Executive overview
NIKI is BTG’s autonomous research and invention programme. Its ambition is not merely to answer questions but to propose computational methods, test them against defined evidence, learn from failed candidates, repair them where possible and retain only demonstrated capabilities. It remains subject to human authority and independent qualification.
01 — NIKI’s intended role
Given an authorized problem, NIKI may study allowed knowledge, generate candidate algorithms or Circuits and seek an executable answer. It is intended to work with FarmNodes, DataChips, BTDU, ENTITY and ADAM rather than replace them.
An articulate explanation or passing compilation is not sufficient. The important distinction is whether the new method performs against independent test cases and reproduces its result in a controlled environment.
02 — The research mandate
A responsible sponsor describes the actual problem, permitted data, desired output, independent acceptance criteria, computational budget, safety rules and qualified evaluator.
Without a clear oracle or measure of success, NIKI may produce interesting material that cannot be defended as a verified invention. The mandate must also prohibit unauthorized access, dangerous actions and leakage of protected information.
03 — Candidate generation and preservation
The system should retain the original candidate, its input assumptions, resource conditions and test results. Any revision becomes a separately identifiable candidate so a later success does not erase an earlier failure.
A suggested repair that leaves the source unchanged or cannot execute due to missing dependencies is not a successful repair. A promising design remains a hypothesis until tested.
04 — Independent tests and adversarial evaluation
A strong evaluation uses tests not simply memorized or supplied as worked examples to the inventing process. It should include boundary values, failure cases, alternative runtimes and resource restrictions appropriate to the task.
A perfect result over a bounded local oracle does not prove all possible inputs or transfer to a new problem class. Reviewers need to see test size, candidate identity, failed runs and whether test data were held out.
05 — Learning from errors
A useful research cycle identifies where a proposal deviated from the expected result, forms an explanation and produces a meaningfully altered candidate or testable revision. Retaining unsuccessful approaches helps avoid repeating the same nonrepair.
The goal is evidence-based improvement, not unrestricted self-modification. Proposed changes must pass controlled admission gates before they affect production applications.
06 — DataChips and Circuit Genomes
A validated NIKI method can be packaged as a versioned reusable capability with a clear input/output contract and associated evidence. A Circuit Genome can preserve the relationships among several such capabilities.
Promotion depends on compatibility, ownership, licence rights, and operator approval. A local success must not automatically provide access to unrelated devices, financial systems or personal data.
07 — FarmNode collaboration
A research task may be worked on by one logical farm and challenged by another. Independent host or institutional review can test whether a candidate works outside its originating environment.
Multiple processes on the same desktop do not establish independent scientific replication. A valid external challenge must document what the reviewer actually reran or independently re-created.
08 — The role of language models
A language model such as a Qwen-family system may help with reasoning, coding and idea generation, if its contribution is measured against the existing NIKI algorithm-invention baseline. It is a tool within the architecture, not a replacement for NIKI’s experiment, validation and ownership framework.
Model size does not demonstrate practical improvement. Hardware cost, reproducibility, measured algorithm quality and safety must govern whether a model is introduced.
09 — Failures and bounded evidence
BTG’s public historical records include a sensor repair later passing all 65,536 cases of one local oracle and a distinct Granby candidate that produced 256 incorrect cases out of 65,536 in its recorded test. The two results address different tasks and cannot be conflated.
The existence of generated repair source is not the same as a qualified changed algorithm. Independent runtime and dependency qualification must accompany future claims of success.
10 — Safety, intellectual property and release
New invention records should document provenance, permitted source materials, potential third-party rights, dangerous or prohibited uses, disclosure restrictions and lawful commercial classification.
NIKI should not grant itself rights, issue investment securities or decide that a technically functioning output is legally marketable. Responsible owners and independent reviewers must approve those actions.
11 — Current status and the next evidence gate
The intended end-state includes reproducible invention, robust self-repair, independently qualified execution environments and release-ready reusable components. Public material shows meaningful bounded progress but not unrestricted general invention or fully qualified physical multi-node production.
