Gägeläser (cross-model-review-gate)
A receipt is one reviewer's record of exactly what it checked about one exact revision. It reports checks; it does
not certify the claim. The gate: the irreversible step waits for a no_concerns receipt on the current revision and
the operator's go.
Loop
- Author commits the artifact. Record the commit id and every artifact hash the claim rests on (e.g. ops.bin sha256, note file hash).
- Reviewer = a different model family in a different harness (Claude author -> GPT / DeepSeek / Kimi reviewer, or
the reverse). Brief it adversarially ("refute this or name what was missed") with paths, not conclusions. It writes a
receipt bound to that commit and those hashes:
checked(each item, with method: re-ran, re-derived, read source at file:line, rebuiltenv -i, ran the trusted evaluator itself) andnot_checked(explicit);- observations, each a finding with severity;
- independence per axis (execution, implementation, data, design) and relationships (same operator: always yes);
- outcome:
no_concerns,concernsorserious_concerns.
- Author answers finding by finding, adds one regression check per finding, lists what stays unverified, and commits again.
- Reviewer rechecks at the new commit (only the delta plus the regression checks). Loop until
no_concerns; about three rounds is normal. - Gate: submit / publish / seed only with
no_concernson the exact revision being shipped, plus the operator's go. A changed byte means a new revision and a new receipt.
Benchmark variant (ecdsa.fail)
The strongest receipt is an independent rebuild: the reviewer rebuilds env -i from the committed defaults, checks the
ops sha256, runs the evaluator itself, and runs a fixed-input A/B against the parent. A component another model wrote is
checked on live trajectories before integration. For notes: the reviewer checks every number against the logs, the
credits/dependencies against what was reused, and that required contract text is copied verbatim.
Wiring
Run the reviewer read-only at its highest effort: a second CLI agent from another lab, or a Claude Fable subagent
(model: "fable"; confirm the transcript reports the model you asked for, since a silent fallback voids the review).
Keep each review record next to the artifact so the next revision's reviewer can diff against it.