AI Creation Council — Verified Engineering Status

01 — What is actually new

BTG has a local AI Creation Council and a FarmNode research-demand controller. The controller reads authenticated research requests, consolidates related needs, and routes eligible tasks to NIKI. Research candidates are preserved with source and evaluation evidence; they are not automatically published as algorithms, admitted as production Circuits, or turned into investment instruments.

The latest completed local acceptance record is dated 23 October 2026 at 18:26 UTC. It reports 114 passing tests across nine suites. The total includes the earlier council and controller tests, rather than being additional to them.

02 — Measured acceptance evidence

Across the nine recorded suites: 40 council/sandbox tests, 12 source-request controller tests, 6 operator dashboard tests, 5 bounded-dispatch tests, 4 source/provenance gate tests, 3 ontology reconciliation tests, 5 real-data evaluator tests, 5 reviewer-repair proof tests, and 34 physical-host security contract tests. Total: 114 passing. These results concern test-defined local behavior, not a qualified international laboratory service.

The source ledger identifies 5,125 registered type definitions, 2,305 registration-only bindings, 16 existing locally executable sector processors, and six original FarmNetwork request messages reconciled to four distinct needs. A design registration is not an operating independent machine or an independently validated laboratory.

03 — What the council demonstrated with generated algorithms

Two different original local Qwen3-8B-generated bounded algorithm candidates were preserved in ADAM/BTDU and tested through independent WebAssembly execution. Each passed 65,536 of 65,536 cases for its particular fixed small-input task. The results demonstrate specific task correctness and the local source-to-evidence pipeline; they do not demonstrate scientific novelty, real-world utility across industries, or a superior universal algorithm.

For the Granby data-quality mask task, the original model-authored candidate produced 65,280 correct outputs out of 65,536: 256 failures. Later model-authored repair attempts failed to qualify. A separately labelled engineer-reviewed deterministic source correction then passed 65,536 of 65,536 cases in isolated WebAssembly. That success must not be represented as an autonomous NIKI repair or unmodified AI invention.

04 — Three sector evaluation packs and 25 historical cases

The council’s research controller has three source-pinned evaluation packs for forestry, agriculture, and historical emergency-response data. These use 25 holdout observations in total: eight forestry forecasting cases, eight agriculture forecasting cases, and nine historical emergency alert schema/timestamp checks. These cases establish baseline/evaluator groundwork and source traceability. They are not 25 independently invented or validated algorithms.

Forestry: 13 historical observations, including 8 held-out cases. A simple median-change baseline modestly beat a naive predictor on the reported error measure; this is not validated anomaly detection. Agriculture: 18 observations, including 8 held-out cases; the median-change method did not beat the naive baseline. Emergency response: 15 historic alert rows and 9 timestamp/schema checks passed, without independent ground-truth event labels or authorization for operational public alerts.

Before council-generated methods can be judged for these fields, the owner must approve data rights; domain specialists must specify independent outcome labels, permitted error and false-alarm tolerances, held-out samples, negative controls and evaluation procedures.

05 — Hosted-model cost and data controls

An anonymous live API catalogue query on 23 October 2026 at 18:17:56 UTC returned 469 model IDs, including 17 zero-price free-labelled entries. Four locally approved free-labelled models were observed to have zero prompt and completion prices in that catalogue snapshot: NVIDIA Nemotron 3.5 Super VL, Poolside Laguna S 2.1, Cohere North Mini Code, and NVIDIA Nemotron 3 Ultra. This does not prove perpetual free access, rate availability, model quality, absence of data retention, or zero charges outside the checked route.

No authorized fresh API credential was enrolled, no real hosted chat-completion algorithm response was verified, and no genuine multi-provider blind proposal / cross-critique comparison was completed. The council is configured to block paid fallback and limit public research prompts; no private personal, commercial, child or restricted data is authorized for third-party submission.

06 — Physical FarmNode and authority gate

The remote FarmNode test suite passed 34 of 34 defined tests for identity/owner-grant policies, mutual-TLS reference rules, nonce and revocation checks, malformed commands, and rejection controls. Passing a simulated contract is not an end-to-end cryptographic hardware attestation.

Zero of the three required physically independent FarmNodes were enrolled and qualified. No owner-signed ENTITY/Genome production admission was issued. Hardware signatures, separated machines, external-network operation, reproducible execution and failure/revocation behavior on actual hosts remain unproven.

07 — What remains HOLD

Production release: HOLD. No 25-task independently verified algorithm invention benchmark; no qualified live free hosted model comparisons; no verified autonomous Granby repair; no three independently attested physical FarmNodes; no signed production rights; no validated commercial or regulatory approval for turning research results into market instruments.

Next controlled steps: enroll a new credential securely without publishing it; run matched independent real model proposals and adversarial critiques using explicitly approved public data; complete the 25-task multi-domain invention and comparison benchmark with validated independent evaluators; qualify three genuine hosts with cryptographic and operational evidence; preserve human and ENTITY owner sign-off for any production promotion.

08 — Records and interpretation

Primary internal source records: BTG_FARMNODE_CREATION_COUNCIL_FINAL_GATE_AUDIT_20261023.json and .md (18:26 UTC), NIKI_HOSTED_ZERO_COST_CATALOG_LOG_20261023.log (18:17 UTC), NIKI_FARM_ORIGINAL_SEALED_EVALUATORS_20261023.json (18:19 UTC), NIKI_PHYSICAL_FARMNODE_GATES_20261023.log, and the original council acceptance and Granby reviewer-repair receipts. Public statements summarize locally observed records; no independent third-party certification is claimed.

Audit figures are a point-in-time research qualification record, not assurances of future performance, service availability, investment returns, privacy compliance, or fitness for safety-critical field use.