KFD-6 Implementation Notes
Authoritative decision · Formal reference · Documentation map
KFD-6 defines a proposed autonomous discovery loop that is grounded in causal
experience and compares plural generation methods rather than canonizing one. The
authoritative text is decisions/KFD-6.md; its registry status is draft.
Package Surface
decisions/KFD-6.md: authoritative proposed procedure.schemas/kfd-6/autonomous-discovery-loop.schema.json: version 4 experiment contract.standards.json#/standards/kfd-6: schema identity, interface, and concepts.scripts/check.mjs: package wiring and anti-self-certification checks.
Version 4 Interface
Version 4 requires generationExperiments. Each experiment records its
observation perspective, current ontology, generation methods, causal basis,
procedure, evidence cut, resource budget, outcome, and disconfirming test.
Perspective methods additionally declare transformed views and replay details. Other methods remain first class rather than being translated into a fictional perspective origin story. This prevents an autonomous loop from flattening Episodes into an anonymous corpus while keeping method comparison open.
methodComparison binds candidate output to kfd-5-primitive-discovery at
candidateSchemaVersion: 3. It requires at least two generation methods,
fixed-ontology and no-new-primitive baselines, a shared evidence cut, a
declared resource budget, qualification metrics, limitations, and held-out
evidence.
boundaryHypothesis remains conditional. Boundary pressure is useful when
present, but it is not a prerequisite for every genesis method. Replacing the
version 3 perspective-only experiment contract requires version 4 rather than
silently changing its meaning.
Episode Asset Stewardship
Systems that expect future ontology discovery should treat qualified Episode corpora as durable historical assets rather than disposable telemetry. At minimum, stewardship should make the following inspectable:
- stable Episode and evidence-cut identities;
- integrity checking, corruption detection, and recovery boundaries;
- portable export, exact or explicitly lossy import, and provider migration;
- observer, perspective, consequence, omission, and degradation metadata;
- retention, deletion, privacy, consent, and redaction policy;
- the distinction between authoritative Episode records and derived indexes, summaries, embeddings, or training projections.
These capabilities preserve future replay options. They do not imply that every Episode must be retained, that a corpus is complete, or that storage volume predicts Primitive yield.
Experimental Use Only
The schema records a bounded experiment. It does not certify that the corpus is reality-complete, a generated perspective is faithful, an agent discovered a primitive, or a candidate may be promoted.
An experiment must declare its corpus cut, observer, capture boundary, missing evidence, generation experiments, method comparison, autonomy boundary, fixed-ontology and no-new-primitive baselines, held-out evaluation, and promotion authority. The generator cannot be its sole verifier, and generated evidence cannot be the only evidence.
Early Case and Reproduction Boundary
The Consequential Settlement case is an early feasibility case for KFD-6. It records Agent-generated ontology during real product work, later cross-lineage replay, provisional KFD-5 qualification, and separated human/repository admission. Read its development lineage for the exact mapping and missing gates.
Do not encode that history as a passing version 4 autonomous-loop record. Its generation methods, evidence cut, budgets, and baselines were not declared before execution. A future adopter may use the immutable public history as a seed corpus or retrospective fixture, but must freeze a new experiment cut, declare plural methods and baselines, separate held-out evaluation, and retain negative results before making a KFD-6 claim.
Activation Evidence
KFD-6 remains draft until a real adopter compares multiple generation methods under bounded conditions; rejects attractive false candidates; distinguishes objects from perspective, anomaly, model-class, causal, and compression artifacts; preserves causal and responsibility boundaries; and leaves a reviewable path from experience through method comparison to promotion.