One Recorded Checkpoint, Any Model, Same Checks

Orchestrate your agent's hops. agent-flight-recorder tapes every step of a run: see where it failed, for evals — or checkpoint at the failure and swap in a larger model. From one recorded checkpoint, run it forward with different models, or the same model handed different context, with exactly one declared change per run. The same deterministic executable checks grade every output — no model judges anything — and the traces, grades, and a SHA-256 manifest export as a receipt bundle anyone can re-grade from the files alone. Agent builders who today cannot prove whether a change fixed a run or just changed it. A passing re-run is an anecdote; this harness turns "it worked this time" into a recomputable receipt: one recorded run, one checkpoint, one declared delta per arm, and a scoreboard rendered from the exported bundle - not from live state - so a reviewer, a teammate, or a judge can verify every grade from the artifact alone, with no trust in the harness and no live database. The paired-arm structure has prior receipts in Corey's public Context Lift work: selected cards took a model from 3 of 30 to 19 of 30 (oracle arm 23 of 30), all graded by executable assertions. Demo video: https://www.youtube.com/watch?v=UZw837rUVBk. No live arm run is claimed; the grades shown are fixture grades, labeled as such. The demo is the artifact itself: the public repo, the exported fixture receipt bundle, and the scoreboard rendered from that bundle alone (not from live state). On the fixture traces the deterministic grader produces the hand-computed grades: bare 2/6, generic 2/6, selected 6/6, oracle 6/6 - the oracle labeled upper-bound reference, never a fair arm. These are FIXTURE grades from the test suite, stated as such. Anyone can re-verify the bundle from its files with one command, no server and no database: .venv/bin/python examples/counterfactual_replay/verify_bundle.py <pack dir> A replay on a hosted model is a re-run, not a bit-identical copy. The score, not bit-identity, is the truth claim. Built Oct 30, 2026 on agent-flight-recorder (public), five steps, each committed with its test suite green: Step 1, fork ticket (one declared delta per arm, second delta rejected before anything is written): f98a5a7, 132/132 tests. Step 2, arm runner (one handler, all four arms, model calls written into the trace): 904e948, 138/138 tests. Step 3, deterministic grader (six executable assertions per arm, derived from the recorded run; tamper tests flip the grade): 9002b38, 145/145 tests. Step 4, receipt bundle (traces, grades, SHA-256 manifest; verifies files-only, names the tampered file): 45c8259, 154/154 tests. Step 5, scoreboard renderer (renders from the exported bundle alone): 8611c96, 161/161 tests. Suite re-verified after the Nov 3 push merge: 175 passed. Repo: https://github.com/cwwjacobs/agent-flight-recorder (hackathon files on main as commit bceae47, verified by read-back Nov 3).