Receipts

live: https://receipts-check.fly.dev receipts checks every claim a coding agent makes in its "done" message against what actually ran. - gemini is the coding agent: it writes the code, its tests and its final message - the tests are actually run; that output is the session log - meta muse spark reads the message + the whole log and pulls out each claim, quoting the log line behind it (quotes checked word for word) - subq snap rules on every claim in one call: backed / contradicted / never ran, with probabilities, in ~0.4 s - first real run: 7 claims, all 7 backed by 10 passing tests. no planted lies