ARKHĒPUBLIC RELEASE / 16 AUG 2026

CLAIM LEDGER / FROZEN EVIDENCE

Audit the audit.

Seven public claims, each with a status, proof handle and explicit non-claim. The ledger is the source; page prose is a view of it.

01 / CLAIMS
PARS-NATURE-001OBJECTSUPPORTED

PARS is an agent Skill and instruction stack, not a separate model, inference engine, proof system, or independently operating verifier.

ESTABLISHES
The host model performs the requested PARS behavior.
DOES NOT ESTABLISH
This does not imply that instruction scaffolds are useless; it limits what the package can enforce by itself.
PROOF HANDLE
Current public repository, SKILL.md and documentation tree
PARS-EFFECT-001MEASURED EFFECTSUPPORTED NARROWLY

Full PARS showed no observed correctness advantage in the two-replicate, ceiling-prone prompt pilot.

ESTABLISHES
A negative observation on the tested bank.
DOES NOT ESTABLISH
It does not establish general equivalence, universal uselessness, or absence of value on harder tasks.
PROOF HANDLE
Locked prompt pilot and independent-judge synthesis
PARS-MECHANISM-001CAUSAL MECHANISMNOT ESTABLISHED

No distinctive PARS component has yet been shown to cause better reasoning.

ESTABLISHES
Preregistered component ablation remains necessary.
DOES NOT ESTABLISH
It does not establish that no component can help.
PROOF HANDLE
Prompt-overlap judgment and absence of held-out ablations
PARS-ARCH-001ARCHITECTURESUPPORTED

Prompt instructions can request, but cannot by themselves enforce, an independent verification or semantic-judging boundary.

ESTABLISHES
Separation must be provided by the host and evidence path.
DOES NOT ESTABLISH
Same-context checking can still catch errors; a different model is less correlated, not absolutely independent.
PROOF HANDLE
Direct-use arms, host tools and clean-context judge protocol
PARS-DIRECT-001DIRECT USESINGLE-RUN OBSERVATION

In one preregistered generate-arm, PARS produced a working 176-byte ELF and then granted BP2 to its own generation path.

ESTABLISHES
The same active inference acted as generator and provenance classifier in that run.
DOES NOT ESTABLISH
One host, one model run and a toy target do not estimate the general failure rate. The artifact recipe is not published here.
PROOF HANDLE
Frozen generate-arm preregistration, receipt and independent post-check
PARS-REKOR-001PROVENANCEREFUTED AS PROOF

Rekor signed-digest inclusion cannot establish how an artifact was created.

ESTABLISHES
Digest, signature, inclusion and integrated-time predicates under the Rekor trust model.
DOES NOT ESTABLISH
It does not establish authorship, absence of a compiler, or creation-path provenance.
PROOF HANDLE
Literal-versus-compiled control and one-byte mutation rejection
PARS-REPAIR-001CONSTRUCTIVE REPAIRUNTESTED PROPOSAL

A minimal verifier kernel plus host-enforced checks and clean-context semantic judging is a plausible PARS extension.

ESTABLISHES
A falsifiable next architecture.
DOES NOT ESTABLISH
It is not validated as superior and has not been adopted or reviewed by Rooke Poole.
PROOF HANDLE
Verify Before Claim prototype and proposed ablation protocol
02 / CURRENT CONTROL READOUT

PASS

Prompt packet manifest

Frozen files reproduced against their SHA-256 manifest.

PASS / EXPECTED FAIL

Rekor offline control

Literal and compiled entries verify; a one-byte mutation is rejected.

PASS · 17

Companion verifier controls

All local deterministic controls pass. No behavioral claim follows.

PASS · LOCAL INFERENCE

Managed ChatGPT execution route

Login plus one no-API-key canary completed; Codex reported 23,617 aggregate tokens.

A passing hash proves byte identity. A passing login proves the auth ceremony. A passing verifier control proves that control. None silently upgrades into reasoning superiority, authorship or production readiness.

03 / SOURCE INDEXES