# Independent assessment — frozen SOFR documentation executor 01

**Recommendation:** Suitable as a **clearly labeled recorded documentation demonstration with this assessment and finding F-01 visible**. It is not suitable for presentation as an unqualified, fully fulfilled methodology result. The document is substantial and the principal A/B comparison is well supported by the supplied numerical evidence, but it accepts an inconsistent day-count sensitivity interpretation from the source. No institutional validation, regulatory approval, production acceptance or broad agent-reliability conclusion follows.

**Assessment result:** out — completed bounded independent evaluation, with an unmet subject requirement identified. Subject implementation was not repaired.

| Question | Judgment |
|---|---|
| Did execution produce the requested artifacts? | Yes, directly observed: complete twelve-section `document.md`, four supporting records in `reviews.md`, `result.md`, retained inputs, figures and manifests. No artifact-generation failure is apparent. Upstream process completion and recovered tool failures are executor reports, not independently observed tool events. |
| Is the selected agreement wholly fulfilled? | **Partly fulfilled.** Most documentary obligations are met. F-01 prevents unqualified fulfillment of the numerical-meaning/claim-support duties and makes the all-met readiness declaration too strong. This is a substantive interpretation issue, not an omitted artifact or a demand for an unauthorized model rerun. |
| Is the result adequate for its intended use? | Broadly useful to a quantitative reviewer; adequate for a recorded demonstration only when the outstanding finding and evidence limits remain visible. It is not a basis for live pricing or real institutional acceptance. |

## Target, identity and review boundary

Evaluator: Codex delegated agent `/root/sofr_evaluator_01`, assigned by `/root`. The served model identifier and usage are not exposed; none is asserted. First observed clock: **2026-10-12 14:46:29 UTC / 10:46:29 America/New_York**. That is a clock observation, not an invented dispatch timestamp. The assignment permits at most eight minutes from dispatch and requires an absolute stop at 10:55 ET. Completion time is recorded at the end.

Frozen target: `../executor-01/`, specifically the files hashed below. Original executor output path is the temporary directory recorded in its `result.md`; I evaluated the coordinator's frozen copy. Observable worktree branch: `codex/imp-130-sofr-demo`. Ownership here is limited to this assessment directory; no subject, source or shared task record was edited, and no commit or publication was made.

The subject agreement is the public commission at `https://openprose.ai/demos/sofr/v1/program.md`, selecting kernel and Quant release `0.4.6-rc.1`, base requirements, house requirements and fictional institution dossier. I read the frozen agreement copies, including all five Quant definitions and the kernel, full base/house requirements, developer brief, institutional dossier, source register and manifest. Their public origins remain their reference origins; local paths do not relocate adopted definitions. Source revision is `01ae3bdac6f61ecbabcd72e0116ebb2860e93157`; numerical as-of is 2026-09-18, spot 2026-09-22; scenario revision is HER-scenario-1.

I read the full submitted document and all four review records and result. I inspected its three supporting Python scripts without running them, inspected selected financial source functions, read retained numerical summaries and selected receipts, performed independent arithmetic on the frozen rows, and visually inspected all four copied PNG figures. No earlier answer, other evaluator report or private source repository was used. No nested agent, source-script execution/import, direct model API call, credential access, installation, model rerun or publication occurred in this assessment. Own checks use the available Python 3.14.6 standard library and write only here.

Full upstream tool-event coverage was not supplied. I cannot independently establish what the executor actually read, its complete context, absence of forbidden actions, whether its context was isolated, its model identity, or the claimed dispatch-duration compliance from final artifacts alone. Its acquisition times, redirect observations and action history remain executor reports supported by retained manifests, not observations of those HTTP/tool events. Agreement-byte identity can be checked locally; actual enactment is a separate question.

## Findings

### F-01 — Material: day-count sensitivity label is inconsistent with the implemented transformation

**Locations:** `document.md` lines 368–379, particularly Table 11's “Fixed leg interpreted Act/365F rather than Act/360” row and its explanatory sentence; the corresponding assumption in Table 13; `reviews.md` audit A39 and its all-met readiness conclusion. Supporting source: frozen `inputs/model-source/bootstrap.py` lines 287–291, `ql_helpers`/`hw_instruments`, and the par relation in document §3.1.

The code retains Act/360 instruments but multiplies each quote by **365/360**. The document reports that implementation detail accurately. It nevertheless calls the resulting maximum DF differences, 0.00500070 and 0.00500142, the sensitivity to interpreting the unchanged fixed coupon under Act/365F. The source comment asserts the same relationship, so copying it does not resolve the inconsistency.

