R1GPT — Connection Approval Audit Engine

R1GPT is an AI audit engine for R1 connection approval submissions in Australia's National Electricity Market (NEM). You upload your R1 submission package — GPS baseline, FAT report, OEM metadata and optional PSCAD/EMT studies — and R1GPT returns a structured, clause-by-clause audit report: a calibrated approval probability with confidence bounds, a NER S5.2 clause scorecard, severity-ranked findings with traceable evidence chains, predicted AEMO/NSP RFI questions, and a rectification plan with effort estimates. Every finding cites the exact source document, source field, and the specific section of AEMO's Power System Model Guidelines v3.0 (September 2025) it maps to.

Hardening notes (read before pitching). A judge-lens review flagged several overclaims; the fixes below are now applied to the code:

• METHODOLOGY.md — the score is now a deterministic readiness index (0–100): "calibrated", the confidence interval, the sigmoid and the month percentiles are removed (lib/scoring/engine.ts).
• LIMITATIONS.md — the honest extraction/score/demo boundaries. Extraction failures now surface as an extraction_warnings banner in the report (was a silent server log).
• CITATIONS.md — the DELTA-001 §6.3/§6.4/§6.2.1 mismatch and the SCR > 10 "hard exemption" overclaim are corrected; thresholds are now labelled as R1GPT triage thresholds. Remaining CONFIRM rows still need a line-check against the controlled PSMG v3.0 PDF.
• DEMO_SCRIPT.md — lead with the evidence chain, not the gauge; Q&A kill-shots and answers.
• STRESS_TEST.md — the deterministic-core stress harness is committed (npm run stress, 24/24 passing) and the floating-point threshold-boundary bug is fixed.

How is this different from ChatGPT?

A general LLM gives you a conversation. R1GPT gives you a structured audit report. If you upload these documents to ChatGPT, it reads them and offers opinions in prose. R1GPT instead runs six deterministic assessors — impedance/firmware delta, clause evidence matrix, mandatory-artefact evidence checks, firmware DMAT comparison, NSP staleness, and EMT/RMS model adequacy — each grounded in a specific PSMG v3.0 section. The output is reproducible: the same documents always produce the same findings, the same severity-weighted readiness index (a ranking of what to fix first — not a probability of approval, not a sentiment guess), and the same PSMG citations AEMO would reference in a real RFI letter. The clause scorecard shows pass/fail/partial/missing per NER S5.2 clause, and every finding is traceable to a document field and a regulation section — not an opinion.

Setup

npm install
cp .env.local.example .env.local
# edit .env.local and set OPENAI_API_KEY=sk-...
npm run dev

Open http://localhost:3000.

The extraction layer and RFI generation use OpenAI gpt-4o. Document extraction runs at temperature 0 with strict JSON schema validation (Zod) on every AI response.

Demo

The upload page offers two bundled synthetic packages:

• "Load failing demo" — Ironbark Solar Farm 400MW, a realistic IBR solar project that fails the audit instructively, exercising all six assessors. Expected findings include a DMAT-triggering transformer impedance delta (PSMG §6.3/§6.4), a major firmware bump since the DMAT baseline (PSMG §4.8), absent harmonic filter design and block diagrams (PSMG §4.6.1, §5.2.2), a stale GPS baseline (PSMG §4.8), and unstated SCR (PSMG §3.4). Result: a low readiness index (severe RFI-cycle risk), dmat_triggering materiality.
• "Load passing demo" — Wattle Creek BESS 200MW, a fully PSMG-compliant grid-forming BESS submission (GPS + FAT + OEM + PSCAD). As-built impedance within 1% of design, firmware matching the DMAT baseline, harmonic filter design and block diagrams provided, RUG provided, SCR 15.2, recent GPS. Result: 0 findings, readiness index 95 (the checklist ceiling), all applicable checks passed, minimal RFI-cycle risk, all seven clauses PASS.

Pick one, then click "Run Audit" to run the full pipeline end to end.

To regenerate the demo .pdf fixtures from their .txt sources:

npm run gen:fixtures

Regulatory grounding

Clause assessors are grounded in AEMO Power System Model Guidelines v3.0, effective 25 September 2025. All PSMG section references are cited in findings and predicted RFI questions.

R1GPT audits against PSMG v3.0 only — the psmg_version field is hardcoded to "3.0" throughout the report so the basis of every verdict is explicit.

Architecture

app/
  page.tsx                 Upload page + streaming progress
  audit/[auditId]/page.tsx Full audit report (7 sections)
  api/audit/route.ts       Streaming pipeline (NDJSON events)
  api/audit/[auditId]      Report retrieval
  api/demo, api/health
lib/
  extraction/              PDF parse + gpt-4o extraction + Zod schemas
  assessors/               delta, clause, evidence, firmware, nsp, model_adequacy
  scoring/engine.ts        Severity-weighted probability + timeline
  report/                  builder + PSMG-grounded RFI generation
types/                     documents.ts, report.ts
synthetic/fixtures/        Ironbark Solar Farm demo package

The audit API streams four NDJSON events in order: extraction_complete, assessment_complete, scoring_complete, report_complete (carrying the full AuditReport). The completed report is stored in a module-level map keyed by audit_id for retrieval by the report page.

Scoring

The headline number is a deterministic readiness index, not a calibrated probability. Each finding carries a transparent severity weight (low 4, medium 12, high 25, DMAT-triggering 45) and the index is plain subtraction — readiness = clamp(0..95, 95 − Σ weight) — so a reviewer can read the arithmetic. There is no sigmoid and no confidence interval (AEMO publishes no clause-level R1 outcomes to calibrate against).

The index is deliberately capped at 95, not 100: passing every automated check means "nothing in our finite checklist fired", which is not proof of completeness — the reserved headroom stands for the chartered-engineer review R1GPT can't automate. It also saturates at 0: once findings stack up the number stops discriminating, so among failing submissions the signal lives in the findings list, materiality and RFI band, not the index. To keep both rails honest the report always shows an "X of N checks passed" coverage line (only applicable assessors count toward N). Instead of fabricated P10/P50/P90 month percentiles, it shows a qualitative RFI-cycle risk band (minimal → elevated → high → severe). See lib/scoring/engine.ts and METHODOLOGY.md.

Deployment

R1GPT is Vercel-ready — no database or external services beyond the OpenAI API.

vercel deploy           # preview
vercel deploy --prod    # production

Set OPENAI_API_KEY in the Vercel project's environment variables.

Tech stack

Next.js 14 (App Router) · TypeScript (strict) · Tailwind CSS · OpenAI SDK (gpt-4o) · pdf-parse · Zod · nanoid. IBM Plex Mono for data values, clause references and PSMG section numbers; Inter for body text.