Triskelion

Private experimental harness for Metalogic / LOGOS verified-learning research.

Triskelion tests whether externally verified experience can change an explicit capability system: compose what already exists, construct capability when old closure is insufficient, learn applicability, revise/revoke under counterevidence, change future discoverability, and optionally compile verified competence into neural form.

Evidence policy

The canonical authority is results/ATTESTATION_LEDGER.md. A prose summary, successful CI job, V-number, or result filename is never sufficient by itself. Every headline is now classified as primary-attested, bounded with pinned lineage, result-only/no-CI, negative/incomplete, or not attested in this repo.

The provenance audit is complete to the current repository. This does not mean every historical experiment has been freshly predicate-audited; only claims marked ATTESTED received that stronger treatment. There are no remaining silently asserted PASSes with unknown evidence location.

Bare V-numbers are not unique identifiers. In particular, V49/V50/V51 are used by both the sealed/operator line and an unrelated quotient-refinement line. Use descriptive canonical IDs.

Strongest current evidence

TWO_GENERATION_COMPOUNDING_V54 — BOUNDED PASS. Run 31761530951: under the frozen one-new-generator budget O2 had zero cold survivors before O1, became discoverable after O1, and the final target required both. This is the strongest bounded developmental result.

SOURCE_DISTINCT_DISCOVERABILITY_V34 — ATTESTED, BOUNDED. Run 31724783160: on fixed source-distinct Requests → Rich → Flask functions under a supplied AST frontier vocabulary, prior acquired operators causally expose later Raise then With frontiers; ancestor ablation moves the deepest frontier backward.

VERIFIER_MODEL_LAYER_V56C — BOUNDED PASS. Frozen 12-case small-model applicability task: 6/12 no-memory, 6/12 raw-memory, 12/12 explicit verifier-controlled capability layer.

VERIFIER_INDUCED_INVOCATION_IKKF_V4B / VERIFIED_ROUTING_IKKF_V2D — BOUNDED PASS. External verification can refine applicability in V4b; on the exact V2c C/J worlds, verifier-controlled routing/execution reached 100%/100% vs learned selector 50% routing and 0% joint execution.

PORTABLE_CAPABILITY_IKKF_V1 — ATTESTED WITH CORRECTION. Fresh compilation, a second fresh compilation call, uninstall and reload are separately executed. Do not claim stochastic independence: C1/C2 training trajectories are bit-identical and the gate does not test it.

ADAPTIVE_CONSOLIDATION_V18B — ATTESTED. Primary artifact supports bounded protected consolidation and reload persistence on the measured skill set.

V49–V51 sealed/operator line — audited

The earlier README overstated these before direct artifact inspection. They have now been audited.

• SEALED_TRANSFER_V49 — ATTESTED WITH SEALING QUALIFICATION. Audited run 31749682731: Requests cold FAIL → warm PASS → ablation FAIL; later in-category Django counterevidence → REVOKE. Sealing is supported by phase-separated workflow/code, not by hard-coded sealing flags.
• OUTCOME_LABELED_V50 — ATTESTED WITH SEALING QUALIFICATION. Original run 31749833138: HELP/HARM is genuinely derived from executable before/after transitions (2 HELP + 2 HARM), not supplied labels; Requests transfers causally and later Django HARM revokes. Same sealing qualification.
• OPERATOR_INVENTION_V51 — ATTESTED WITH SEALING QUALIFICATION. Run 31759857216: two executable obstruction episodes have unique common repair < -> <=; it is outside the frozen token-permutation closure, causally repairs withheld Requests, ablation restores failure, and later Django evidence causes revocation. This is not invention outside every supplied meta-language.

The files named V49_EXTERNAL_QUOTIENT_REFINEMENT_RESULT.*, V50..., and V51_RESULT.txt belong to a different experiment family.

Other reconciled historical evidence

The repo now preserves primary lineage for V15–V17, V19–V21, V23, V33/V34, V40–V43 and V70–V74. Those are classified in the ledger rather than lumped together as “historical claims awaiting audit.” Important boundaries:

• V70–72 are bounded positives on a manually normalized 51-event corpus; V73 is a clean hierarchy negative; V74 is mixed and does not establish macro-level incremental math value.
• V45–V66 has a bounded external quotient ledger pinned to parent artifact 9190497808; V68 has a committed result/precommit but no CI run located.
• V36/V37 has a committed constructor checkpoint but no CI run located for that checkpoint.
• V22 has no Actions run located.
• The reported ffmpeg→SHAP EXCEPTION_FLOW_EXTERNAL_STREAM is not attested in this repo: no commit/file/code-search hit containing EXCEPTION_FLOW was found. Do not use a Triskelion ZIP as evidence for it.

Natural / constructor frontier

HISTORICAL_COMPOUNDING_V55A — INCOMPLETE. Only two usable historical worlds survived, both Black; neither yielded a unique O1.

CONSTRUCTOR_COMPOUNDING_V55B — NEGATIVE under tested substrate. K0 obstruction was reached but the frozen insertion meta-substrate failed to synthesize K1.

HISTORICAL_BLIND_EDIT_V56A / CONSTRUCTOR_RETRY_V56B — HARNESS FAILURES. No scientific verdict.

MODEL_CARD_V55C — MIXED/NEGATIVE. Prompt-level scope text over-applied the capability; protected negatives failed.

SORRYDB_V77 — NEGATIVE at current solver/budget. Exact-commit/verifier plumbing works; five frozen solver arms solve 0/3 selected tasks.

Current thesis

A useful research state is:

A_t = (O_t, L_t, S_t, Π_t, K_t, D_t, V)

where O are capabilities, L scoped laws, S applicability/revision/lifecycle conditions, Π operational distinctions/quotients, K constructor machinery, D discovery policy and V external authority. Neural realizations are downstream implementations rather than automatically authoritative state.

The central bounded result is V54-shaped compounding: acquired capability structure can change which later capability becomes discoverable under a frozen search protocol.

Remaining scientific capstones

These are now scientific gaps, not documentation/provenance debt:

1. Natural development: reproduce V54-shaped compounding on pre-existing heterogeneous worlds with solutions sealed.
2. Constructor development: establish Constructible(K0) ⊊ Constructible(K1) under an externally meaningful protocol.
3. Scale/economic value: matched sequential comparison of base, episodic memory, ordinary adaptation and LOGOS on meaningful coding-agent workloads.

Attestation remains a precondition for any capstone claim counting.

Canonical navigation

• results/ATTESTATION_LEDGER.md — authoritative scientific status and allowed wording.
• results/BRANCH_RUN_INDEX_2026-08-14.md — canonical provenance pointers.
• RESEARCH_STATE.md — synthesis, subordinate to ledger.
• results/CURRENT_FRONTIER_2026-08-14.md — concise frontier.
• results/PROGRESS_V49_V56_2026-08-14.md — chronology.
• results/, protocols/, experiments/, .github/workflows/ — primary/reproducibility material.

Research discipline

Closure first. Preserve negatives. Do not infer scientific success from CI success or scientific failure from a harness crash. A gate name cannot claim more than its predicate. Hard-coded booleans do not prove experimental properties. If the ledger says result-only/no-CI or not attested in this repo, downstream prose must preserve that boundary.