Replay manifests

Reconstruct a captured launch and verify data, runtime, capacity, and real training-path identity with replay manifest v14.

halo-forge replay <run_dir> inspects a captured run and reconstructs its launch command. Add --launch to run it again.

halo-forge replay ./models/raft
halo-forge replay ./models/raft --launch

Every shipped trainer writes replay.json beside training_summary.json. Current manifests use format v14; v1-v13 remain readable.

What v14 captures

Replay v14 records everything through v13, plus the exact real training-path revision and certification. V13 added the managed runtime digest, host/device identity, core qualification, and accelerator coexistence decision. V14 adds:

  • pinned certification fixture and exact model commit;

  • real trainer and capacity-adapter versions;

  • before/after parameter evidence and reloadable artifact evidence;

  • host identity and current path-certification identity; and

  • workstation beta-qualification identity when that full evidence gate exists.

  • canonical run ID, modality, model, timestamp, literal launch arguments, seeds, full resolved configuration, backend, platform, and package versions;

  • local dataset hashes, pinned Hugging Face source identity, or immutable Dataset Lab role bindings and content-addressed trainer artifacts;

  • immutable verifier profile, implementation, configuration, reward contract, qualification, and runtime identity;

  • reward-system revision, ordered auditors, reward mapping, retention protocol, integrity profile, selected boundaries, signal capability, trace manifests, audit decisions, and runtime compatibility; and

  • corpus extraction/version identity, tokenizer and packing hashes, budget semantics, LoRA/full adaptation mode, and CPT training-artifact identity;

  • proof outcome and reviewed full-run decisions;

  • controlled adaptation study protocols, arms, assignments, and deviations;

  • specialized task, label-schema, model-head, processor, loss, and retrieval identities; and

  • deterministic environment, fixture, tool, episode, snapshot, and trajectory identities.

Versioned sections are present only when the corresponding workflow was used. Large datasets and signal traces are hashed or referenced, not embedded in the manifest. Their checksum bundles remain in Dataset Lab and Reward Integrity storage.

Drift checks

Replay displays environment differences before showing the reconstructed command. Launch behavior is stricter:

DifferenceDefaultExplicit override
Python, platform, backend, or package environmentrefuse--force
Local dataset contentrefuse--allow-dataset-drift
Verifier identity/runtimerefuse--allow-verifier-drift --verifier-drift-reason "..."
Reward system, auditors, protocol, profile, boundaries, capability, or trace identityrefuse--allow-reward-drift --reward-drift-reason "..."

Verifier and reward overrides require a reason. Halo Forge appends that reason and the exact differences to replay_overrides.jsonl. The launch is then intentional, but it is no longer an exact replay.

Legacy raw-verifier runs stay readable and carry legacy_unqualified rather than fabricated qualification. Older runs without training-signal traces are never described as reward-integrity monitored.

Manifest versions

VersionAdded identity
v1run, seed, config, dataset, environment, and arguments
v2managed dataset roles and trainer artifacts
v3immutable verifier reliability and qualification/runtime state
v4reward system, auditors, protocol, integrity profile, signal capability, traces, boundaries, and decisions
v5corpus extraction/version, tokenizer and packing, budget, adaptation, and CPT artifact identity
v6proof outcome and full-run decision
v7adaptation study protocol and assignments
v8specialized task and task-artifact contract
v9environment, episode-suite, snapshot, and trajectory identity
v10prepared evaluations, study assignments, grounding verification, environment execution, and reviewed decisions
v11immutable dataset repair, repaired-record set, source fingerprint, workstation readiness, and platform capability
v12immutable training plan, prepared-model commit, capacity evidence, safe adjustment, confirmation, and run binding
v13managed runtime, host/device, core qualification, occupancy, and contention evidence
v14real training-path certification, fixture, model commit, adapter versions, host, and workstation qualification

The current replay format v14 uses MANIFEST_VERSION = 14. The loader supplies empty versioned sections when reading older manifests, preserving history without inferring evidence. A future-format manifest loads with a warning; only fields the current client understands can participate in its checks.

Limits

Replay captures launch identity; it does not claim bit-for-bit numerical equivalence across accelerator families. Reward Integrity separately audits the exact retained outputs used during verifier-guided training. See Reward Integrity for same-output sentinel behavior, capture coverage, and reviewed pause rules.