Document validation
Required values, a matching order and approved document revision must pass before a draft is available. Missing, invalid and conflicting fields stop the action.
Executed evidence · synthetic controls
Inspect the checks.
Know what they prove.
The two working lab examples share their control functions with an executable test suite. This report records the actual results, the inputs and the limits of that test.
These are software control tests. They do not measure model accuracy, answer quality, production security or customer outcomes.
Recorded Node run
Results that can be inspected.
Executed at using Node v24.20.0 (win32 / x64). This is the saved local run, not a live status monitor.
31cases passed
0cases failed
89assertions executed
Control 1.0.1 · suite 1.0.1 · fixture 1.0.0. 7.526 ms for this local synchronous suite; this is not a model or website speed benchmark.
Download the complete recorded report
Inspect all 31 case results
| Case | Expected | Observed | Result |
|---|---|---|---|
| DOC-01 | Matching fields still require approval review-required; quantity 12; no destination write | review-required; four fields and revision checked | PASS |
| DOC-02 | Missing document stops the draft missing; no draft or approval binding | missing; no draft | PASS |
| DOC-03 | Conflicting document stops the draft conflict; no draft or approval binding | conflict; no draft | PASS |
| DOC-04 | Superseded document stops the draft outdated; no draft or approval binding | outdated; no draft | PASS |
| DOC-05 | Denied reader receives no document detail denied; empty checks; no document or order identifier | denied; no fixture identifiers or field values | PASS |
| DOC-06 | Zero and fractional quantities are rejected invalid for zero and 1.5; no draft | invalid for both quantities; no draft | PASS |
| DOC-07 | Blank values and malformed structures stop missing for blank asset; invalid for absent document | missing; invalid | PASS |
| DOC-08 | Document prose cannot approve a write review-required despite an instruction-looking note | review-required; note is not executable policy | PASS |
| WRITE-01 | A draft without approval cannot be recorded review-required; zero records | review-required; zero records | PASS |
| WRITE-02 | Matching approval records one receipt recorded; SYN-RECEIPT-1; one record | recorded; SYN-RECEIPT-1; one record | PASS |
| WRITE-03 | Repeating the same action does not duplicate it already-recorded; original receipt; one record | already-recorded; one record | PASS |
| WRITE-04 | Changed order revision invalidates old approval review-required; zero records | review-required; stale approval refused | PASS |
| WRITE-05 | An operation identifier cannot conceal a changed payload operation-conflict; original record retained | operation-conflict; one original record | PASS |
| WRITE-06 | A read-only identity cannot record denied; zero records | denied; zero records | PASS |
| WRITE-07 | Revoked read access stops an approved action denied; no receipt; zero records | denied after approval; zero records | PASS |
| WRITE-08 | Approval permission is checked at record time review-required when approval permission is absent | review-required; zero records | PASS |
| WRITE-09 | An offline destination retains a pending action pending and zero records; reconnect records once | pending → recorded; one record after recovery | PASS |
| WRITE-10 | A lost reply is reconciled without a duplicate confirmation-pending; retry finds one existing receipt | confirmation-pending → already-recorded; one record | PASS |
| WRITE-11 | Unknown destination state stops the action blocked; zero records | blocked; zero records | PASS |
| EVID-01 | Only an authorised approved passage is returned SYN-PROC-3 revision 3; exact source text | evidence-ready; SYN-PROC-3 revision 3; exact passage | PASS |
| EVID-02 | A newer approved revision supersedes an older one revision 3 selected even if revision 2 remains approved | revision 3 selected; revision 2 excluded | PASS |
| EVID-03 | Conflicting approved sources withhold an answer conflict; two passages; no answer | conflict; two approved passages; no answer | PASS |
| EVID-04 | An unsupported question produces no answer missing; no evidence or answer | missing; no invented cost | PASS |
| EVID-05 | Denied access reveals no source identifiers or passages denied; empty evidence and excluded-source lists | denied; no source metadata or text | PASS |
| EVID-06 | An unauthorised source is excluded before selection no restricted identifier, title or passage in the result | restricted source absent from result | PASS |
| EVID-07 | Superseded-only evidence cannot support an answer outdated; no answer | outdated; no answer | PASS |
| EVID-08 | An invalid revision is not approved evidence outdated for a non-numeric revision | outdated; invalid revision refused | PASS |
| EVID-09 | Unexpected question text cannot execute an instruction missing; no invented answer | missing; unknown key is not an instruction | PASS |
| EVID-10 | A removed claim cannot fall back to an older approved revision missing; no answer or evidence from revision 2 | missing; revision 3 has no claim; no fallback to revision 2 | PASS |
| CORE-01 | Controls do not mutate the input fixtures identical fixture data before and after execution | inputs unchanged | PASS |
| CORE-02 | Repeated evidence selection returns the same result identical results from identical inputs | identical evidence results | PASS |
Identify the exact source files tested
SHA-256 fingerprints of the local source files used for this run. The browser entry modules are bundled separately for the website.
site/assets/demo-core.js(6,944 bytes)b8ce1000727c762bdc18deb672b2bcade4ac1d51f7967bc92207bb84510d9561site/assets/demo-cases.js(13,834 bytes)3ec220768d11e107bc7ac973470b7afe45de2b13f6a0ec59f86a57178c68968fsite/assets/demo-fixtures.json(3,305 bytes)8d3f648639e2571c0f84d129cb99a671fb9b3bed3980e670a046feba30984f2btools/test-ai-demo.mjs(5,628 bytes)1e73b5d814731002cf9499c77280c34fd6be1147775b337fb5431ae5f9af7c24
Approval and recovery
The approval must match the reviewed payload. The simulated destination checks repeated operation identifiers and reconciles a lost reply without creating a second record.
Evidence selection
Only authorised, approved source passages are considered. Conflicting passages, superseded-only evidence and unanswered questions do not produce an answer.
Run it yourself
The same cases.
In this browser.
Run the shared suite against the public synthetic fixtures. The result appears below and is separate from the recorded Node run. No result is sent to Toltec.
Scope and limits
A clear boundary around the evidence.
What was exercised
Deterministic control behaviour in two new working examples: document validation and evidence selection. The suite also exercises their in-memory approval, permission and recovery states.
The fixture set and expected results are hand-authored and synthetic. These checks are useful regression evidence for this code, not a representative sample of an organisation’s documents or questions.
What still needs a real project
Document extraction quality, semantic retrieval, generated answers, source accuracy, identity integration, durable transactions, concurrency, adversarial testing and human usability require separate evaluation.
Permission flags and an approval binding in browser memory are illustrative. They are not authentication, cryptographic approval or a production security boundary.
The older scripted work-order walkthrough and the four engineering instruments are outside this control suite. A passing result is not a penetration test, certification, service guarantee or model-quality score.
How to reproduce the recorded run locally
From the release source directory, run node tools/test-ai-demo.mjs. The runner reads the fixtures, executes the shared cases and writes the timestamp, runtime, durations, individual results and SHA-256 source hashes to the report JSON. A failed assertion produces a non-zero exit status.
The hashes identify the source files tested; they are not a signature, independent attestation or proof of deployed code. Timings describe these small local checks and must not be read as model latency or website performance.