Built at The Agent Harness Hackathon

Lighthouse

Lighthouse predicts which applicants will complete a program AND come back as alumni volunteers, cheerleaders, donors and recruiters — with critical-analytical reasoning, a fairness veto, and a human on every action. Built on TrueForge. THE PROBLEM Admissions optimises for the next four years and hopes for the next forty. Nobody predicts alumni engagement; when they try, they end up predicting wealth. WHAT IT DOES Reads an applicant's own record (statement, activities, academic trajectory, institution interactions, interview notes) and produces a Long-Term Fit Report: calibrated estimates with evidence, counter-evidence and confidence intervals for five outcomes over a 10-year horizon. Reasoning follows the four pillars of Stanford GSB LEAD's Critical Analytical Thinking course: competing hypotheses, graded evidence, the cheapest falsifying experiment, and analogies to past cohorts with their limits stated. Every claim cites the input field it came from. Donor = future capacity × generosity. A trajectory prediction (ambition, persistence, reciprocity), never current wealth. A statement leaking "Atherton, my father's firm, our family foundation" is flagged as a circumstance signal and excluded — the estimate does not move. That's a live test. OBSERVE IT — three saved TrueForge agents, each run an inspectable session: critical analyst (GPT-5.5) → independent fairness auditor (GPT-5.4-mini, veto power) → action proposer. A local MCP server exposes applicant data, base rates, outcome recording and one write tool. CONTROL IT — pipeline_propose_action is @write-gated, so TrueForge pauses with Allow/Deny before anything person-affecting happens. The input schema physically rejects 20 protected attributes. Cost caps: $0.15/report, $10/eval; actual $0.054/report. Nightly TrueForge schedule re-scores every report against recorded outcomes and refreshes calibration. TEST IT — 14 Gherkin scenarios written first; 51 offline + 6 live tests passing; a traceability test fails if any scenario lacks a test. Eval scoreboard (n=30) vs base-rate baseline: beats it on every outcome; 849 evidence claims, 0 hallucinated fields; run-to-run std 0.016; recruiter ranking AUROC 0.97. Two calibration gates fail at n=30 and are reported honestly — that's what the nightly rescoring loop is for. An adversarial self-review with known limits is in the README. SCALE — domain packs swap university admissions for startup recruiting (hire/retain/refer/advocate) or corporate talent with the same analyst, auditor, gate and evals. Seasonal on-ramp: Lighthouse Coach turns the analyst toward the student's own record for families and tutors during application season. Stack: TrueForge (agents, MCP, approvals, schedules, sessions), OpenAI GPT-5.5 / GPT-5.4-mini, TypeScript, Zod, Vitest, Model Context Protocol SDK.