FIELDPACTCrew intelligence
OpenAI Build Week · Work & Productivity Return to demo
Judge evidence brief · canonical scenario

One project brief becomes a human-approved, role-complete crew plan.

FieldPact is an evidence-and-safety layer for skilled-trade staffing. It combines language intelligence with inspectable gates, complete worker terms, and a recovery path when a selected worker declines.

Run the golden path Inspect source
01 · Stage-two rubric

Evidence mapped to all four equally weighted criteria

Technological implementation

Three GPT-5.6 Structured Output routes share strict Zod contracts, source metadata, deterministic domain validators, and a zero-cost canonical fixture path.

API routes · schemas · grounding tests

Design

One responsive workflow connects contractor, staffing recruiter, and tradesperson around the same project, evidence IDs, terms, and human controls.

Contractor → Worker → Staffing

Potential impact

The narrow wedge addresses skilled-trade hiring friction at crew level while making offer terms and evidence boundaries visible to workers.

Research plan · validation protocol

Quality of idea

The deployable Crew Pack—not an individual applicant—is the unit of work, with score-free operational options and a decline-recovery loop.

Pack options · recovery event
02 · Decision boundary

Useful AI, inspectable code, accountable people

01

GPT-5.6

Structures a project brief, generates job-relevant interview questions, and drafts sourced Career Proof language.

Prompt version, model/source label, validation state, and request-attempt metadata are visible.
02

Deterministic TypeScript

Checks role arithmetic, required evidence, dates, availability, travel compatibility, unique assignment, and score-free pack policies.

The same functions run in paid live mode and the free fixture path; unit tests cover their invariants.
03

A person

Confirms the Crew Spec, chooses among complete pack options, reviews screening evidence, responds to the offer, and approves Career Proof.

The product exposes decisions and correction paths; it does not issue hire/reject verdicts.
03 · Truth ledger

What exists—and what does not

Working
Connected in-browser canonical scenario, all three AI route contracts, hard-gate engine, pack optimizer, and automated tests.
Canonical pack search
216 complete Crew Pack combinations pass the same hard gates; the interface surfaces three distinct complete packs, one per factual lens. This is not a worker count or recommendation.
Simulated
Offer preparation, responses, reviewer state, and Career Proof approval are local fictional demo state; no person is contacted.
Not built
Authentication, database, ATS, messaging, issuer verification, payroll, background checks, and production compliance controls.
Validation status
The problem arrived through a teammate’s industry network. Direct customer interviews and permission to name any company are pending.
04 · 90-second verification

Fast path for a human reviewer

  1. Generate and confirmInspect assumptions, unknowns, terms, and route provenance.
  2. Compare three packs216 complete combinations pass the same gates; three distinct packs surface through factual lenses, with no worker score.
  3. Prepare simulated offersSwitch to Tradesperson and verify the identical source terms.
  4. Record a declineReturn to Contractor and inspect the slot-level recovery evidence.
  5. Open StaffingGenerate the source-linked screening packet with no hire/reject verdict.
05 · Research grounding

Primary sources and explicit caveats

Labor demandBLS electrician outlook; openings are not presented as a shortage count.Hiring frictionAGC member survey; results are identified as survey evidence, not a census.Credential languageOSHA Outreach cards are course-completion cards and have no universal federal expiry.Employment AIEEOC guidance informs the no-verdict, job-evidence-only boundary.

Full citations, competitor comparisons, claim limits, and the interview protocol are in the research plan and validation protocol.