KFD-6 Formal Reference
Authoritative decision · Formal model · Usage · Documentation map
- Status: experimental
- Normative: no
- Formal model version: 1
- Authority:
decisions/KFD-6.md - Decision status: draft
Imported vocabulary
CausalExperience, ExperienceCut, OntologyState,
GenerationExperiment, MethodComparison, HeldOutEvaluation,
PromotionAuthority.
Domain objects
At loop step n:
E_n immutable cut of causal experience
O_n declared current ontology
G_n set of bounded generation experiments
C_n method comparison under a shared evidence cut and budget
Q_n KFD-5 qualification result
H_n held-out or independent evaluation
A_n promotion authority
The proposed loop is:
(E_n, O_n)
-> G_n
-> candidates and no-candidate results
-> C_n
-> Q_n
-> H_n
-> A_n
-> O_n or candidate O_(n+1)
-> new action and causal experience E_(n+1)
Relations and predicates
CapturedBy(e, b) experience e belongs to capture boundary b
Uses(g, E_n) experiment g declares its experience cut
Comparable(g1, g2) experiments share declared evidence and budget rules
Independent(H_n, G_n) evaluation is not solely produced by the generator
Promotes(A_n, x) authority A_n admits candidate x into the ontology
SelfCertifies(x) generator output is its only proof or promotion basis
Invariants
I1 Experiment(g) -> Uses(g, E_n) and Declares(g, O_n)
I2 Compare(G_n) -> at least two declared methods or method configurations
I3 Compare(G_n) -> fixed-ontology and no-new-primitive baselines
I4 Candidate(x) -> QualifiedByKFD5(x)
I5 Promoted(x) -> Independent(H_n, G_n) and Separated(A_n, G_n)
I6 SelfCertifies(x) -> InvalidPromotion(x)
I7 GeneratedEvidenceOnly(x) -> InvalidPromotion(x)
I8 Promotion may preserve O_n when no candidate survives qualification
I9 New experience cannot be silently projected backward into E_n
Loop transition
experience cut fixed
-> ontology and capture boundary declared
-> plural bounded experiments run
-> candidates and negative results retained
-> methods compared against baselines
-> candidates qualified through KFD-5
-> held-out or independent evaluation
-> separated promotion decision
-> bounded intervention
-> consequences enter a new experience cut
Proof obligations
- Preserve action, consequence, observer, capture boundary, and missing facts.
- Declare the ontology being tested.
- Compare methods under explicit evidence and resource budgets.
- Retain negative and false-candidate results.
- Use fixed-ontology and no-new-Primitive baselines.
- Keep generator, verifier, and promotion authority separable.
- Evaluate against held-out evidence and later consequences.
- Keep autonomy bounded and revocable.
Early feasibility mapping
The Consequential Settlement live case can be mapped retrospectively onto the loop without being accepted as a conforming experiment:
| KFD-6 object | Early-case mapping | Status |
|---|---|---|
E_n |
maintainer-held task transcript plus public Kungfu PR, test, review, and dogfood history | reconstructed after action |
O_n |
Fact Admission, KFD-2 trust execution, and Mission/Go work semantics | recoverable |
G_n |
Agent generation of Claim/Assessment/Decision followed by cross-lineage replay and compression with Admission | real but not predeclared or method-plural |
C_n |
no shared-budget comparison against fixed-ontology and no-new-Primitive baselines | absent |
Q_n |
provisional KFD-5 Consequential Settlement cut | partial |
H_n |
repository checks and independent PR review | present but not held-out evaluation of discovery |
A_n |
maintainer authorization plus protected merge admission | separated from genesis |
O_(n+1) |
draft KFD-11 Consequential Settlement procedure | admitted as draft, not activated |
This mapping supports feasibility of Agent-generated ontology change under
real pressure. It does not establish I2 or I3, leaves discovery-specific
H_n incomplete, and does not establish an autonomous trigger for beginning
ontology review. It is a retrospective fixture for designing a future
experiment, not a KFD-6 witness.
Invalid states
- A model-generated corpus is treated as the only reality evidence.
- The generator is its sole verifier or promotion authority.
- Method comparison changes evidence cuts or budgets without declaration.
- New evidence is projected backward into an old experiment.
- Ontology change is rewarded without a no-new-Primitive baseline.
- Schema conformance is presented as proof that autonomous discovery exists.
Machine mappings
| Formal statement | Decision source | Schema or check | Verification |
|---|---|---|---|
E_n, O_n, G_n, C_n |
Experimental contract | schemas/kfd-6/autonomous-discovery-loop.schema.json v4 |
Machine for structure |
I3, I5-I7 |
Anti-self-certification | schema constants and scripts/check.mjs |
Machine for declaration |
| KFD-5 handoff | Qualification boundary | candidate contract/version linkage | Machine for linkage |
| Autonomous discovery capability | Status and activation gate | no current package witness | Not yet proved |
Non-claims and extension points
KFD-6 is a draft experiment contract. It does not claim that causal experience is complete, that a dominant genesis method exists, or that autonomous Primitive discovery has been demonstrated. Activation requires real adopter evidence, false-candidate rejection, independent evaluation, and accountable promotion. Later formal versions may refine the loop without weakening its grounding or anti-self-certification boundaries.