exoskeleton/code/experiments/README.md

Evidence pipeline

This package is the experiment/evidence layer. The pre-prototype executors in experiments.executors adapt immutable trial records to H1 retargeting, H2 synthetic sensitivity, and H3/H4 rigid-body simulation backends.

Minimal workflow

From the repository root with PYTHONPATH=code:

python -m experiments.cli plan \
  --spec code/config/experiments/h1_smoke.json \
  --output /tmp/h1-plan.json

python -m experiments.cli run \
  --plan /tmp/h1-plan.json \
  --batch-dir output/experiments/h1-smoke \
  --executor experiments.executors:execute_h1_retargeting

Available executor/config pairs are:

execute_h1_retargeting       h1_smoke.json / h1_calibration.json
execute_h2_synthetic         h2_smoke.json / h2_calibration.json
execute_bilateral_simulation smoke.json / bilateral_calibration.json

Auditable second-stage calibration specifications are:

execute_h1_retargeting       h1_calibration_v2.json
execute_h2_synthetic         h2_calibration_v2.json
execute_bilateral_simulation bilateral_calibration_v2_energy.json
execute_bilateral_simulation bilateral_calibration_v2_network.json

The v3 redesign separates branch-crossing evidence from contact/energy stress:

execute_h1_retargeting       h1_calibration_v3.json
execute_bilateral_simulation bilateral_calibration_v3_stable_contact.json
execute_bilateral_simulation bilateral_calibration_v3_energy_challenge.json

h1_calibration_v3.json is a deterministic branch-regression fixture, not statistical tail-latency evidence. The bilateral v3 stable grid uses three exogenous seed groups shared across haptic gains. The synthetic energy challenge shares the corresponding k=3200 N/m groups, is explicitly ineligible for H3, and must be analyzed only with its H4/challenge metric configuration.

The bilateral network specification is a gated Stage B template. Its requires_stage_a_selection flag means the haptic parameters are placeholders; do not execute it as a locked study until the energy/gain Stage A acceptance gate has passed.

An executor callable receives one immutable trial mapping and returns:

TrialPayload(
    samples={"time": time_array, "...": sample_array},
    events=[{"sample_index": 10, "event": "contact"}],
    metadata={"backend": "simulation"},
)

Every sample array must have the same first dimension. Object arrays are rejected. Successful trials are committed by one atomic directory rename; failures are retained separately and may be retried with --resume.

Validate a completed batch:

python -m experiments.cli validate \
  --batch-dir output/experiments/h1-smoke

Recompute independent endpoints and paper source-data:

python -m analysis.make_paper_artifacts \
  --batch-dir output/experiments/h1-smoke \
  --metric-config code/config/experiments/metrics_h1_calibration.json

Use the matching independent metric configuration:

h1_calibration.json                  metrics_h1_calibration.json
h1_calibration_v2.json               metrics_h1_calibration_v2.json
h2_calibration*.json                 metrics_h2.json
bilateral_calibration.json           metrics_bilateral.json
bilateral_calibration_v2_*.json      metrics_bilateral_v2.json
h1_calibration_v3.json               metrics_h1_calibration_v3.json
bilateral_calibration_v3_stable_contact.json
                                        metrics_bilateral_v3_stable_contact.json
bilateral_calibration_v3_energy_challenge.json
                                        metrics_bilateral_v3_energy_challenge.json

The bilateral configurations derive H3 for every mapping/supervisor condition, but their H4 tables contain only tank-supervised methods; PO/PC and bypass conditions cannot be silently mixed into a tank audit.

The declared minimal storage contract is JSON for manifests/plans, NPZ for numeric sample arrays, JSON Lines for events/trial metrics, and CSV for paper source-data. No Parquet dependency is required.

Pairing and random numbers

pair_id excludes the method and therefore identifies common inputs. trial_id includes the method. The trajectory, sensor, model, and network streams are derived independently with NumPy SeedSequence; their serialized states are identical across methods in the same pair.

Calibration, pilot, and locked studies must use separate specifications. A locked plan is immutable: changing a factor, method, trajectory, or seed invalidates its hashes.

Files named *_locked_template.json are deliberately not confirmatory plans. Copy and freeze them only after calibration thresholds, safety limits, repeat counts, the source commit, and the analysis configuration have been approved.