Validation ledger

04 OPEN RECORDS

Questions, observations, and boundaries stay attached to the project that generated them.

Evidence is allowed to stop at not yet established.

The register records what is being tested, what has actually been observed, and what the current evidence still cannot support.

4 of 10 systems carry an open validation program. The remaining 6 are documented engineering systems without an open research question — their claims and limits are on their own case studies, not here.

#Research questionStatus

AgentGuard

DEV harness implemented

How should authorization remain bound to execution when retries, crashes, and ambiguous provider effects intervene?

Method: authorization-instance lineage · durable ledger baseline · failure schedules · stale-reconciler fencing · exact statistical decision rules
Observed: The development harness and failure model exist.
Not established: The confirmatory claim is not validated.

Project case study

InferenceLedger

External pilot not started

When does a model, provider, or execution-policy change satisfy every requirement that mattered before the comparison?

Method: freeze → capture → ledger → verify → decide · retained outcomes 2 INCONCLUSIVE · 1 VALIDATED_MISMATCH
Observed: The retained snapshot contains two inconclusive outcomes and one validated mismatch.
Not established: External validation is not established.

Project case study

ProofDiff

External validation in progress

Can change-conditioned evidence selection reduce evaluation work without missing protected behavioral regressions?

Method: independently labeled changes · protected behavioral contracts · deterministic trace replay · omission analysis · selection reduction · provenance-backed impact hypotheses
Observed: Synthetic/self-oracle coverage and an initial external batch are recorded.
Not established: Broader external validity remains open.

Project case study

PermitDiff

Public reference published

Can a permission change be shown to grant more authority before release, without treating an unresolved rule as safe?

Method: same scenarios on baseline and candidate · bounded static check · fail closed when a changed rule cannot be shown to be non-expanding
Observed: The public reference demonstrates a bounded authority-expansion example.
Not established: No external corpus result is claimed.

Project case study