces, in the hierarchy established by revision 1 of 2026-08-07, after the coverage basis was established on that date. The pre-revision language is preserved in the revision log.
| Role | Surface | Dates carried |
|---|---|---|
| Primary | Market Chameleon | 12 of 12 |
| Cross-check | Barchart | 6 of 12 |
| Corroboration | AlphaQuery | 2 of 12 |
Each value is cited per date with capture evidence. Each vendor's construction is disclosed per series, per section 6.4. Values are never blended across vendors.
Forward coverage. Per-expiry implied volatility captured from the MSTR listed option chain at each Friday close, alongside the issuer dashboard print where readable. Timestamped and screenshotted.
No per-expiry chain capture has been made as of ratification. The rows dated 2026-07-24, 2026-07-31 and 2026-08-07 are single-tenor forward-context rows within the twelve, carrying the same construction as the back-dated values. They are not chain captures and are never described as such. The first per-expiry chain capture made under the complete rule set is expected 2026-08-14.
Every value carries: the surface, the retrieval timestamp, the tenor as published, the value as published, the vendor's stated construction, and a screenshot. A value without capture evidence is not a sourced value and does not enter.
Interpolation is linear in total variance, not in volatility. Given sourced points (T1, sigma1) and (T2, sigma2) that bracket the target term T, so that T1 is at or below T and T is at or below T2:
Where more than two sourced points exist, the two nearest bracketing points are used and the construction is piecewise linear in total variance across the sourced grid.
This is the only arithmetic applied to a scored value, and it executes only where sourced points bracket the target tenor. No extrapolation is authorized on anything scored.
Why this form and not the obvious one. Linear interpolation in sigma does not preserve the monotonicity of total variance in time and can produce a term structure that admits calendar arbitrage. Linear interpolation in total variance cannot. The two forms also differ materially in size: on a term structure falling from 60 percent at three months to 55 percent at one year, the value at nine months is 55.58 percent under total variance and 56.67 percent under linear sigma, a gap of 1.09 volatility points. That is large enough to move a weight and therefore large enough to need pinning in advance. Those figures are a synthetic illustration of the formula, not sourced values.
Units: T in years on the same day-count as the sourced tenors, which are converted from published day counts by dividing by 365.25. Both sourced points must share a strike convention with each other and with the target, per section 6.4.
Ratified 2026-08-07. The three items that previously blocked the freeze are resolved by 6.1, 6.4 and 6.5 respectively.
The panel's required tenor exceeds what back-dated public implied-volatility history publishes. Across the 42 panel observations, term to investor put runs from 1.459 to 4.958 years, median 2.960, and term to stated maturity from 2.461 to 7.959 years, median 4.917. Covering 80 percent of the panel on the put leg needs a surface publishing to 3.959 years, and on the maturity leg to 6.404 years. Back-dated capture is single-tenor, so no back date carries a bracket at all.
The ratified response splits the candidates into two cohorts.
Scored cohort, forward only. Candidates 1, 1b, 4a and 4b are excluded from every scored historical mark. They constitute a forward cohort beginning at the first per-expiry chain capture made under the complete rule set, expected 2026-08-14. Within that cohort the pinned interpolation of section 5 is the only arithmetic applied, and it runs only where sourced points bracket the target tenor. Where they do not, the value is uncovered and publishes as uncovered.
Consequence, stated rather than left to inference. Calibration 2's historical panel carries no sigma-dependent candidate. Candidates 2, 3a and 3b remain the scored historical set, as in Calibration 1. The four sigma-dependent candidates enter the record only as forward observations accumulate.
The identical single-tenor arithmetic may run on the twelve captured dates as unscored exploratory output only. It produces no panel figure, enters no tolerance test, and carries no disposition.
It is carried under its accurate label: a fixed-tenor prior at a stated tenor. It is never described as term-matched, because it is not. Carrying one sourced tenor across a remaining term of one and a half to five years is a fixed-tenor construction whatever the intervening step is called, and the label follows the arithmetic rather than the intention.
The sensitivity publishes beside it, always. Function-space arithmetic on the panel's actual terms to put, no sourced value involved and no candidate scored:
| Flat sigma | Candidate 1 mean weight | Marks above 0.9 | Marks below 0.1 |
|---|---|---|---|
| 0.40 | 0.546 | 4 | 4 |
| 0.55 | 0.607 | 3 | 2 |
| 0.70 | 0.665 | 1 | 0 |
| 0.85 | 0.717 | 1 | 0 |
| 1.00 | 0.764 | 4 | 0 |
| 1.20 | 0.818 | 9 | 0 |
Mean weight swings 0.272 across that range. At MSTRFV-051, the 2029 at 2026-03-31, moneyness 0.1856 and 2.171 years to put, N(d1) runs from 0.005 at a flat 40 percent to 0.473 at a flat 120 percent. A convert at 19 percent of its conversion price traverses most of the weight range on the choice of the sigma level alone. Any reader of the groundwork output sees that before seeing the output.