For a fixed coupon rate `r` over `d` days, Act/365F pays `r*d/365`. The equivalent rate in an unchanged Act/360 instrument solves `r_equiv*d/360 = r*d/365`, giving **r_equiv = r*360/365**, the reciprocal of the implemented scaling. As a scalar check only, a 4.5% coupon over 360 days pays 0.04438356164 per unit notional under Act/365F; the code's scaled Act/360 coupon pays 0.045625. This uses the supplied cash-flow relation; no curve was reconstructed.

Thus the retained numbers support the implemented upward coupon perturbation, not the stated unchanged-coupon day-count switch. Their exact values for the intended switch are **unknown within this assessment**. Because this is a nonlinear recalibration, one must not merely reverse their signs or reuse their magnitudes. The quarterly sensitivity independently remains larger than the A/B maximum, so this finding does not invalidate that separate comparison or the headline USD 227,458 single-payment disagreement.

The commission did not authorize repairing the financial model. The documentation could have fulfilled its conflict-reporting duty by preserving the historical numbers with the literal transformation, marking the intended day-count interpretation unresolved, and carrying that finding into the audit/readiness result. It instead marks A39 supported without qualification. Relevant duties: numerical-evidence meaning and conditions; claim-evidence asserted relationship/conflict reporting; R-05, R-09, R-31, R-40 and R-62. These are partly met where affected. A qualified annotation in a new document would produce new bytes and require a newly identified assessment target; this frozen version is unchanged.

### P-01 — Minor presentation issue: inline mathematical markup needs a viewing check

The source uses ordinary parentheses around TeX expressions, for example `\alpha`, `\log`, subscripts and `8.0\times10^{-10}` in §§3 and 7.4, rather than inline math delimiters. A plain Markdown rendering may display these commands literally. Displayed equations are separately delimited. I inspected Markdown source and figures, not a rendered document in the final presentation surface, so the actual rendering effect is **unverified**. This is a readability/presentation concern, not evidence that the underlying equations are wrong.

## Independent evidence checks

The reproducible local arithmetic records are `independent_checks.py` / `independent_checks.json` and `supplement_checks.py` / `supplement_checks.json`. They do not execute any supplied financial code.

- **Byte identities:** All 73 entries in the submitted artifact manifest match their frozen files. All 61 acquisition-manifest entries match. All 53 present files declared in the public bundle SHA256SUMS match; `presenter.md` is not cached. The latter is optional presenter guidance, not an adopted methodology dependency, so its absence is not a fulfillment defect. All four local figure copies match their input PNG bytes.
- **Structure and notices:** Twelve exact numbered house headings, substantive sections, front matter, three-sentence conclusion, assumptions and control tables, citations, 14 institutional gap markers, verbatim demonstration disclaimer and complete provider notice are present. The advisory word target is not treated as a hard limit. There is no process checklist substituted for the model narrative.
- **Calibration:** Independently compared every Table 5 entry with every one of the 22 CSV rows, both error columns and input rate. No mismatch. Every absolute error is below 0.01 bp; CSV maxima are A 3.86e-9 and B 2.755e-10 bp. The higher-precision JSON numbers support the finer displayed maxima elsewhere. This checks the retained error columns, not the underlying calibration engine.
- **A/B economic comparison:** Monthly grid has 361 rows, 2026-09-18 through 2056-09-18. Largest retained difference row is 2050-04-18, A greater than B. Full-precision JSON difference 0.0022745819731463945 times USD 100 million is USD 227,458.19731463946. The document's rounded USD 227,458.20 and single-cash-flow interpretation are supported. Rounded CSV DFs alone cannot establish cents, and the document says so.
- **Method A implementation comparison:** Recomputed all 7,482 absolute row differences. Each agrees with the retained difference column; maximum is 1.9251267247000214e-13, mean of magnitudes 1.9170748626330293e-14. Exact bounds 2026-09-18 through 2056-09-26 are preserved. Business-day identity follows inspected source/calendar selection; I did not independently reconstruct the SOFR holiday calendar. This is correctly distinguished from A/B method disagreement and from institutional independence.
- **Residual populations and statistics:** Independently counted 523 input-tenor spot prints, 140 other-maturity spot prints and 320 forward-start prints, hence 460 held out. Upper-middle absolute medians reproduce A 0.4599424411 and B 0.4046620120 bp; both meet the selected 1 bp criterion. Ordinary medians differ as disclosed. All ten historical signed-summary cohort/method results and all twelve paired-comparison rows, including p-values and robustness selections, independently reproduce from the supplied residual CSV with the inspected conventions. B/A closer counts are 269/189 plus two ties. This does not create independent market observations.
- **Distance bins:** Recomputed all six cohort/bin counts, medians and inclusive IQRs from expiration-to-nearest-OIS-payment-pillar calendar-day distances, excluding deposits. All match the retained JSON and displayed rounding. The document preserves 0–30, 31–90, >90 boundaries and keeps cohorts separate; it identifies B's worse nearest-bin forward-start median and does not call distance a risk or independence measure.
- **Shape/locality:** From rounded retained forward CSV, verified the A/B maximum daily moves, counts above 2 bp, and five-business-day moves at displayed precision. Inspected the source's grid construction through the last node, positivity collar and inactive-collar result, solver guard, extrapolation and plotting code. Bump maxima strictly outside the inclusive 4Y–6Y node window reproduce 0.0305/4.1475 bp. Absolute ladder shares reproduce 5.86%/11.94% after rounding. The default QuantLib blend and failed paper-setting solver attempt are correctly distinguished; the failure is retained, not hidden.
- **Precision limit:** Summing changes from the rounded A forward CSV gives 421.3685 bp rather than full-precision JSON 421.3288736704 bp. Accumulated rounding can explain that gap: a conservative bound from 7,481 differences and six-decimal-percent rounding is 0.7481 bp. I therefore do not call this a contradiction or claim exact reproduction of the unrounded total variation. The document correctly uses JSON precision.
- **Figures:** Four observed PNGs have readable axes, units and legends; captions supply historical date, calculation/data identity and relevant interpretation limits. The required discount-factor, forward and bump/ladder content is present with local relative references. I did not regenerate figures or numerically recover every plotted pixel.

