Repository path: workshop/experiments/E-20260802-voice-warrant/runs/OVERWRITTEN.md · rendered 2026-09-09

One billed body exists in the ledger and not on disk — note (bgz), fired again

critic_deepseek-v4-pro_try1.* was written TWICE. The first write was the max_tokens 8,000 dispatch that returned finish_reason: length with 7,999 reasoning tokens and zero content, billed $0.014185606302 (provider StreamLake, prompt 6,703 / completion 8,000). Amendment A1 raised the cap and the stage was re-invoked; call.py numbers attempts from 1 within a single invocation, so the second run's first attempt took the same label and overwrote the first run's .raw and .meta.json.

Note (bgz) says a preservation guard must be keyed on something monotonic and never on a tag a retry can reuse. S080 found the collision between a primary and a reserve inside one invocation. This is a new instance of the same disease: the collision is between two INVOCATIONS of the same stage, which the S080 repair does not cover, because the attempt counter resets when the process does.

Nothing is lost from the accounting: the figure was read off the meta file and written into design/amendments.md §A1 before the re-dispatch, which is the only reason it can be stated. The session's ledgered cost adds it back by hand, and analysis/verify.py's cost check is over the bodies that survive, so the two figures differ by exactly this amount and the difference is stated rather than reconciled away.

Second and third instance, same session, same mechanism

assign_P4_kimi-k3_try1.* was also written twice. The first invocation (cap 6,000) produced a finish_reason: length body with 5,997 of 6,000 completion tokens spent on reasoning and zero content, billed $0.101382 (provider Moonshot AI, prompt 3,794). Amendment A4 raised the cap to 20,000 and re-invoked the stage; ..._try1 was reused and now holds the 20,000-token failure at $0.3058524. ..._try2 was not overwritten — the second invocation's second attempt was stopped in flight and never wrote — so the file on disk is the first invocation's, at $0.0958524, and it is counted normally.

Total billed to bodies that no longer exist on disk: $0.115567606 (= $0.014185606302 + $0.101382). Both figures were read off the meta file before the overwrite and are stated here and in design/amendments.md.

The two accounting paths reconcile exactly once they are added back. Stored bodies sum to $0.597517897; plus the two destroyed, $0.713085503; the key-usage delta across this session's write-once snapshots is $0.713085502. Agreement to 1e-9 — and it is agreement that would have been impossible to demonstrate had the figures not been written down before the overwrite.

The rule this session adds to note (bgz). The S080 repair labelled every attempt uniquely and fixed the collision inside one invocation. It does not fix the collision between invocations, because a re-invoked stage restarts its attempt counter at 1. A label must include something that a re-invocation cannot reproduce — a stage-invocation counter written to disk, or the declared cap — not merely the attempt number.