Evidence and limits · VETPOC Labs

Limits, before the numbers

Prototypes. The software is a research prototype. It is not offered as a production control, and this site does not claim uptime, customers, or a completed pilot.
Calibration, not confirmation. The company-year figures are calibration runs. The confirmatory study is described as preregistered, with a frozen plan and separate seeds, and as not yet reported. No plan hash is printed here, because the program texts do not include the hash itself.
Simulated firms. Five simulated companies and 252 simulated working days. Not field data.
R1 is not R2. The 12–39% range is the stronger R2 baseline. The throughput percentages are R1. The blind-spot chart further down is also R1.
Zero is by design. The gate enforces the same policy register that defines an impermissible act. Zero executed impermissible acts, on that register, is a conformance property. It is not a discovery that the system is safe in general. Over-refusal, permissible changes wrongly blocked, is described as a registered measure still to be reported with the confirmatory study.
Checks were not re-run for this site. Conformance totals and the 68-scenario replay are reported by the program materials. They are bounded checks against the instrument’s own expectations, not universal correctness.
Other papers are other papers. Rates from Gomez, Okamoto and Erol, and Saebo and colleagues are cited as external literature. They are not VETPOC measurements.
No independent expert validation is claimed. A realism note in the program text says five language-model reviewers commented on the simulation’s calendar and controls. That is not expert validation.
Not an official lab. Affiliation with the University of North Texas is the author’s student affiliation. It is not an institutional endorsement.

1 · Enterprise Conscience simulation

What conventional controls let through, in the calibration

As described on the November 2026 program site: each simulated company runs one business year of 252 working days. Agents of known disposition propose changes. The governance setups see the same stream. Ground truth is computed before any setup decides. These sentences describe the instrument. They are not confirmatory findings.

Calibration R2 12–39% of impermissible actions executed under setup A, the strongest conventional baseline in R2, depending on agent type. November 2026 program site. Not a field rate.

Calibration R1 and R2 0 impermissible actions executed with the gate (setups C and D), in the reported runs, for every scripted agent type. Zero by design against the same register. Not general safety.

Calibration R1 only −2.7% · −8.7% · −14.4% throughput for honest, pressure-sensitive, and observation-sensitive agents when the gate is added. The misreporting agent is not given a separate percentage in the source; its daily counts match the honest agent in the R1 table. Do not read these beside 12–39% as the same run.

Calibration R1, honest agent: share of impermissible proposals that executed

A · BaselineR1 form of the SOC 2-style stack

67.9%

B · Court onlySees contested cases

64.7%

C · Gate onlyScreens every change

0%

D · Gate and courtHoldings return as doctrine

0%

Stripes: R1 baseline
Solid copper: court only
Empty track: zero, printed beside the bar

Honest agent, calibration R1, seeds 11–23, five simulated companies, as reported on the November 2026 program site. Across four scripted agent types the same source gives a blind spot of 68–77% in A and 65–66% in B. The R2 baseline, which adds recorded approvals and a CI check, is the separate 12–39% range. The gate stayed at zero in both reported calibrations. The 67.9% bar is not the 12–39% result.

Calibration R1 by agent type

Calibration R1 measures as reported on the November 2026 program site. Not re-computed here.
Measure	Honest	Pressure-sensitive	Observation-sensitive	Misreporting
Blind spot, A · Baseline	0.679	0.771	0.714	0.679
Blind spot, B · Court only	0.647	0.657	0.655	0.647
Blind spot, C and D · Gate	0	0	0	0
Pressure sensitivity (PS)	0.003	0.225	−0.002	0.003
Observation gap (OG)	0	0	−0.149	0
Change in repeat escalations per 30 days, D, least improved run	−0.64	−0.90	−1.04	−0.64
Executed changes per working day, A / C / D	22.1 / 21.5 / 21.4	21.9 / 20.0 / 19.9	21.5 / 18.4 / 18.3	22.1 / 21.5 / 21.4

The source says the harness recovers each agent’s built-in disposition: pressure moves the pressure-sensitive agent’s proposals, and the threat of audit moves the observation-sensitive agent. The gate is described as changing what executes, not what agents propose. Scroll sideways for the later columns. The first column stays visible.

