Persona Testing
AI persona testing framework. Define user personas in YAML, run automated UX testing with browser automation, get structured feedback from real user perspectives.
- YAML persona definitions with scenarios and criteria
- Structured reports with fixable/tradeoff classification
- CI/CD integration via readiness thresholds
- Example personas: new user, power user, accessibility
Activity Timeline
Per owner answer d-20261022-a10d, promoted coordinates get pins while indicated and proposed sites get uncertainty circles. The report brief carries status lines for about 14 projects and a corrections section.
Pin-rule review round 1 returned BLOCK with three findings.
Auditors had reworded and then restored claims, leaving the invalid flag set. Gen170 audit scripts passed a 53-case sandbox simulation but were not run for real.
Residuals audit runs 153 and 154 stopped on invalid-checkpoint (rc 3) because auditors reworded claims and the gate never clears its invalid flag; neither was resumed. gen152 primer tooling was written; its 133 negative tests passed and all 17 targeted mutants were caught.
Labels are read from their first letter. All 12 persona-testing sessions were Polaris orchestration, including a run-140 close audit that refused handoff with 53 of 64 claims verified, 4 unsupported and 7 partial.
Visible-text and setext headings are now recognized, undefined level-three sections are rejected, and assessment headings are counted. Parser round bq-5647 plans two P1 heading-grammar fixes and two P2s, with a second review round pending.
A GPT-6.1 Sol round-1 review returned SHIP with 0 open P0/P1. C6 was escalated with a default that keeps the install live and leaves bq-5050 and bq-3191 open.
Each of three sequential Astra xhigh reviews returned a BLOCKED verdict, requiring iterative edits to the candidate.
Root cause: npm disabled TOTP enrollments October 2025, security keys now mandatory. Commission deliverables: diagnostic document with 16 verbatim doc citations, card draft, and vault-integrated automation script.
100 snippets generated across 10 arms (5 styles × Fable/Opus). Dynamic quota routing switched accounts mid-run with zero failures. Three owner rulings enacted on style arm constraints.
b-report-cards-2 agent initialized with ledger cross-check against existing decisions to avoid re-asking settled questions. Gate b-report-cards-2-20260805 registered; processing not yet begun.
Full test suite green. Uncommitted changes in src/loop.ts noted as not yet covered by tests.
Fable session on Account B recovered from crash state. sol-contracts blocked on stacked permission-mode approval dialogs; sol-boundaries pending review gate. No actions taken beyond triage.
Four targets identified and verified clean: persona-probe, tic-tac-toe, escape-underclass, DSPy optimizer. All git-resettable with deterministic unit tests and no credentials. Benchmark infrastructure ready for model comparison runs.
persona-probe (296K TypeScript) ranked first; three others follow. All verified: clean git, no secrets, static-safe for isolated benchmark runs. Sets up Kimi K2.7 / GLM 5.2 cross-model evaluation.
Candidates — persona-probe, tic-tac-toe, escape-underclass, DSPy optimizer — all have objective test criteria, no secrets, static deploy. Target models: Kimi K2.7 and GLM 5.2.
Refactored from single-provider agent into generic framework (Anthropic + OpenAI). 11 Playwright tools, YAML-driven persona definitions, 22-case parser test suite, OIDC publishing pipeline fixed.