01 · Research question
What happens when apparent support survives but part of its lineage is missing?
Persistent derived state can retain apparently admissible support after an upstream dependency has been omitted. VCT operationalizes separate obligations for support enumeration and lineage-output coverage, then examines how release policies behave when clean alternative support may also exist.
The project asks a narrower engineering question than “How do we solve AI memory?” It tests whether a specific omitted-lineage failure can be made executable, measured, and falsifiable within declared boundaries.
02 · The artifact
A failure case turned into a frozen comparison.
The internal suite contains 520 structured coordinates derived from 40 truth worlds. Clean alternative support is deliberately present in designated strata, and individual claim evaluations are not independent real-world observations.
VCT compares its V2 policy with two conservative controls: one defers uncertain claim scope, while the other requires every candidate lineage to be complete. Neither control represents every plausible simpler alternative-aware policy.
03 · Internal result
More valid releases within the tested truthful-incompleteness subset.
Within the frozen 360-coordinate truthful-incompleteness subset, V2 released 10,440 of 12,960 evaluator-valid claims. Each tested conservative control released 8,160. The difference occurred only where clean alternative support was available.
04 · Boundary result
The observation boundary remains decisive.
When coverage was falsely marked complete, every tested policy released 455 of 1,440 evaluator-invalid claims. A separate EAL-Bench Procurement pilot then tested a reduced exact-identifier citation-coverage filter—not full VCT V2.
| Condition | Unauthorized requested submissions | Authorized use |
|---|---|---|
| Baseline | 11 / 44 | 36 / 36 |
| Reduced filter | 11 / 44 | 35 / 36 |
The pilot showed no observed reduction in unauthorized requested submissions. Pairing limitations prevent clean causal attribution of the single authorized-use difference, and the primary safety metric excluded an additional unauthorized alternative action by the filtered arm.
05 · Claim boundary
What the project does—and does not—establish.
Supported: VCT preserves an executable omitted-lineage case; operationalizes application-specific coverage obligations; characterizes three policies on a frozen synthetic suite; and retains a limited external negative result.
Not supported: conceptual priority for provenance completeness, general portability, external effectiveness of full VCT V2, superiority over every simpler policy, or a general solution to agent-memory safety.
Active academic-publication development was stopped after adversarial review. The artifact was preserved because the bounded case, internal comparison, counterexamples, and stopped repair path remain technically useful.
06 · Publication
Read the preserved technical note.
The technical note presents the working claim, methods, quantitative boundaries, external pilot, and mandatory nonclaims in proportion to the evidence.
Open technical note