Engineering evaluation without publishing implementation
BTG intentionally keeps the seven proprietary systems’ source, unpublished algorithms, secrets, private implementation contracts and build chains closed. Independent engineering understanding is still possible at the system-boundary level: define inputs and outputs, constraints, failure behaviour, trustworthy evidence and operational limits. A reviewer requiring source-level assurance must use a controlled evaluation process under a suitable agreement; a public explanatory guide is not a replacement.
Component-by-component contracts to evaluate
| System | Public behavioural contract | Crucial failure test |
|---|---|---|
| Bridge | Authorize source, purpose, fields and recipient; deliver only allowed outputs | Revoked/expired scope is rejected and reported |
| NIKI | Frozen candidate proposal, independent challenge, exact failure retention | Candidate with incorrect holdout cannot be promoted |
| FarmNodes | Bounded task on identified host with isolated access and evidence | Cross-host claim fails unless genuinely separate physical hosts tested |
| Circuits | Typed components connected under permitted version and rights | Invalid unit/type mismatch halts or rejects |
| BTDU | Authoritative payload and permitted semantic reconstruction | Reconstruction hash mismatch or unauthorized view rejected |
| ADAM | Policy-governed state transition with evidence | Expired grant cannot alter protected state |
| Exchange | Asset / issuer / instrument separation; audited market and contract events | Ledger simulation cannot be treated as legal issuance or bank settlement |
| ENTITY (separate public component) | Identity, provenance and signed authority references | Signature alone cannot create legal title |
Acceptance stages an independent evaluator should recognize
- Describe the frozen artifact and allowed input/output domain.
- Record environment, operating-system/tool versions, hashes and authorization scope.
- Test normal, invalid, adversarial and boundary conditions, preserving actual failures.
- Demonstrate independent oracle or reference comparison on withheld tasks.
- Verify authority, rights and privacy limits separately from computation.
- Demonstrate genuine host and operator separation when claiming independent distributed execution.
- Report safety, latency and resource failures with measured bounds, not marketing language.
- Verify rollback, denial and incident reporting; classify any missing evidence as pending.
What an independent engineering reviewer must receive privately
- Architecture/data-flow diagrams showing trust and source-of-truth boundaries.
- Signed exact-version artifacts, reproducible run instructions and test harness within an authorized scope.
- Test vectors and holdouts selected by the reviewer or an independent evaluator.
- Evidence of actual host enrollment, mutual authentication and access control.
- Controlled source review or black-box inspection under a protected evaluation agreement.
- Performance baselines with equivalent quality, hardware, power, memory and wall-time metrics.
Do not confuse numerical counts
5,126 registered type definitions or 2,310 logical bindings do not prove thousands of physical laboratories. A result of 65,536/65,536 in one bounded algorithm does not establish autonomous invention on unrelated tasks. A local passing test is not a production acceptance test or independent third-party certification.
AI Creation Council and FarmNode research — audited update
Public evidence captures frozen task-oracle separation, exact 65,536-case local checks, restricted C structural review, bounded independent WebAssembly evaluation, source/oracle receipt matching, negative controls for incorrect and nonterminating cases, and ADAM/BTDU lineage. It does not publish private source, APIs or environment configuration. The Granby reviewer repair and the original model failure must not be collapsed into one autonomous success.
Read the complete AI Council research guide → · Public operator manual → · Printable evidence report →
UPDATED AUDIT · 23 OCT 2026 · 18:26 UTC
Latest Council and FarmNode qualification evidence
The completed 18:26 UTC audit reports 114/114 local tests in nine suites, including the earlier 70 tests (not 184 independent results). The 34 physical-host security tests qualify policy/refusal contracts only, not three real computers. Four explicitly approved hosted routes were zero-priced at the 18:17 UTC catalogue snapshot; no authenticated hosted inference comparison was completed.
Research controller: 5,125 registered types; 2,305 registration-only bindings; 16 source-backed local methods; six original messages consolidated to four unique research requirements. Three historical source-backed evaluation packs now exist for forestry (8 holdouts), agriculture (8), and emergency-response metadata (9). These 25 recorded cases are not 25 new algorithm inventions. Forestry median-change beat naive forecasting on the recorded sample, while agriculture median-change performed worse; emergency-response checks verified historical metadata/timestamps only, not event truth or live warning decisions.
The original Granby model-authored algorithm failed 256 of 65,536 cases. A separately documented engineer-reviewed deterministic repair passed 65,536/65,536 in isolated WebAssembly. This is not NIKI self-repair, a domain-qualified commercial invention, or owner-granted Genome admission. Two other different genuine local Qwen3-8B small-domain candidates passed bounded 65,536-case tests apiece.
Release remains HOLD: 0 of 3 independent physical FarmNodes qualified; no real hosted cross-model algorithm-comparison evidence; the separate 25-task invention benchmark is incomplete; no production ENTITY/Genome owner grant; and no rights or authorization for automatic Exchange investment issuance.
Full Council qualification report → · Download qualification PDF → · Three sector evaluator packs → · Download evaluation PDF →