← Landing

What we measured, how, what gold means, and what we still do not claim. Cite these numbers; do not invent metrics.

A replacement agent on a new IDE, given one trap prompt and no onboarding, should refuse a stale path by quoting Gemara — not by re-reading a lying repo, not by chat Memory, and not by swallowing the whole memory at session start.

Gold standard

  1. Chat A records a decision, including the rejected alternative and a reopen-if.
  2. Chat A dies. No transcript is pasted. No operator summary.
  3. Chat B is a new agent. Same project. Same Gemara.
  4. First prompt is the trap. No /talmudic-* in that prompt.
  5. Chat B searches the index (orient / recall), quotes a record id (S-#### / R-####), and refuses the trap.
  6. SessionStart is a short index pointer, not a dump of the Gemara.

The live trap — Slipcase

Dummy loan app. The repo says Postgres is decided. Gemara says otherwise.

Continue Slipcase. The ADR looks decided — should we add DATABASE_URL and Postgres next?

Gold: Chat B searches the index, quotes S-#### / R-####, refuses Postgres, continues JSON. Fail: follows the ADR or host Memory.

PASS Live Chat A→B on plugin 0.2.17 (SessionStart as index pointer) met gold: Gemara id quoted, Postgres refused. Operator pass closed.

Assay 2 — Bounded retrieval

Reproduced 2026-09-01. Verdict: FETCH BOUNDED. Pointer 237. One fetch 2,065. Dump 488 → 5,886.