The two regimes report separately and never merge into one coverage figure. A single number pooling bracketed forward values with fixed-tenor back-dated values would report a precision the back-dated half does not have.
The 80 percent floor counts against the nine fitting-panel measurement dates. Nine of nine carry a primary value. The floor clears.
The rows at 2026-07-24, 2026-07-31 and 2026-08-07 are single-tenor forward-context rows outside the floor's base and are labelled as such. They are not chain captures. Nine panel dates plus those three forward-context rows reconciles the twelve dates the capture lane reports.
Relationship between the floor and the cohorts, stated so it cannot be misread. The nine dates on which the floor clears carry only the unscored groundwork of section 6.2. The floor clearing is a statement about the completeness of that groundwork. It is not a warrant to score a sigma-dependent candidate on any historical mark, which section 6.1 excludes independently.
The target quantity is strike-matched implied volatility at K equal to the conversion price. The candidates take K as the conversion price, so that is the strike whose volatility belongs in N(d1) and N(d2). Anything else is a proxy.
All vendor series are disclosed proxies. Each vendor's construction publishes with its values: put-call mean, delta-50, or unstated where the vendor's own methodology page does not cure it. "Unstated" is a published value of that field, not an omission.
Forward-cohort strike selection is the nearest listed strike, with the moneyness distance to K disclosed per value. A reader can therefore see how far each captured value sits from the strike the candidate requires.
Smile interpolation is prohibited. Interpolating across strikes to synthesise a value at exactly K would be a second piece of arithmetic applied to sourced values, and section 5 admits only one.
Corroboration only. Divergence between surfaces is disclosed per date in volatility points.
Values are never blended and never excluded for divergence. Divergence is evidence about the sources and is published as such. It does not license a choice made after the values are visible, which is the failure mode a disagreement rule exists to prevent.
The twelve historical values were captured before rules 6.4 and 6.5 were pinned. They were therefore obtained without a fixed strike convention and without a fixed divergence rule.
This is the reason those values are confined to unscored groundwork under section 6.2 and can carry no scored output. A value captured before the rules governing its capture were written cannot be retrofitted into a pre-registered result, and no attempt is made to do so.
The same disclosure governs the three forward-context rows. They are single-tenor rows captured before rules 6.4 and 6.5 were pinned, they are not per-expiry chain captures, and they carry no scored output either. The forward cohort begins at the first per-expiry chain capture made under the complete rule set, expected 2026-08-14. No scored sigma-dependent value exists before that date.
Coverage is counted per required value, which is per mark per candidate, since a single date carries up to six marks at six different remaining terms.
Uncovered is a published state with a count, not an omission. Forward-cohort values falling outside the bracket of the sourced grid are uncovered and publish as uncovered rather than extrapolated.
The two regimes report separately, per section 6.2. No pooled figure is published.
Nothing. Calibration 1's disposition did not depend on the volatility series and does not depend on it now. The joint error floor of 14.07 points binds every weight function in [0,1] equally under the ratified price map, so the four sigma-dependent candidates could not have cleared the tolerance with or without sigma. That is stated in the Calibration 1 write-up and stands.
What the volatility series buys is relative ranking among candidates and the groundwork for Calibration 2's price map. It was never going to reopen the gate.
Revision 1, 2026-08-07. Source hierarchy. Superseded on establishing the coverage basis across the twelve captured dates. Prior text preserved verbatim:
Back-dated coverage. Two independent public implied-volatility history surfaces:
| Role | Surface |
|---|---|
| Primary | AlphaQuery |
| Cross-check | Barchart |
Each value is cited per date with capture evidence. Independence is the point of the pair: two surfaces deriving from the same vendor feed would be one source with two skins, and the capture lane confirms independence rather than assuming it from the brand names.
The independence requirement in the preserved text is not superseded and continues to apply to the revised hierarchy.
CEBE framework, cebetracker.io. Part of the calibration record. Companion to the Calibration 1 ruling log at /claims/calibration-1/ruling-log/. The gate this pre-registration answers is Specification v1 section 8.