Desk / Evidence ReviewREAD-ONLY REPLAY
THE EDITOR’S EVIDENCE WORKSPACE · READ-ONLY SHOWCASE

Know what the sources actually say.

Desk is the private AI editorial tool behind idhetc.com. This site replays real saved runs exactly as they happened.

An editor brings a small bundle of exact news excerpts. One structured call to gpt-5-mini-2025-08-07 turns them into a Claim Ledger: atomic claims classified from verified to speculative, each tied to quotes that code checks character for character. A person approves statements one at a time, and only approved statements reach the drafting call. Desk never publishes and never writes to the catalog.

PRODUCTION RUN · SEP 30, 2026

Mythri × G Squad: announcement or speculation?

6 excerpts from 3 reports that all trace back to one announcement. 6 atomic claims, 5 human decisions and the grounded draft that followed.

gpt-5-mini-2025-08-07 · evidence-analysis-v6 · 17.7 s · ~$0.0055

Open the review replay ↗
EVALS · 12 FULL SUITES ON EIGHT GOLD CASES

Measured, not assumed.

GPT-5 nano cleared the gates in 0 of 7 suites across 6 prompt versions. GPT-5 mini cleared them in 4 of 5. Every failure is kept and inspectable.

precision ≥ 95% · recall ≥ 85% · citations 100% · critical hallucinations 0

Inspect the eval runs ↗

How one review runs

  1. Freeze the evidence. Excerpts, publisher, source type, origin and independence are locked with a SHA-256 checksum before the call.
  2. One structured call. A versioned prompt and a strict JSON Schema generated from Zod; no tools, no browsing, no automatic retries.
  3. Assume it is wrong. Zod re-parses the output; every quote must occur verbatim, “verified” needs a primary source and “corroborated” needs two independent origins. Failures are saved, never patched.
  4. The Claim Ledger. Atomic claims with rationale, classification confidence and citations located by the server.
  5. Human review. Actions are derived in code; approving an action never approves a sentence.
  6. Grounded draft. The writer sees only approved statements, and every paragraph must cite their claim IDs.

Exported Oct 10, 2026. Raw provider responses and gold-case inputs are omitted; the gold inputs are fictional and listed on the evals page. News excerpts belong to their publishers and link to the original reports.