Sop Ft Orchestrate

SOP Fine-tuning Orchestrator

Prime Directive: Never Stop Until the Goal Is Reached

IMPORTANT: Never stop unless iterations_substantive >= budget.max_pipeline_iterations, accuracy criteria met, the failure requires locked external systems, or hardware changes.

This means:

• Phase fail → diagnose, fix, restart. Do not report and wait.
• /sop-rca delegation FAILED (skill missing or errored) → use the fallback diagnostics in Step 8b for that one iteration; document the failure in run_state.yaml; retry /sop-rca next iteration.
• max_pipeline_iterations reached → write PARTIAL report (Step 9) with remaining gaps and recommended fixes.
• Hang/crash → watchdog diagnoses and restarts autonomously, including code-level fixes. All code fixes follow Override Policy — never modify plugins/.

Iteration Budget — ENFORCED

The budget is max_pipeline_iterations (default 8, set in inputs.yaml). Track it explicitly under run_state.yaml:

iteration_budget:
  max_pipeline_iterations: 8        # from inputs.yaml
  iterations_substantive: 0         # cumulative count of retraining iters (categories below)
  iterations_eval_only: 0           # threshold sweeps, etc — DO NOT count against budget
  iterations_remaining: 8           # = max - substantive
  rca_runs_completed: 0             # invariant: must equal count(eval_history entries that miss success_criteria) when run_rca: true
  attempts:                         # one entry per substantive iter — for the pre-PARTIAL gate
    augment_config_change: []       # list of distinct configs tried (e.g., ["DMCQ confusion", "DMCQ adjacent"])
    training_config_change: []      # e.g., ["LR corrected to 5e-6"]
    ddm_training_config_change: []  # e.g., ["epochs 30", "epochs 30 + RandomResize bilinear"]
    code_change: []                 # infra fixes — free, do not count against budget

Iteration types — only "substantive" counts against the budget:

Only substantive counter matters — 2 retrains + 3 threshold sweeps = 2/8 budget used. Update iteration_budget after EVERY iteration.

(Diagnostic heuristics for the rare case where /sop-rca is unavailable have moved to Step 8b-fallback. Do not consult them as a routine substitute for RCA)

Watchdog Policy

After every job launch: run with run_in_background=true; do not poll or sleep.

Training watchdogs remain file-based. Eval uses the blocking skill envelope (Step 7a/7b) — no watchdog log to monitor; the skill returns when eval-ms reports a terminal status.

If Monitor silent >10 min (training): check watchdog process alive; restart if dead.

Override Policy: Never Modify Plugin Files During a Run

Files inside plugins/ are read-only. Copy any file needing changes to <run_dir>/overrides/ first, then modify the copy. Pass the override via env vars

Record in run_state.yaml under overrides:. Never use plugins/ paths.

Pipeline Overview

DDM and VLM are independent retry axes — each can be retrained without touching the other. The table shows which steps run for each action type:

[Step 1] Prerequisites
      │
      ▼
[Step 2] Initialize run_state.yaml
      │
      ▼
[Step 3] Import dataset ── run ONCE, never repeated
      │
      ▼
[Step 4] Data Augmentation  ◄── augment-config-change (DMCQ, confusion_map, MCQ max_chunk_len)
         BCQ · MCQ · GoldenGQA · GQAs · DMCQ · DS            [DDM unaffected — skip Step 5]
      │
      ▼
[Step 5] DDM-Net Training  ◄── DDM training-config-change (resolution, RandomResize, epochs)
         ResNet-50 boundary detector                         [skip Step 4 + Step 6]
      │
      ▼
[Step 6] VLM Training  ◄────── VLM training-config-change (LR, fps/max_frames)
         Cosmos-Reason2-2B SFT + LoRA merge if applicable   [skip Step 4 + Step 5]
      │
      ▼
[Step 7] Evaluate
      ├── 7a By-Action  (VLM in isolation, perfect segmentation)
      └── 7b E2E        (DDM + VLM full pipeline)
            │  [DDM-only retrain → Step 7b only; VLM or augment → both 7a + 7b]
      ▼
[Step 8] RCA + Decision Loop
      │
      ├── ALL criteria met ─────────────────────────────────► DONE ✅
      │
      ├── eval-config-change ──────────────────────────────► Step 7 → Step 8
      │
      ├── DDM training-config-change ─────────────────────► Step 5 → Step 7b → Step 8
      │   resolution · RandomResize · epochs                 (augment + VLM unchanged)
      │
      ├── VLM training-config-change ─────────────────────► Step 6 → Step 7 → Step 8
      │   LR · fps/max_frames mismatch                       (augment + DDM unchanged)
      │
      ├── augment-config-change ───────────────────────────► Step 4 → Step 6 → Step 7 → Step 8
      │   DMCQ · confusion_map · MCQ max_chunk_len           (DDM unchanged — skip Step 5)
      │
      ├── code-change ──────────────────────────────────────► override fix → retry phase
      │
      └── max iterations or manual ──────────────────────────► Step 9 PARTIAL + hand off