## Coverage of the agreement and supporting records

| Selected scope | Independent assessment |
|---|---|
| R-01, R-02, R-03: purpose, use/exclusions, construction | Met on inspected content: plain-language explanation, single-curve scope, lagged cash-flow relation, dates, time basis, A/B construction, solver and extrapolation account are substantive. No attempt here to prove general interpolation properties. |
| R-04: inputs and provenance | Met as documentation of supplied evidence. Real fixing, reported prints, derived estimates and scenario fiction are distinct; raw-source truth is not newly verified. |
| R-05: retained tests | Partly met because F-01 misinterprets one scenario; other checked tests are accurately labeled historical, with all calibration errors supplied. |
| R-06: outside-input and implementation comparisons | Met on the extensive numerical checks above; signs, populations, shared adjustment/schedules and lack of licensed benchmark are explicit. |
| R-07, R-08: limits and controls | Met: one date, sparse long end, locality, common adjustment, extrapolation, missing event steps, calendar and stress limits; proposed fictional controls and absent effectiveness are distinguished. |
| R-09: assumptions | Partly met: substantial reason/evidence/consequence/monitoring register, but day-count consequence inherits F-01. |
| R-10, R-11: fictional monitoring/governance | Met as scenario documentation; pending approval, unsigned block and actual-evidence gaps preserved. Fiction does not establish real controls. |
| R-12, R-13: components and retrievable evidence | Met as historical identities and public-origin references. Selected external paper/code assertions remain attributed and explicitly unchecked. |
| R-20: every principal/smaller choice | Seven principal and 27 smaller entries are represented, with reasons, alternatives and limits. I compared the full brief's choice inventory with the register and document destinations; no missing choice identified. F-01 remains an interpretation finding, not an omitted choice. |
| R-21: principal comparison | Met: specified single payment, signs/units and smoothness/locality tradeoff supported. No profit, superiority or portfolio-loss inference. |
| R-30: reproducibility and identities | Met as supplied provenance and scope explanation; raw archives and original external papers unavailable locally, no offline raw-data rerun claimed. |
| R-31: measure traceability | Mostly met, partly affected by F-01's scenario meaning. Dates, grids, units and scripts otherwise extensively supplied. |
| R-40: factual claim relationships | Partly met: a substantive audit and generally precise support, but the source/meaning mismatch in A39 was not surfaced. |
| R-41: permitted sources | Compatible with observed artifacts; actual full acquisition behavior cannot be established without upstream tool events. External contents are not asserted as independently verified. |
| R-50, R-60, R-61: institutional gaps and readiness distinctions | Met in document presentation: scenario labels, 14 real gaps, disclaimers, executor versus independent assessment and actual authorization kept separate. |
| R-62: requirement-by-requirement report | Structurally met, judgment partly met: every base requirement is mapped, but all-met/full-fulfillment declaration should be qualified by F-01. |
| Supporting section plan | Full requirement mapping with document/report locators, not merely an outline. |
| Supporting choice register | 34 identified entries with reason, alternative/explicit absence, limitation and evidence; substantive. |
| Supporting citation audit | 45 checks plus explicit unchecked/unavailable categories. It goes beyond a citation list; F-01 demonstrates its limitation and prevents blanket acceptance of every “Supported” conclusion. |
| Supporting readiness/result | Exact artifact identities and unresolved broader work recorded; self-readiness is correctly labeled as executor work. Its completeness judgment is too strong in light of F-01. Process/boundary claims remain reports rather than independently observed events. |

