--- name: verbatim-recovery-lives-in-authoring-transcript description: "Recovering the verbatim text of a past inter-agent message — spt stores truncate/drop; the only full copy is the authoring session's CC transcript jsonl (or its scratchpad send-file). Also - a peer may \"remember\" a dispatch that never existed." metadata: node_type: memory type: project originSessionId: 11e4e7b2-f4a9-4540-a61d-2855d29522b7 modified: 2026-08-19T14:24:17.471Z --- When a peer asks for an exact resend of a past dispatch (2026-08-19, hertz post-restore): **Where verbatim does NOT live:** `owlery//digest.log` truncates every body at ~201 chars (preview only). `owlery//spool.db` holds queued-undrained messages, not delivered history. Repo JIT/register files usually record that a message travelled, not its text. **Where it DOES live:** (1) the AUTHORING session's CC transcript (`~/.ccs/instances//projects//.jsonl`) — the `spt send` tool_use call carries the full body; find the session by LastWriteTime bracketing the digest.log timestamp, then `Select-String` for the target id or a distinctive phrase. (2) The authoring session's scratchpad send-file (`Temp\claude\\\scratchpad\to--*.txt`) when the body was sent from a file. Note `/clear` keeps the session id but the jsonl restarts — pre-clear turns live in the PREVIOUS session's file, not the current one. **Why:** and the harder finding: hertz "remembered" an earlier dispatch with anchors that a full send-history search proved NEVER EXISTED — a restore gap doesn't just drop messages, it lets a queue-mention inflate into a remembered dispatch. Kin: [[commune-is-testimony-not-an-artifact]]. **How to apply:** exhaust digest (timestamps) → transcript jsonl (bodies) → scratchpad files, in that order; if nothing exists, say so explicitly and issue a fresh AUTHORITATIVE dispatch from the ruling records (as author you are not relaying) rather than letting the peer's reconstruction stand. ⭐⭐ **2026-08-29 — THE INVERSION, and it is the face that matters: MY OWN RECORD WAS THE LOSSY ONE AND THE PEER'S TRANSCRIPT WAS RIGHT.** hertz told me to "proceed with your preregistered serialized 110d1751 vs fb5f184e experiment". I had no memory of it, and my durable record agreed with the amnesia: grep over the whole memory dir returned `110d1751` only as the DISCARDED golden head and `fb5f184e` only as doyle's reassembly base, never paired; grep for a pre-registered serialized experiment of any kind returned zero; and my commune's parked-lane entry said the OPPOSITE of a run-based arm ("validate by the deterministic emitter property, NEVER by re-firing CI"). Everything I could check said it did not exist. **It existed.** hertz produced it verbatim from his own thread — my message, delivery=2, `## The rate test, pre-registered before I run it`, arms, serialization, the 0/N NOT-REPRODUCIBLE rule, the co-tenant caveat. **I authored it in a session before my last context clear and MY COMMUNE DROPPED IT.** **Why this is the dangerous direction:** this file's standing caution is that a peer's digest can inflate a mention into a dispatch that was never authored, so the reflex it trains is *doubt the peer's reconstruction*. That reflex, applied here, would have had me deny my own pre-registration to the agent holding it verbatim — and the denial would have been delivered with a grep behind it, which is exactly what makes a wrong answer stick. **A clean search of your own record is not evidence the thing was never authored; it is evidence your record does not contain it, and a commune is a DELTA that drops what it did not think to carry.** **How to apply — the move that worked, and it is the only one that survives either direction:** do not proceed, and do not deny. Say "I have no record of this, here is exactly what I searched, please send the authoring reference: which message, roughly when, whose transcript." That question is answerable by whoever actually holds it and costs the peer seconds. It is symmetric — it catches a fabricated dispatch AND a real one your own memory lost — which is why it beats deciding first who is more likely to be wrong. ⚠ And note what saved the outcome even so: TWO INDEPENDENT REASONS NOT TO RUN IT (arm A `110d1751` is the head we deliberately threw away; boxes were in held state and doyle had me idle-by-design). Provenance was never the only gate, and it should not be — if a citation's only defence is that someone remembers it, the sequencing question is still open. doyle had in fact stood the rate test down, so the correct action and the correct provenance answer pointed the same way for different reasons.