Median task time
- Traditional
- 39 sec
- JobPilot
- 18 sec
OpenAI Build Week · Apps for Your Life
Codex helped turn application fatigue into five reproducible product repairs.
The build story is not “AI generated an interface.” Codex helped observe defects, reproduce them with evidence, repair the product boundary, and preserve each correction in regression tests.
Build 173bf67 · Study Evidence RC
- 1DISCOVER
- 2PRIORITIZE
- 3UNDERSTAND
- 4APPLY
- 5TRACK
Moderated directional usability study, n=5
Directional usability study
Five participants compared a traditional job posting with JobPilot.
Factual accuracy
- Traditional
- 75%
- JobPilot
- 100%
Self-reported clarity
- Traditional
- 3.6 / 7
- JobPilot
- 6.2 / 7
Small directional study, n=5. All participants completed the traditional condition first and JobPilot second. Different roles prevented same-role carryover, but practice or order effects may contribute to the observed difference.
Confidence: 3.8 → 6.0 out of 7. Decision alignment: 40% → 100%. These are aggregate usability observations, not hiring-outcome evidence.
View methodology
Participants reviewed two synthetic roles: Analytics Operations Engineer and Data Enablement Engineer. Different roles were used between conditions to prevent same-role carryover.
Four factual questions covered required experience, preferred experience, work arrangement, and the weakest-supported qualification. Apply / Review Further / Skip was analyzed separately as decision alignment. Confidence and clarity used seven-point self-reported scales.
Collection was anonymous and browser-local under jobpilot-decision-utility.v2. It collected no PII, and participant-level rows are not public.
Limitation: All participants completed the traditional condition first and JobPilot second. Different roles prevented same-role carryover, but practice or order effects may contribute to the observed difference.
Observed development evidence
Five proof moments
01
Incorrect work-mode interpretation
- Observed
- A practical work-mode mismatch could read like a technical capability failure.
- Reproduced
- Frozen cases showed identical technical evidence receiving a different user interpretation when only work mode changed.
- Repaired
- Technical Fit Score and practical compatibility were separated in data, UI, and explanation.
- Regression tested
- Work-mode mutation cases prove unchanged technical points, score, and receipt arithmetic.
02
Scoring-budget defect
- Observed
- Preferred weighting could cancel itself and an earlier allocation did not reconcile to the full 100-point budget.
- Reproduced
- Micro-point ledgers exposed both the cancellation and the missing allocation deterministically.
- Repaired
- Bounded class transfers restored Preferred contribution while preserving the exact 100,000,000 micro-point total.
- Regression tested
- Property cases, full-score fixtures, class caps, transfers, and rounding now reconcile exactly.
03
Receipt and verifier construction
- Observed
- A displayed score could not independently prove its source inputs, arithmetic, evidence IDs, or canonical hash.
- Reproduced
- Tampered receipts demonstrated that valid-looking JSON was not sufficient evidence.
- Repaired
- JobPilot added frozen inputs, a canonical Score Receipt, and independent web, API, and CLI verification.
- Regression tested
- All bundled receipts reproduce; adversarial changes fail closed without external requests.
04
Missing Tailwind production pipeline
- Observed
- The development interface looked styled while the production CSS bundle lacked required Tailwind output.
- Reproduced
- A production build and route capture exposed the missing pipeline rather than a component-level styling bug.
- Repaired
- The Tailwind/PostCSS production path was restored and its generated bundle bound to a hash.
- Regression tested
- Dedicated CSS pipeline tests and production visual checks now guard the compiled result.
05
Dense Job Detail transformation
- Observed
- The Job Detail read like an audit report before answering whether the role deserved attention.
- Reproduced
- Decision-critical facts competed with arithmetic, evidence, and provenance at the same visual level.
- Repaired
- Fit, priority, strongest match, biggest gap, and next action moved first; proof became progressive disclosure.
- Regression tested
- Responsive browser checks preserve both the fast decision view and the complete inspectable evidence path.
Production reality
A read-only, public-safe snapshot proves current aggregate catalog and source coverage without copying private records.
Decision utility
Today’s Shortlist turns an endless feed into three explainable next decisions with no hidden aggregate score.
Bounded AI
Gemma remains primary, code owns every score, and four gpt-5.6-sol reviews are prepared through Codex, hash-bound, not live, and zero-API.
Neutral differentiation
From fragmented searching to one controlled decision loop.
Traditional job search often separates discovery, comparison, explanation, employer handoff, and tracking. JobPilot connects those steps without claiming complete-market coverage or automating the user’s application decision.