RL Task Specs: myflow -> Harbor

Concrete task contracts for the Harbor RL loop. These map directly to scripts currently in ~/repos/laude-institute/harbor/scripts.

Task 1: Validate Snapshot

• Script: scripts/myflow_validate_snapshot.py
• Purpose: fail fast on broken snapshot/split artifacts before reward labeling or gating.
• Command:

python3 scripts/myflow_validate_snapshot.py \
  --snapshot <snapshot|latest> \
  --myflow-dir data/myflow \
  --prepared-dir data/myflow_prepared \
  --require-train-events \
  --report-out data/myflow_prepared/<snapshot>/validation_report.json

• Inputs:
  • data/myflow/<snapshot>/assistant_sft.jsonl
  • data/myflow/<snapshot>/train_events.jsonl (required when --require-train-events)
  • data/myflow_prepared/<snapshot>/manifest.json
• Outputs:
  • validation report JSON
• Exit contract:
  • 0 = pass
  • non-zero = integrity failure

Task 2: Build Reward Labels

• Script: scripts/myflow_build_reward_labels.py
• Purpose: produce versioned per-event rewards for RL training and canary gating.
• Command:

python3 scripts/myflow_build_reward_labels.py \
  --snapshot <snapshot|latest> \
  --myflow-dir data/myflow \
  --out-dir data/myflow_rewards

• Inputs:
  • data/myflow/<snapshot>/train_events.jsonl
• Outputs:
  • data/myflow_rewards/<snapshot>/train_event_rewards.jsonl
  • data/myflow_rewards/<snapshot>/reward_summary.json
• Exit contract:
  • 0 = labels generated
  • non-zero = parse/IO/schema error

Task 3: Canary Promotion Gate

• Script: scripts/myflow_eval_canary.py
• Purpose: gate promotion based on reward quality and optional baseline deltas.
• Command:

python3 scripts/myflow_eval_canary.py \
  --candidate data/myflow_rewards/<snapshot>/train_event_rewards.jsonl \
  --baseline data/myflow_rewards/<baseline_snapshot>/train_event_rewards.jsonl \
  --report-out data/myflow_reports/<snapshot>/canary_gate.json \
  --min-candidate-mean 0.55 \
  --min-delta-mean 0.00 \
  --min-delta-ci95-low -0.02

• Inputs:
  • candidate rewards JSONL
  • optional baseline rewards JSONL
  • optional --rollouts for action-dominance gate
• Outputs:
  • canary gate report JSON
• Exit contract:
  • 0 = promotion gate pass
  • 1 = gate fail (expected for regressions)
  • non-zero other = runtime error

Task 4: Mine Hardcases

• Script: scripts/myflow_mine_hardcases.py
• Purpose: mine regressions/low-reward canary samples and produce next-cycle seed set.
• Command:

python3 scripts/myflow_mine_hardcases.py \
  --snapshot <snapshot|latest> \
  --prepared-dir data/myflow_prepared \
  --candidate-rewards data/myflow_rewards/<snapshot>/train_event_rewards.jsonl \
  --baseline-rewards data/myflow_rewards/<baseline_snapshot>/train_event_rewards.jsonl \
  --out-dir data/myflow_hardcases \
  --top-k 100

• Inputs:
  • prepared canary/train splits
  • candidate rewards
  • optional baseline rewards
• Outputs:
  • hardcases.jsonl
  • next_train_seed.jsonl
  • hardcase_summary.json
• Exit contract:
  • 0 = hardcases emitted
  • non-zero = missing inputs / parse errors

Recommended Harbor Flow Task Names

1. myflow-validate-snapshot
2. myflow-build-reward-labels
3. myflow-eval-canary
4. myflow-mine-hardcases

Use these names in Harbor task orchestration so docs/runbooks stay stable.

Executed Verification (2026-02-18)

Executed end-to-end on a deterministic fixture snapshot:

1. prepare_myflow_dataset.py (30 rows input)
2. myflow_validate_snapshot.py -> PASS
3. myflow_build_reward_labels.py -> labels generated
4. myflow_eval_canary.py -> Promotion gate: PASS
5. myflow_mine_hardcases.py -> hardcases + next seed generated

Artifact root used during verification:

• /var/folders/.../tmp.5arfBojfhp/* (temporary run directory)