README.md

Example runs

Real runs of the engine. Nothing here is staged: every source was fetched live, the verifier ran independently in its own context, and the trace is what the engine produced. (One change, disclosed for honesty: run.json was later extended with the per-cycle grounding history — cycles_detail and refused_ever — that the engine now records natively; every value in it is this run's real, unchanged history, recovered from controller/loop-log.md.)

20260625-135009-ai-notetaker-sea-gtm

A GTM market-opportunity brief — "Assess the market opportunity for an AI sales-call notetaker targeting SMB sales teams in Southeast Asia (Vietnam, Indonesia)."

Read it in trace order: brief.md → decompose/sub-questions.md → select/ → retrieve/ → consolidate/synthesis.draft.md → verify/contestation.md → controller/loop-log.md → synthesis/result.md.

What this run shows, end to end:

• The verifier refused a real claim. In cycle 1 the independent verifier caught CLM-02 — a Gong positioning claim — citing a web page that never mentions Gong, and refused it. A genuine defect, not a staged one. Several pricing claims were downgraded for the same root cause: borrowing a citation to a page that did not contain the figure. The cycle-1 refusal is recorded in controller/loop-log.md, in run.json (refused_ever: true, cycles_detail), and in verify/contestation.md (its cycle-1 note and the disposition table). This run predates the engine's per-cycle contestation archival, so cycle 1's verdicts live inside the final contestation.md rather than a separate contestation.cycle-01.md; newer runs write one file per cycle.
• Contestation shaped the research. The refusal drove the controller to re-research rather than synthesize over the weak claim (see controller/loop-log.md, cycle 1 → cycle 2). The remediation re-grounded the claims to vendor-primary sources and materially corrected stale figures (Otter Pro $10 → $8.33; a phantom Fathom tier removed). Grounding rose from 26/34 to 33/33.
• It refused to fabricate. No open source scoped to the SMB-sales-notetaker SEA market, so the engine left the market size honestly unanswered (CLM-31) instead of inventing a number — and dropped a paywalled figure rather than rely on a source a stranger cannot reach. (That figure still surfaces in search-result snippets, but is not confirmable on the open page itself — the engine refuses to ground a claim to a number it cannot fetch.)
• It is reproducible. Every load-bearing source is free and keyless; the central result can be re-run from the same brief. (Open web pages may drift over time; the OpenAlex and World Bank spine is exactly reproducible.)