AI Council — Public Operator Manual

Current Council engineering status — 25 October 2026

Provider roster/identity admission regression passed 7/7 tests. Genuine task execution admitted one local creator and zero independent reviewers, so the target 25-seat Council is not qualified. A model-selected public cutblock dataset was fetched, but the original compiler rejected an unconnected required dataset socket. Studio system-import reconstruction and adaptive DataChip invention still require real independent full-cycle proof.

Updated Council qualification manual · Current evidence

Audience and access

This user-facing manual is for research sponsors, authorized FarmNode operators, independent reviewers and governance staff. It explains decisions and evidence that an authorized product should expose; it does not publish internal commands, access tokens, private code or administrator setup instructions.

1. Identify a real research request

Record which laboratory is responsible, the real-world problem, what inputs the organization may lawfully use, who owns the research mandate, the intended output and the impact if it is wrong. If the problem has no reliable pass/fail test, mark EVALUATOR_DESIGN_REQUIRED before model generation.

2. Review the council candidates

For each candidate, look for a unique task identifier, generating model, local-versus-hosted label, frozen source fingerprint, model response custody, stated method and constraints. If the alleged different models are actually fixtures or mock replies, mark peer-review status as SIMULATED, not verified.

3. Understand safety and correctness

Read static review as a screening stage, not proof of correctness. Compare pass counts with the fixed number of independent cases in the oracle. For executable candidates, require a legitimate isolated execution receipt, source/oracle matching and errors retained. An unknown result is not a failure count and not a pass.

4. Review repairs fairly

Preserve the original failed candidate. Record who proposed each correction (the generating model, a different model, or a human engineer), whether the code actually changed, what re-ran and whether it passed. The Granby review corrected the local model output; it was not a successful model self-repair.

5. Examine FarmNode results

A registered farm type is only a definition. A logical running lab on one machine does not become an independent research organization. Ask whether at least one separate physical environment and independent assessor repeated the result before using terms such as independent validation.

6. Consider controlled reuse

Do not move a candidate into a DataChip or Genome merely because its bounded oracle passed. Inspect domain-relevance, software-safety, regressions, licenses, intended hardware, institutional review and owner-signed ENTITY authority.

7. Report a failed or missing step

Record the task identifier, result stage, timestamp, receipt reference and error category without including credentials or confidential inputs. Report failed tests, unsupported evaluator syntax, remote provider unavailability and timeouts as separate conditions. Keep the previously accepted baseline unaffected.

8. Investor and regulator interpretation

A verified algorithmic operation is one element of technical diligence. Ownership, third-party rights, privacy, commercial performance, real customers, financial statements, issuance, market operation and investor protections require separate evidence. The AI council never grants investment rights or performs live securities trading.