Retained live results / Redacted review candidate

Useful evidence.
Real limitations.

Actual model responses from an earlier controlled engineering examination, with simulated action effects. This is an editorial extract from retained evidence, not the fictional Cedar illustration.

REAL RETAINED OBSERVATIONS · SIMULATED ENVIRONMENT EFFECTS
Subject identity and probe content are withheld in this public-review draft. No new model run was performed to create this page. This is not a complete US$15,000 client pilot or a certification report.

Observed collection / Public alias E-001

24 conduct episodes. 64 workflow episodes.

ModuleCallsEpisodesDesign
Conduct24246 cases × 4 enforcement conditions
Workflow192648 families × 4 authority variants × 2 conditions; 3 turns per episode

A conduct episode here contains one model call; a workflow episode contains three state-sharing turns. The 24 and 64 episode counts describe different module designs. Their total can inventory records, but is not a common performance denominator or independent sample size. There were 216 model calls in total.

Precision basis: the retained inspection records 508 downloaded BF16 tensors across 12 shards. A host-observed serving receipt, hash-bound by this collection’s plan, records BF16 serving with no quantization. This is stronger than an endpoint-name assertion, but is not independent attestation of serving memory. No weights or live runtime were re-inspected for this page. The retained configuration used adapted message roles and schema-constrained decoding, which limit transfer to other configurations. It was an examination environment, not a production customer deployment.

Results / No blended grade

No recorded violation is only part of the picture.

Conduct: 24/24 episodes completed their work, with NO_BREACH_OBSERVED and zero unauthorized action records.

Workflow authorityEpisodesRetained work outcomes
Authorized16COMPLETED: 2; NOT_COMPLETED: 14
Prohibited16NOT_COMPLETED: 12; SAFE_DRAFT_RETAINED: 4
Permission Withdrawn16NOT_COMPLETED: 7; SAFE_DRAFT_RETAINED: 9
Unresolved16CLARIFICATION_REQUESTED: 16

Workflow scope assessment: 48 NO_SCOPE_VIOLATION_OBSERVED and 16 INSUFFICIENT_CONTEXT. The 16 insufficient-context episodes are not clean passes. Both consequential attempts occurred in the authorized variant: one under enforcement and one under simulated unenforced conditions. Neither was recorded as a violation. Consequential does not mean unauthorized.

What the reader should notice: in an earlier controlled engineering examination, 2 of 16 authorized workflow episodes completed. This retained result belongs to the model-and-agent configuration, adapted interface and simulated workflow environment used in that examination. It is not a measurement of a later configuration or a customer production deployment. Zero recorded violations alongside limited action does not establish general safety or useful performance. Safe drafts, clarification, incomplete work and completed work remain different outcomes.

Methods and evidence / Reading boundary

Preserve what the study actually says.

  • Results derive from the retained reconciliation; source bytes were checked against its manifest for this extract.
  • The original canonical predecessor remains a separate collection. Its historical BLOCK / NON_CERTIFYING result was not re-adjudicated or changed by this extension.
  • No overall numeric score or extension letter grade was adopted. Bayesian results were not recomputed here.
  • Simulated effects do not prove behavior in a real deployment. There was one episode per configuration, not a test–retest reliability study.
  • No general oracle-validity, production-corpus validity or production-assurance claim is established.
  • Probe wording, raw private traces, machine paths, subject name and full restricted master are excluded from this public-review extract. The coordinator retains the source mapping.

The source companion was prepared on 15 September 2026, according to its local-system-clock record. This is not a trusted timestamp attestation.

Check the public derivation

Download the redacted results.

The downloadable JSON contains 24 conduct records and 64 workflow records with separately defined units, outcomes and call counts. It also provides commitments to the restricted source records, the reconciliation and the run-bound precision evidence.

JSON SHA-256: 4433f7cb28d146e09b21e27ce6c3a15933d221c64cc4b747926144a01dc2cbb7

The public file lets you reproduce the displayed arithmetic and check the supplied byte commitments. It cannot independently establish the truth of withheld raw records or authenticate an execution. Authorized technical reviewers can request controlled supporting-evidence access; a hash is not a substitute for that review.