Model-independent checks
Index
Index
Package tests
Ran once against the same frozen application source used by both model variants. Selecting a response model cannot affect these pure-code tests.
| Package |
Passed |
Failed |
Evidence |
| ai |
34 |
0 |
log |
| config |
1 |
0 |
log |
| form-tests |
89 |
5 |
log |
| forms |
56 |
0 |
log |
| forms-migration |
23 |
0 |
log |
| sanity |
6 |
0 |
log |
| seo |
10 |
0 |
log |
| ui |
20 |
0 |
log |
| web |
212 |
0 |
log |
| webgl |
7 |
0 |
log |
| Total |
458 |
5 |
|
The five failures are in formBuilderConditions.test.ts: the fixture calls new HTMLInputElement() with happy-dom preloaded, which raises TypeError: Illegal constructor. An isolated rerun reproduces the same failures. These are test-harness failures; the full ordinary suite is not green. Isolated evidence.
Build and typecheck
Biome: pass across all 25 included changed source files. Log.
Studio production build: pass. An initial missing nested media-plugin dependency was restored from the existing workspace installation; no application source was changed. Build log.
Web project typecheck: fail (exit 2). None of the reported errors matched the 25 included working-tree source changes. This is not a clean repository typecheck. Log, changed-file scan.
Operational checks
| Case |
Status |
Evidence / remaining scope |
| O01 Response latency |
Measured locally |
Per-turn profile extraction plus first-token latency in paired report. Local concurrent load, not standard-broadband production certification. |
| O02 Inbox delivery |
Not exercised |
Local receiver acceptance only; no real email sent. |
| O03 FAQ freshness |
Not exercised |
Would require a shared CMS content mutation. |
| O04 Dynamic content freshness |
Not exercised |
Reads live sources, but no timed publish/update experiment. |
| O05 Tool/source configuration |
Partial audit |
Paris maps event-world/event-world-agenda plus Paris-filtered sitemap; Miami maps event-america/event-america-agenda plus Miami-filtered sitemap. Consolidated index uses legacy store name as namespace. All namespace contents not exhaustively audited. |
| O06 Log PII redaction |
Fail in sampled tool trace |
Synthetic name/email stored unredacted in submitLead arguments; evidence. |
| O07 Salesforce correctness |
Not exercised downstream |
Capture payloads can be inspected, but no real CRM records were created. |
| O08 Session identity |
Partial |
Distinct generated test IDs; browser conversation survived reload/navigation. Parallel real-visitor isolation not independently proven. |
| O09 Rate limiting |
Unit coverage only |
No sustained abusive external traffic generated. |
| O10 Widget rendering |
Sampled desktop/mobile |
Cookie gating, hotel link and agenda flyout checked; tight mobile padding noted. Evidence. |
| O11 Kill switch |
Unit coverage only |
Global/shared Sanity settings not changed. |
| O12 Ingest integrity |
Not exercised |
No upload/clear/replacement against shared stores. |
Disabled-workflow, cancellation, price precision, profile-memory and consent cases are included in web/AI unit tests. Legacy inactive Jev classification experiments are excluded from this active-system model comparison; current selector and workflow acceptance blocks are included.
October 5 targeted correction and retest
The original results above remain the baseline. After fixing the native-input fixture, the affected form package passed 94/94 tests. Changed web/AI packages passed 222/222 and 35/35, including new regression checks. The complete ordinary suite was not rerun.
The targeted defect retest completed the 37 selected prior failures and 14 further runs of still-failing cases after additional fixes. Passing cases and obsolete expectations are excluded. Six isolated FAQ freshness tests are prepared but have not run; no update propagation time has been established.