This document is a public account of the programme, not a disclosure of confidential models, private algorithms, source repositories or production credentials.
Related manuals and use
The companion User & Operator Manual describes the decisions and reviews applicable to this system. References to proposed user flows are explicitly conditional on qualification of the relevant product release.
Expanded system explanation
| Question | Clear answer |
|---|---|
| Audience and purpose | NIKI — Cognitive Research & Invention is described for the public, institutional evaluators and prospective partners; this is not a software implementation reference. |
| What exists | An experimental research system intended to propose executable solutions, test them against independent rules, diagnose failure, and retain only qualified knowledge. |
| Analogy | A research assistant in a supervised laboratory that must show working experiments and failed attempts, not just convincing answers. |
| Authority | The described process requires actual contracts, human authority and technical qualification at every relevant boundary. |
Practical examples — what a reader should infer
Bounded success
An original sensor repair matched 65,536 of 65,536 local test cases in a specified WebAssembly experiment. That supports a narrow algorithm result, not a claim that NIKI can invent arbitrary correct software.
Bounded failure
The Granby research task previously reported 256 incorrect cases out of 65,536. Such results must remain visible and rejected or under repair, not silently counted as successes.
Reusable component
A candidate method may become a DataChip only after frozen-source tests, rights review, runtime qualification and authorized admission.
System lifecycle — each separate decision
- Receive an authorized problem and success criteria
- Search permitted prior knowledge and propose a candidate
- Freeze the exact candidate before evaluation
- Execute in a bounded test environment
- Compare independent oracle and negative cases
- Preserve failures, repairs and qualification decision
People, organizations and responsible roles
- Research sponsor defines objectives and lawful sources
- NIKI proposes candidates; proposals are not certifications
- Sandbox enforces execution restrictions
- Independent oracle/reviewer judges actual behaviour
- Rights/owner authority authorizes any governed retention or release
Where the system can fail
- Language model size alone does not establish algorithm correctness or autonomous discovery.
- A successful narrow experiment does not establish general invention reliability.
- Untrusted generated code needs sandbox containment and adversarial tests.
- Repair attempts producing unchanged code are not evidence of successful correction.
Evidence a serious evaluator would require
These are acceptance requirements, not claims that each artifact has been issued or independently certified:
- Frozen candidate hashes and prompt/mandate references
- Independent test oracle and holdout definitions
- Raw pass/fail counts and observed counterexamples
- Sandbox cgroup/rlimit/network negative evidence
- Three-host replication if distributed claims made
- Owner-signed governed admission and provenance
Terms explained in ordinary language
| Term | Meaning |
|---|---|
| Oracle | A separate criterion or evaluator for deciding correctness. |
| Holdout | A test withheld from generating or fitting the solution. |
| Candidate | A proposed method not yet accepted as correct. |
| Promotion | Controlled admission of a validated capability into a reusable system. |
How to read the current status accurately
The architecture and local engineering records are distinct from production operation. A named component or test figure cannot establish authorized commercial rights, fully qualified infrastructure or independently accepted financial activity. Read the dated corporate evidence register for each claimed version; ask for a scoped, access-controlled evidence bundle for decisions that rely on private implementation.
Corporate system information → Engineering evidence and limitations → Publications → ENTITY (the separately published open component) →
AI Creation Council and FarmNode research — audited update
BTG has implemented a local, bounded AI Creation Council workflow. It can preserve local Qwen3-8B proposals in ADAM/BTDU, perform restricted reviews, independently test executable candidates inside a sandbox and retain only research evidence in FarmNetwork. Two different model-generated methods passed 65,536/65,536 bounded tests each. The Granby model candidate failed 256 of 65,536; the subsequent full-pass correction was reviewer-authored, not an autonomous NIKI repair.
Cross-model critiques were also exercised using fixtures; no real authenticated hosted multi-model inference or 25-task cross-domain comparison has been qualified. Local tests do not establish novelty, field utility, global superiority or production rights.
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.
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 →