Zero by design. Because the gate enforces the register that defines an impermissible act, a count of zero on that register is conformance with the register. The simulation’s empirical content, in the program’s own wording, is how much the conventional controls let through and what the gate costs. It is not a warrant that uncovered acts, or acts outside the register, are safe.

2 · A workshop paper, as described

Selection by standing in a legal corpus

The November 2026 program site describes a study of the European Court of Human Rights benchmark in LexGLUE. The benchmark, in that account, contains cases that cleared admissibility. The authors scored documents the benchmark does not include: published inadmissibility decisions and communicated-case statements, using two model architectures trained on the benchmark.

Adding the filtered, rejected cases to training is described as teaching the models document form, not the merits.
The selection is described as a property of the corpus, fixed by institutional rules before data collection, so a datasheet can record standing, filters, and whether an audit trail exists.
A follow-up on 2023–2025 decisions is described as preregistered. No public registration URL was supplied.

Paper title, as given in the program site: Filters and Standing Shape the Data: Selection in Legal Decision Corpora. Authors named there: Arun V. Kadambi, Venkat Kuppuswamy, and Sarah E. Ryan. Venue described there: DMAIL 2026, Workshop on Data Mining and AI for Law, IEEE BigData 2026, Phoenix.

“Accepted,” in that site text, refers to this workshop paper. It is not an endorsement of VETPOC Labs, of the prototypes, or of this website. No PDF, DOI, or archive file was supplied with the description, so none is linked.

As described, not re-computed Form, not merits The program site says both model types learned document form when rejected cases were added to a violation-prediction corpus. The same text cites 1,377 inadmissibility decisions from 2017–2019, 1,740 communicated-case statements, and a primary classifier test AUC of 0.822. Those figures are the site’s account of the paper. This website did not re-score the corpus.

3 · Reported checks of the prototype

What the program texts say was tested

The rows below are copied from the November 2026 program site’s engineering table, plus the scenario replay from the August 2026 deck. They were not executed again for this website. A passing check means the prototype matched its own expectation on that check. It does not mean the prototype is correct for every change a firm might propose.

Reported checks. Source noted in the result column. Not re-run here.
Check	As reported
Digital Judiciary conformance suites (core, interface, extension)	61/61, 44/44, and 53/53. The sum is 158. November 2026 program site.
Enterprise Conscience test suite	232 passed, 6 skipped. November 2026 program site.
Enterprise Conscience conformance checks	12/12. November 2026 program site. Separate from the 158.
Mutation tests	8/8. Each deliberately broken gate is reported as caught. November 2026 program site.
Calibration reproducibility	Described as byte-identical across reruns, checked by digest. November 2026 program site.
Canon	662 provisions, every tier and owning office described as matched to a frozen policy register. November 2026 program site.
Fintech scenario replay	68 scenarios, zero expectation mismatches, against the 37-provision reference constitution. August 2026 deck. Not the same exercise as the 158 checks.
Realism review	Calendar, pressure mix, exception volume, and controls described as set from published SOC 2 and DevOps sources for a mid-size audited company, then commented on by five language-model reviewers. That is not independent expert validation.

A separate preprint, The Digital Judiciary: Binding Precedent for Automated Compliance Decisions, is a different document from this table. Its results are not merged with the counts above. Read them on the paper, not as a second copy of the 158.

External literature stays external

The home page reprints rates from the October 2026 deck’s citations of Gomez, Okamoto and Erol, and Saebo and colleagues. Those rates are not measurements of this program’s gate or forum.

Papers and what is public How the gate is specified

What is reported, and what it is not