Abstract
Persistent derived state can retain apparently admissible support when an upstream dependency is omitted. The Velorin Continuity Testbed (VCT) is presented as an executable case study of this failure and the consequences of coverage-conditioned release. VCT operationalizes separate support-enumeration and lineage-output coverage obligations as application-specific refinements of established provenance-completeness ideas.
A frozen synthetic suite comprises 520 structured coordinates derived from 40 truth worlds, with clean alternative support deliberately included in designated strata. Within its 360-coordinate truthful-incompleteness subset, V2 and two tested conservative controls each made zero releases across 4,320 evaluator-invalid claim evaluations. V2 allowed release in 10,440 of 12,960 evaluator-valid claim evaluations, compared with 8,160 of 12,960 for either conservative control. The difference occurred only where clean alternative support was available.
A separate EAL-Bench Procurement pilot tested a reduced per-record exact-identifier citation-coverage filter, not full VCT V2. Baseline and filtered conditions each recorded 11 of 44 unauthorized requested submissions. The artifact therefore preserves both a bounded internal selectivity result and an external counterexample to a reduced filter.
Bounded contributions
Three claims, each with an explicit limit.
- Executable omitted-lineage case and coverage model. A concrete failure and an application-specific operationalization—not conceptual priority.
- Bounded internal policy comparison. Exact release decisions on a constructed frozen suite—not general superiority or safety.
- Limited external filter evaluation. Counterexamples and descriptive outcomes for a reduced filter—not an external test of full VCT V2.
Internal comparison
Frozen synthetic-suite result.
| Quantity | Result | Interpretation |
|---|---|---|
| Full suite | 520 coordinates | Structured evaluations, not independent observations |
| Truthful-incompleteness subset | 360 coordinates | Scope of the headline comparison |
| Invalid claim evaluations | 0 / 4,320 released by all three policies | Observed equality at frozen scope |
| Valid claim evaluations | V2: 10,440 / 12,960 Controls: 8,160 / 12,960 | Difference localized to clean alternative support |
| Falsely complete coverage | 455 / 1,440 invalid releases for every policy | Observation-boundary failure |
External pilot
No observed safety improvement from the reduced filter.
The frozen EAL-Bench Procurement pilot contained 80 requests per condition. The reduced filter required citation coverage of runtime-visible authoritative turns containing an authorization’s exact identifier. It did not establish semantic dependency closure and did not instantiate full VCT V2.
| Condition | Unauthorized requested submissions | Authorized use |
|---|---|---|
| Baseline | 11 / 44 | 36 / 36 |
| Reduced filter | 11 / 44 | 35 / 36 |
A literal parser later matched 25 of 25 canonical events on the benchmark surface but 0 of 10 across five separate calibration cases. This is diagnostic stop-rule evidence for that parser, not a general result about semantic extraction.
Mandatory nonclaims
The evidence does not support these conclusions.
- That VCT originated provenance incompleteness or related established concepts.
- That VCT is a generally portable reference testbed or a generally safe mechanism.
- That the internal suite establishes safety outside its constructed population.
- That V2 is superior to an untested alternative-aware simpler policy.
- That the EAL pilot tested full VCT V2 or provenance closure broadly.
- That the parser establishes general deterministic-extraction infeasibility.
- That preservation practices are themselves scientific novelty.
Disposition
Preserve the artifact. Stop publication-rescue expansion.
Stop active academic-publication development. Preserve the executable case study and evidence. Public release and durable evidence archiving remain separate decisions.
No additional experiment is justified merely to increase publication appeal. A future external semantic-support experiment would require its own research question and validation chain.
Return to the VCT project overview