Deterministic Replay — at1 replay

When data was produced by a deterministic generator — a seeded simulation, a synthetic-ML dataset, a Monte-Carlo run, a fuzzing corpus — its entire content is described by the recipe {generator + seed + params}. AT-1 Replay stores that recipe and re-executes it at decode to reconstruct the original byte-for-byte. A 15 MB seeded dataset becomes a few hundred bytes.

Pack — recipe or fallback

# the data was produced by a known generator -> store the recipe
at1 replay generators                       # list built-in generators + params
at1 replay pack run.f64 --out run.at1r \
  --generator np_montecarlo \
  --params '{"seed":2025,"n_paths":500,"n_steps":2000}'
#   -> mode: replay   121 B   ~59,000x vs xz   (byte-exact, SHA-256 sealed)

# no generator (captured data) -> never-worse xz fallback, byte-exact
at1 replay pack sensor.csv --out sensor.at1r

If the declared generator reproduces the file's bytes exactly, the container holds only the recipe. Otherwise — or with no generator — it stores a never-worsexz container, so the artifact is never meaningfully larger than xz and never claims a regeneration that doesn't hold.

Recover a lost seed

# operator lost the seed: recover it blind over a candidate pool
at1 replay recover run.f64 --out run.at1r \
  --generator sk_classification \
  --params '{"n_samples":2000,"n_features":20,"n_informative":8,"n_classes":3}' \
  --seed-pool 65536
#   -> seed=41394 reproduces the bytes; stores the recovered recipe (byte-exact)

Blind seed-recovery regenerates over the candidate pool and matches the bytes. Typical CI/run-id seeds (a 2¹⁶ pool) resolve in seconds; if nothing reproduces the data, it falls back to the never-worse container.

Decode, verify, inspect

at1 replay unpack run.at1r --out run.f64   # re-execute generator, fail-closed, byte-exact
at1 replay verify run.at1r --against run.f64  # reconstruct + SHA-256 check + file match
at1 replay info   run.at1r                 # show the stored recipe (generator, params, mode)

Reconstruction is SHA-256-verified. The recipe is bound to a code-hash over the generator source + library fingerprint (python / numpy / sklearn / scipy versions), so a decode on a build whose generator or library differs refuses rather than emit different bytes — fail-closed, not silently wrong.

Built-in generators: sk_classification, sk_regression, sk_blobs (scikit-learn, fixed random_state) and np_montecarlo (numpy PCG64, documented cross-platform stable). Replay ships compiled under the generative engine license; enable it from your dashboard → Licensed engines. ICP: CI/property-test/fuzzing reproducers, Monte-Carlo risk runs, agent-sim / game-DVR replay, and privacy-safe synthetic dataset pools.