AI Agent Evidence, Claims, and Open Gaps

Current conclusions and limitations.

This ledger separates controls that exist from behavior observed in tests and outcomes that remain unproven. Each conclusion retains its scope, limitations, and next test.

Model and cross-harness independence

WTK keeps one package record as source truth when generating its current runtime projections.

MechanismImplementedEvidenceBounded observation

Observed basisImplemented WTK package, contract, and adapter boundaries

Implementation and evidence basis

Mechanism basisTarget-neutral contracts and generated target projections

Evidence basisBounded development observations, reviewed August 2026

Governance independence

The current qualification rules prevent semantic judgment from overriding a deterministic failure.

MechanismImplementedEvidenceBounded experiment

Observed basisA 24-attempt causal probe preserved positive successes and repeated absence failures rather than allowing favorable cases to erase them.

Implementation and evidence basis

Mechanism basisConformance floor separated from evaluator and independent-review evidence

Evidence basisFR-001, one frozen question across two model bindings and four observation conditions

Intelligence-substrate independence

WTK retains a local governance path when an external provider is absent.

MechanismImplementedEvidenceBounded observation

Observed basisOffline structural workflows and explicit provider opt-in

Implementation and evidence basis

Mechanism basisLocal-first provider bindings and provider-neutral evaluator specifications

Evidence basisBounded development observations, reviewed August 2026

Delegation integrity

Current development runs preserve bounded role contracts and attributable handoff records.

MechanismImplementedEvidenceBounded observation

Observed basisWTK team and factory execution paths

Implementation and evidence basis

Mechanism basisCoordinator-led lanes, handoffs, fan-in policies, and evidence records

Evidence basisBounded development observations, reviewed August 2026

Human trust

WTK represents registry discovery separately from qualification and promotion.

MechanismImplementedEvidenceBounded experiment

Observed basisA fixed four-case comparison preserved capability fit, cost consent, discovery, and construction boundaries that an installed-first policy missed.

Implementation and evidence basis

Mechanism basisPackage, registry, qualification, and target lifecycle separation

Evidence basisF-003, deterministic capability-resolution comparison

Governance independence

The current factory design separates provisional improvements from promotion authority.

MechanismImplementedEvidenceBounded experiment

Observed basisTyped repair routing selected the declared layer in five fixed cases, and one package edit retired stale construction and approval authority without deleting history.

Implementation and evidence basis

Mechanism basisIncumbent/provisional lifecycle with qualification, evidence invalidation, and approval

Evidence basisF-002 and F-004, deterministic routing and evidence-retirement experiments

Human trust and governance independence

Current consequential promotion remains human-authorized.

MechanismImplementedEvidenceOperating boundary

Observed basisCurrent WTK lifecycle records require an accountable operator decision before a candidate replaces an incumbent or receives release authority.

Implementation and evidence basis

Mechanism basisPromotion authority remains outside the candidate and improving component

Evidence basisCurrent WTK authority model and bounded development observations

Governance independence, long-term drift, and repair-loop integrity

Can a separate policy-controlled approval process safely authorize specified changes without a human decision on every candidate?

MechanismTarget stateEvidenceUnproven

Observed basisNo public WTK experiment has established safe policy-authorized promotion across repeated improvement cycles.

Implementation and evidence basis

Mechanism basisProposed independent promotion authority with evidence requirements, risk limits, escalation, suspension, revocation, and rollback

Evidence basisArchitecture direction only

Evidence authenticity

Structured evidence can support repeatable local trust decisions.

MechanismImplementedEvidenceBounded experiment

Observed basisA deterministic package revision preserved prior artifacts for inspection while rejecting their construction and approval authority for the changed digest.

Implementation and evidence basis

Mechanism basisDigest-bound artifacts, receipts, provenance, qualification packets, and stale-evidence rejection

Evidence basisF-004, one local source-edit-to-catalog evidence-retirement experiment

Economic value

Does governed autonomy create sustained economic value?

MechanismProposedEvidenceUntested

Observed basisNo sustained organizational deployment evidence yet

Implementation and evidence basis

Mechanism basisOutcome ratings and improvement loops are defined

Evidence basisNo public empirical evidence, reviewed August 2026

How evidence supports a qualification claim

EVIDENCE / QUALIFICATION RENDERER

State only what the evidence supports.

No visual status silently upgrades a mechanism, run, or projection. The evidence scope travels with the claim.

01 / SUBJECTPackage or execution form

02 / RECEIPTSComplete, attributable evidence

03 / VERIFYDeterministic checks

04 / QUALIFYScoped verdict

05 / CLAIMBounded public statement

E / EVIDENCE STANDARD

Claim and qualification language

Evidence views distinguish what ran, what was verified, and what may be claimed.

Canonical packageThe authority-bearing source.

Derived projectionA harness-specific representation.

Enforcement gateA deterministic allow or block point.

Human decisionAn explicit operator judgment.

Evidence pathA receipt-backed observation.

Trust boundaryAuthority changes across this line.