# Eve SOTA Task 10.4 — current V1.0 k=10 measurement

Date: 2026-09-12

Initiative: `eve-sota-gap-closure-2026-09-01`

Task leg: `10.4-current-release-leg`

## Outcome

The exact seven release-blocked cases were rerun at k=10 through the production
assistant eval route under the authoritative V1.0 `current` expectation scope.
The registered OpenRouter binding used `deepseek/deepseek-v4-flash-0731`; Baidu
served all reported turns. The run finished 55/70, with zero provider retries
and zero observed forbidden-tool failure runs. It billed $0.0084, had a
4.9-second median turn latency, and reported 643,584 cache-read tokens over
680,407 prompt tokens on 68/70 runs.

| Case                             | Passes |
| -------------------------------- | -----: |
| `veritas-tool-selection`         |   9/10 |
| `family-md-empty-saved-articles` |   2/10 |
| `family-md-course-continue`      |  10/10 |
| `family-md-metis-search`         |   6/10 |
| `family-md-metis-recommend`      |  10/10 |
| `family-md-empty-course-search`  |   8/10 |
| `family-md-empty-recommend`      |  10/10 |

This is an observational advisory result, not a promotion. Failed runs remain
failures. The prior advisory reason was not edited, full-deck/family/builder
floors were not evaluated, and no floor moved. The run used a seeded, run-unique
PostgreSQL database and retained proof of exact teardown without retaining its
identity.

## Retention and controls

The receipt retains the exact case counts, release/mode, provider/model/serving
provenance, spend, cache use, latency, retry count, source hashes, and database
lifecycle. The log retains aggregate grader diagnostics but no prompts, model
response excerpts, credential, database URL, or run-unique database identity.
The provider key remained in process memory only and the external secrets file
remained mode `0600`.

The verifier recomputes receipt arithmetic, compares every per-case result and
run header against the sanitized log, checks source hashes and the shipping
release constant, scans the log for forbidden retained data, and verifies the
advisory/no-floor/future-release boundaries. Eleven semantic mutations are
retained as observed-red controls.

## Remaining release gate

V1.2 has not opened. Accordingly no restored-semantics provider run was made and
no annotation was promoted or retired. Task 10.4, Phase 10, and G8 remain open
until the product release changes to V1.2, the exact restored semantics are
rerun at k≥10, their required floors pass, and the owners record the explicit
annotation decision.