NIKI AI Creation Council

Latest verified engineering — 25 October 2026

Model roster and provider-identity tests passed 7/7 locally, yet the 25-seat Council has not been independently qualified. Actual Grand Forks research used one local creator and zero independent reviewers, so autonomous general project reconstruction is not established.

Oct 25 full engineering record · Contribution/use economy · Council qualification

What the AI Creation Council actually is

Think of the council as a research review panel rather than a single AI that gets to declare its own answer correct. FarmNode research requirements can be grouped into specific problems. NIKI and other permitted models can propose candidate algorithms. Independent validators, not model votes, decide whether a particular candidate solves a defined test.

The council is intended to improve the reliability of reusable algorithms. A more capable or larger model may be consulted when authorized, but parameter count and a convincing explanation are not evidence of correctness, novelty, safety or value.

The proposed research-to-qualified-capability sequence

1. A FarmNode records a real research need, its responsible party, permitted inputs and desired result. Similar messages can be consolidated without inventing new laboratory requests.

2. The evaluator establishes an independently defined task, oracle or challenge dataset and restrictions before asking any model for an algorithm. If no appropriate evaluator exists, the controller holds the request.

3. A bounded council can seek independent candidate proposals and arrange blind cross-model challenges. Council replies are retained as untrusted research with origin, task, version and hashes in ADAM/BTDU; confidential data must not be sent to public hosted providers.

4. Static/structural screening rejects unsafe forms and distinguishes an unsupported evaluator feature from an incorrect algorithm. Promising C candidates must still undergo isolated executable evaluation.

5. Independently defined test cases are run in a restricted WebAssembly environment. Correctness counts, failing cases, resource limits, frozen oracle references and source hashes belong in the evidence receipt.

6. FarmNetwork may retain a research-only result and its failure/repair history. Passing a bounded test does not automatically admit a DataChip, Circuit Genome, service, commercial asset or investment instrument.

7. Broader domain validation, intellectual-property review, independent physical-host reproduction, ENTITY authority and owner-signed production acceptance are separate gates.

Verified engineering evidence — 23 October 2026

Local NIKI council regression: 40 of 40 tests passed. FarmNode demand/invention controller acceptance: 70 of 70 tests across six suites. These suites include overlapping council checks and must not be added together as 110 independent tests.

Two distinct locally produced Qwen3-8B research algorithms, saturating_add8 and xor_parity8, each executed correctly on 65,536 of 65,536 bounded inputs in an isolated Wasmtime process. The xor_parity8 candidate used a loop that an earlier AST interpreter could not score; the separate sandbox qualification demonstrated correctness in the named bounded task.

An independently challenged deliberately incorrect sensor candidate failed 798 of 65,536 cases. A deliberately nonterminating loop exhausted its WebAssembly fuel and did not obtain an acceptance receipt. Missing or cross-task receipts were rejected in negative tests.

A real FarmNode Granby research request led to a local NIKI/Qwen candidate with 65,280 correct out of 65,536 cases (256 wrong). An attempted further model repair timed out. Subsequently a reviewer-authored deterministic single-constant correction passed 65,536 of 65,536 cases in both the task preflight and isolated WASM. This is a repair of model-produced source by engineering review, NOT proof that the model independently invented a correct Granby solution.

FarmNode controller census: 5,125 registered type designs, 2,305 registered-only instance bindings, four locally running logical research environments, 16 existing source-backed research processors, six original research messages consolidated into four distinct needs, 43 reconciled reusable method families across 123 domain families. These are registry/analysis and local processing measurements — NOT physical farm count, proprietary invention count or commercial valuation.

Four original research demands

Natural resources / data quality: routed to NIKI; the Granby original candidate failed, later reviewer repair passed the specified sandbox tests; no production acceptance.

Agriculture / predictive modelling: evaluator design required; do not imply a qualified farm-produced forecast.

Forestry / anomaly detection: evaluator design required; do not claim a tested general anomaly-detection algorithm.

Emergency response / metadata reconciliation: evaluator design required; do not represent a research placeholder as validated emergency software.

Free hosted models: configured versus actually verified

A public-task-only adapter and policy tests have been implemented. It is designed to check explicit free-model identifiers, live zero-price conditions and capped request cost, with paid fallback prohibited. Remote provider availability and privacy terms may change.

At the audited checkpoint there was no fresh authenticated hosted-provider session and no independently qualified real multi-provider council result. The observed successful algorithms were from local Qwen3-8B; mocked cross-model discussions are not evidence of real independent hosted critiques.

Free hosting does not necessarily mean the provider keeps data confidential. No personal, client, protected BTG algorithm/source, partner or child data should be submitted to a public model provider. No API keys or proprietary implementation instructions are published here.

How this affects the other BTG systems

FarmNodes identify and document legitimate research demand. NIKI and its proposed council attempt solution generation and review. ADAM/BTDU retain attributable model proposals, tests and corrected versions. The controlled Wasmtime evaluator checks executable behaviour in restricted tasks.

DataChips and Circuit Genomes may later reuse a capability only after their own compatibility, rights and qualification checks. ENTITY retains actual authorizations and owner-signatures. Bridge permissions do not become research or sale rights automatically. Bit Trade Global may describe a separately qualified asset and properly documented rights; a passing council test does not create a ticker, lawful security, revenue, or investment value.

What remains not established

No real paid-free hosted council parity benchmark, no verified superior/world-best algorithm, no 25-task cross-domain quality result, no independently qualified three-physical-host reproduction and no owner-signed ENTITY/Genome production admission.

The reviewed Granby correction is not autonomous model repair. Incomplete external model calls or model timeouts are not passing or qualified results. No broad agriculture, forestry or emergency-domain evaluator was qualified in the controller audit.

All council records represent local engineering evidence unless an independently identifiable verifier and reproducible external scope are provided. Public access to this document does not authorize model access, data use, service availability, securities issuance or participation.

Public evidence record and validation method

Engineering receipt set: NIKI_AI_COUNCIL_RELEASE_READINESS_20261023.json, NIKI_AI_COUNCIL_FULL_REGRESSION_20261023.log, NIKI_FARM_INVENTION_CONTROLLER_ACCEPTANCE_20261023.json and NIKI_GRANBY_DETERMINISTIC_REPAIR_AUDIT_20261023.json. These remain internal on the authorized BTG workstation. This public summary contains no keys, private algorithm source or exploit-relevant configuration.

For each future public claim, the release record should identify the task, version, actual evaluator scope, pass/failure case counts, source provenance, accountable reviewer, independent replications, disallowed use and current release status. An externally inspectable redacted proof may be published only after rights and security review.