## Remaining uncertainty and demonstration use

I did not acquire external primary papers or versioned QuantLib originals; verify their substantive assertions separately if the demonstration will characterize them as checked facts. I did not recover the omitted raw DTCC/NY Fed data, reconstruct any model, independently build calendars, test additional dates or regimes, validate actual institution records, or audit every line of every source script. Code inspection was targeted to material method, population, sign, tolerance and scenario claims. Arithmetic checks have full row coverage only where explicitly stated above. Historical environmental receipt contents identify supplied historical runs, not proof that those runs were independently reproduced today.

The exact document can be shown with its frozen artifact hash, its own disclosures, and this assessment's F-01 clearly visible. Describe it as a recorded documentation execution subjected to this bounded assessment, with one unresolved material interpretation finding. Do not present its “fulfilled” line without the assessment qualification. Any source correction, new numerical run or document revision should be separately authorized, retain this evidence and receive its own identity and assessment. Real institutional evidence and authorization remain outside this demonstration.

## Exact byte identities

All paths below are relative to the frozen `../executor-01/`. `independent_checks.json` contains SHA-256 values for **every frozen file**, including all selected definitions, all numerical CSV/JSON, scripts, figures and manifests.

| Artifact | SHA-256 |
|---|---|
| document.md | `16d7918bc07263bef7f90b9ed2aa9f407ff01a3a28da4c3085c081f87ddca2bf` |
| reviews.md | `17c4ea54770d26f7fe4c04cc0ce4fb2b942d253fcbe10382b67ddf70c264cde8` |
| result.md | `23fdb6889fa34cec8bfe43c34375ef3d1ffe5cbb48dfb413f076fb33d6a977a0` |
| artifact-manifest.json | `bc916154450fdba44381bc0781839d2be2a56927f75d18cdda41d87cf10e9d63` |
| input-manifest.json | `01a1ddb18fb620fd285715c280a44927aacb281730f39a35e503f486008945f1` |
| inputs/program.md | `ad458104fe99c7e1b0683558dd4cc8843d4756ba902dd7b7b9534a96cca859f9` |
| inputs/kernel.md | `4622c21aa110d36abc9262496f922e578ecd9eb20e87241b28b2060d06df36cb` |
| inputs/model-brief.md | `a3262b78cd21a770b4c446508d1cb681610f2ab760bc23bbff35bf96b5333c74` |
| inputs/requirements.md | `9e0965eb4e8c9c13ad9383049a2377cd053e39bc575a99b693ed882144b4a8bb` |
| inputs/house.md | `85f083dae309d17a4862a27997a56b58ffac5ed7bc95645d8e6c3f2ae8b3c783` |
| inputs/model-source/bootstrap.py | `d897203d1b145f81b60248df422ef79411362a212098e0408a05fb757f3e56b2` |
| inputs/evidence/repricing.csv | `70d1e851114cc2326121fd615dd4da158bd505e12cf29afdb5adf9f791a63f5e` |
| inputs/evidence/oos_residuals.csv | `28dc38fafd4ca94665217a792eb414d0dca3f090b4b0a7a0244f3b0f6214b58e` |
| inputs/evidence/implementation-comparison.csv | `06408929119a4b61259e9114e57951940f204e25068def2099e1727281b550da` |

**Assessment completed:** 2026-10-12T14:52:25.602591+00:00 UTC (America/New_York is UTC−04:00 on this date). Frozen file hashes rechecked at completion: no changes from the assessed target. Completion is before the absolute stop and under eight minutes after the first observed clock; an exact dispatch timestamp was not supplied to this evaluator.
