Every number, and the run that produced it.
This page is generated from docs/proof/claims.toml
and the evidence files it points at. It is not written by hand. The same
generator runs in the project's health check, so a number that drifts from its
evidence fails the build rather than reaching a reader.
Two claims are deliberately unpublished and say so in the
ledger, and several are marked read by a human: they are historical or
rounded statements that no automatic check can settle, and pretending otherwise
would be a false green.
| Claim | Value | Evidence |
|---|---|---|
| The run's own clock, compared against the same work timed by hand. | ||
| Machine time for one deployed run of the whole lifecycle machine time only; the human approval is timed separately and excluded | 45.1 s of machine time | spikes/judge_run/evidence.json |
| The single human approval the same run required | 22.4 s human approval | spikes/judge_run/evidence.json |
| Seconds from the supplier packet to the case parking for a human the recorded take's own run; the narration rounds it to whole seconds | Parked in 11.9 s | spikes/judge_run/evidence.json |
| The budget one deployed run must finish inside | 130s | spikes/judge_run/evidence.json |
| The same work done by hand, timed author-timed, not practitioner-reviewed | 663.5 s | spikes/manual_baseline/evidence.json |
| Steps in the hand-done walkthrough author-timed, not practitioner-reviewed | 20 steps | spikes/manual_baseline/evidence.json |
| Steps of that walkthrough the run removes measured against an author-timed baseline, not a practitioner-reviewed one | 19 of which the run removes | spikes/judge_run/evidence.json |
| Business days the two cases span the lifecycle clock is compressed, and the compression is disclosed on screen | 380 simulated business days | spikes/judge_run/evidence.json |
| Checks this run re-executes every time it reports them, not numbers quoted from an earlier pass. | ||
| Contracts re-verified at capture time | nine contracts | spikes/core_contracts/evidence.json |
| Graded domain cases, all passing a deterministic pass metric; re-run the suite before citing it | 24/24 | spikes/domain_evals/evidence.json |
| Mean score across the graded domain suite a deterministic pass metric; re-run the suite before citing it | 100% | spikes/domain_evals/evidence.json |
| Contract and unit tests re-executed by the run that reports them re-executed by the harness, not quoted from a previous run | 629 passed | spikes/judge_run/evidence.json |
| What that run actually wrote to the ERP (enterprise resource planning: the system of record for suppliers and purchasing), and what it refused to write twice. | ||
| Fields entered into the ERP without a human retyping them | 22 fields | spikes/judge_run/evidence.json |
| Days a non-compliant supplier was actually held from purchasing claimed only where both the hold and its release executed in the ERP | 5 enforced hold days | spikes/judge_run/evidence.json |
| Decisions policy requires a human to make | 1 policy-required intervention | spikes/judge_run/evidence.json |
| Duplicate ERP writes after a retry the judge run itself retried nothing (commands_retried_then_succeeded: 0); the retried ERP write that proves singularity is the sweep probe in spikes/core_contracts/evidence.json, so this number is not spoken or shown in the video (2026-08-25) | 0 duplicate writes after a retry | spikes/judge_run/evidence.json |
| What running this project costs to operate, kept here for traceability even where it hasn’t been turned into a prose claim anywhere else yet. | ||
| Gross running cost for the whole project, month to date before credits, measured from the billing account rather than from published rates | $16.63 | spikes/cost_posture/evidence.json |
| Cloud credit remaining console-only; no API exposes the balance, so it is passed in when the measurement is taken | held (reason in the ledger) | spikes/cost_posture/evidence.json |
| Measured daily uptime of the screening VM | held (reason in the ledger) | spikes/cost_posture/evidence.json |
| A second Google model, measured against this project’s own documents. Reading them correctly was not enough to adopt it, so both halves of that result are published together. | ||
| Documents an open-weights Gemma read correctly, served by Vertex AI paired with gemma_format_conformant and never quotable without it: this number is the content the model read, after key renaming, and says nothing about whether it returned the required structure | 6 of 6 | spikes/gemma_extraction/evidence.json |
| The same documents on which that model returned the required output format paired with gemma_content_correct and never quotable without it: alone it reads as a model that could not extract, which the measurement disproves | 0 of 6 | spikes/gemma_extraction/evidence.json |
| The same Gemma weights run locally, where the runtime constrains generation to the required structure author workstation, not Google Cloud; carries no deployment claim | 6 of 6 | spikes/gemma_extraction/evidence.json |
| Median time for one served-Gemma extraction, against the deployed step's own median | 8.8 s | spikes/gemma_extraction/evidence.json |
| Statements a human read and confirmed by hand, because no automatic check can settle them. | ||
| The ERP hold was written minutes before the rejection committed | read the evidence file | spikes/hitl_reject/evidence.json |
| The graded suite passed 8/8 when the generation pin was measured | read the evidence file | spikes/domain_evals/evidence.json |
| Reasoning tokens fell to effectively none and the timed sequence shortened | read the evidence file | spikes/gemini_37_eval/evidence.json |
| What the published build-piece article already said, checked here against the run it described at the time. | ||
| The build piece states the graded suite passed 8/8 with reasoning pinned off | read the evidence file | spikes/domain_evals/evidence.json |
| The build piece states a two-supplier run dropped from 85 to 57 seconds | read the evidence file | spikes/gemini_37_eval/evidence.json |