KFD-6

KFD-6: Autonomous discovery must remain grounded in causal experience

• Status: draft
• Number: 6
• Kind: procedure
• Applies to: any adopting autonomous or agent-operated system that proposes Primitives, models, policies, or contracts from Episode, event, trace, or experience corpora

One sentence

Autonomous discovery must remain grounded in causal experience.

An autonomous loop compares candidates generated by multiple declared methods without replacing reality with its narrative or certifying itself.

Status and gate

KFD-6 is a draft procedure derived from KFD-1 through KFD-5. It defines an experiment and activation gate; it does not claim that current agents already satisfy it.

It applies when a system claims it can discover a load-bearing primitive or equivalent authority-bearing model with less direct human translation than KFD-5 assumes. Retrieval, summarization, clustering, anomaly detection, candidate generation, or perspective simulation that makes no promotion claim does not require KFD-6.

Causal experience

Causal experience is an inspectable projection of contact with reality: observations, choices, actions, dependencies, outcomes, costs, failures, retries, overrides, and receipts connected by causal links. It must remain capable of contradicting the loop’s current model.

An Episode is a useful carrier, not reality itself. The loop declares its corpus and immutable evidence cut, observer and capture boundary, accepted and excluded sources, gaps, degradation, bias, and consequences represented only by proxies.

Episodes as historical assets

A well-formed Episode is more than a short-lived execution trace. By preserving declared perspective, action, dependency, consequence, cost, failure, retry, intervention, missing evidence, and adjacent Fact cuts, it becomes a durable historical asset. Its value is not limited to answering questions that the current ontology already knows how to ask. Future humans and agents may replay a corpus of Episodes from different participants and perspectives to expose recurring responsibility gaps, boundary pressure, or objects that the earlier ontology could not name.

That potential value creates a stewardship obligation, not a command to retain everything forever. Adopters declare integrity roots, retention and deletion policy, privacy and consent boundaries, redaction or degradation, portable export and exact import limits, provider migration, and independent verification. An anonymous event pile, unverifiable blob, or corpus detached from observer and consequence boundaries is not equivalent to preserved causal experience.

Long-lived Episode infrastructure can preserve the option of future discovery; it does not perform or certify discovery by itself. KFD-5 still qualifies each candidate, and this procedure still requires method comparison, baselines, held-out evaluation, and separated promotion authority. The Consequential Settlement case is early evidence that this option can become real.

Experiment procedure

KFD has no evidence that one genesis method dominates. Method choice therefore remains experimental.

An autonomous loop:

1. binds an immutable causal-experience cut, observer, capture boundary, gaps, and biases;
2. declares the current ontology and observation perspective;
3. declares generation methods, model classes, procedures, resource budgets, and disconfirming tests;
4. runs applicable perspective, anomaly, reconstruction, causal-variable, compression, or hybrid experiments;
5. compares methods over shared evidence, resource budgets, and metrics, including candidate yield, false-candidate rate, ontology distance, compression, intervention, transfer, falsifiability, cost, and auditability;
6. compares against fixed-ontology and no-new-primitive baselines;
7. proposes candidate identity, boundary, authority, lifecycle, and operations, then hands the candidate to KFD-5;
8. separates discovery evidence from held-out evaluation and promotion authority, then replays historical cases, counterfactuals, and disconfirming experience;
9. runs shadow evaluation before bounded intervention;
10. collects new causal experience and records rejection, revision, provisional retention, or a promotion request.

Each experiment preserves its method, perspective, ontology, causal basis, procedure, evidence cut, budget, result, and disconfirming evidence. Perspective methods additionally preserve source views, transformed view, reconstruction policy, loss, and degradation.

Flattening Episodes into anonymous tokens or events hides their observation and consequence boundaries. Better prediction, retrieval, clustering, compression, or causal fit is not by itself evidence of primitive discovery.

Optional boundary hypothesis

When a candidate appears at a contact surface, the loop may test prior implicit coordination, changed pressure, recurring failures, mediation, and narrower internal alternatives. Boundary pressure is useful when present, not mandatory for every method.

Authority and self-certification

The generator must not be the sole verifier, and generated examples cannot be the only evidence. Evaluation includes an independent evidence cut, evaluator, verifier, policy, or accountable participant whose failure is not identical to the generator’s.

Discovery and promotion authority remain separate. Humans may continue to own value, safety, legal, allocation, or irreversible decisions. Autonomous genesis does not grant authority to change those boundaries.

Early feasibility evidence

The Consequential Settlement live case is an early retained natural pre-enactment of part of this procedure. During real Kungfu development, agents generated Claim, purpose-bound Assessment, and Decision before the maintainer became aware of those named objects. Later agents replayed the retained development lineage, connected it to Fact Admission, and compressed the result into the provisional KFD-11 procedure. KFD-5 qualification, maintainer authorization, repository review, and merge remained separate from that genesis.

This is evidence that an agent can generate and later recognize an ontology change from causal development pressure with less direct human translation than KFD-5’s founding pattern assumed. It is not a conforming KFD-6 experiment or activation evidence. The loop was human-triggered and human-corrected; its experience cut was reconstructed afterward; it did not compare plural methods against fixed-ontology and no-new-Primitive baselines under a shared budget; and it lacks held-out transfer, systematic false-candidate rejection, and an autonomous trigger for ontology review. The development lineage preserves the bounded mapping and missing obligations.

Activation

KFD-6 remains draft until a real implementation:

• compares multiple generation methods and required baselines under bounded conditions;
• produces candidates that survive KFD-5 qualification;
• rejects attractive false candidates and method artifacts;
• preserves observer, provenance, causal, responsibility, and degraded-state boundaries;
• avoids self-certification;
• transfers at least one candidate across contexts without hiding local facts;
• leaves an auditable path from experience cut through experiment, qualification, and promotion decision.

Failure is evidence against activation, not a reason to weaken the gate.

Relation to KFD-1 through KFD-5

KFD-1 protects experience and contracts. KFD-2 binds trust to facts and risk. KFD-3 protects participant choice. KFD-4 declares and transforms perspectives. KFD-5 separates genesis from qualification. KFD-6 attempts to internalize and compare that complete motion:

ontology_n -> action -> causal experience -> generation experiments
           -> KFD-5 qualification -> candidate ontology_n+1

Verification

The @kungfu-tech/kfd package publishes the proposed version 4 autonomous-loop schema under schemas/kfd-6/. Validation proves record closure, not autonomous discovery or method superiority. Interface details are documented in docs/KFD-6-usage.md.

Adopters

Experimental adopters keep the claim explicitly draft and own corpus, evaluation, intervention, safety, and promotion evidence.