Reference Documents

• ${SKILL_DIR}/references/inputs-template.yaml — all config fields
• ${SKILL_DIR}/references/decision-logic.md — RCA action type → pipeline response mapping
• ${SKILL_DIR}/references/prerequisites.md — pre-flight checklist
• ${SKILL_DIR}/references/run-state-schema.yaml — full run_state.yaml template with field annotations
• ${SKILL_DIR}/references/gqas-preflight.md — local vLLM launch, probe-validate, and output check (Steps 1–3)
• ${SKILL_DIR}/references/claude-gqas-setup.md — Claude API backend setup: patch, restart, verify
• ${SKILL_DIR}/scripts/watch_api_job.sh — watchdog for API-polled jobs (DDM, VLM training)
• Eval jobs (/sop-by-action-eval, /sop-e2e-inference) are NOT watchdogged — the eval skills block until the eval-ms HTTP job reaches a terminal state and return a JSON envelope.
• ${SKILL_DIR}/scripts/auto_detect_splits.py — auto-detect train/test split subdirectories
• ${SKILL_DIR}/scripts/import_dataset.sh - import annotated dataset
• ${SKILL_DIR}/references/augment-config-guide.md — annotated augment_config.yaml template (DMCQ, MCQ, GQAs)

Resume Protocol

Run at the very start of every session:

ls <output_dir>/run_*/run_state.yaml 2>/dev/null | sort | tail -1

If a run_state.yaml is found:

1. Print full contents with === RESUMING RUN: <run_id> ===
2. done → skip. in_progress → verify: DDM/VLM: check API status; eval: check output file exists; augment: check augmented dir exists.
3. Continue from first non-done phase.

If no run_state.yaml exists — proceed to Step 0.

Step 0: Parse Inputs

Read the inputs.yaml (or parse natural language). Resolve all paths. Set defaults. Print config summary and confirm with user before starting:

Output Directory Contract (mandatory)

There are two trees in an orchestrated run, owned by different things:

1. <run_dir>/ — owned by the orchestrator. Contains run-state, per-iter configs, RCA reports, progress, watchdogs, and a per-iter snapshot of each eval job's output.
2. <host_results_root>/<eval_job_id>/ — owned by eval-ms. Each /sop-by-action-eval and /sop-e2e-inference invocation produces one such directory; eval-ms decides the uuid. The orchestrator never writes here, but it does two things after each eval: (a) records pointers in run_state.eval_outputs, and (b) snapshots the job's output into <run_dir>/iter<N>/{by_action,e2e}/ so the run dir is self-contained and survives any later cleanup of the eval-ms store.

<output_dir>/                              # from inputs.yaml; default ./sop_fine_tune
└── run_<YYYYMMDD_HHMMSS>/                 # = <run_dir>; created in Step 2
    ├── run_state.yaml                     # full run state — incl. eval_outputs pointers (KEPT)
    ├── progress.md, progress.html         # phase logs + live progress chart
    ├── orchestrator_report.md             # final SUCCESS / PARTIAL summary
    ├── overrides/                         # code overrides per Override Policy
    ├── watchdog_*.log                     # DDM/VLM training watchdog tails
    └── iter<N>/
        ├── augment_config.yaml            # snapshot of augment config used
        ├── ddm_train_config.yaml          # snapshot of DDM training config
        ├── train_config.toml              # snapshot of VLM training config
        ├── training.log                   # copy of VLM training log
        ├── inputs_by_action_iter<N>.yaml  # generated per-iter eval input (moved here, off top level)
        ├── inputs_e2e_iter<N>.yaml        # generated per-iter eval input
        ├── by_action/                     # snapshot of the by-action eval job output
        ├── e2e/                           # snapshot of the e2e eval job output
        ├── rca_analysis/                  # /sop-rca helper JSONs
        └── rca_report.md                  # /sop-rca formal report (REQUIRED on failure)

<host_results_root>/                       # docker-compose volume
├── <by_action_eval_job_id>/               # /sop-by-action-eval job output
│   ├── inference_results.json
│   ├── log.txt
│   └── assets/
└── <e2e_eval_job_id>/                     # /sop-e2e-inference job output
    ├── e2e_results.json
    ├── log.txt
    ├── sop_e2e_eval_log.txt
    ├── outputs_action_recognition/
    │   ├── accuracy.json
    │   ├── video_name_to_output_text.json
    │   └── action_recognition_multi_gpu.log
    └── outputs_temporal_segmentation/
        ├── f1_<thr>.json
        ├── video_to_boundaries_debug.json
        ├── video_to_ddm_info_debug.json
        ├── temporal_segmentation.log
        └── <video>.png

Hard rules:

• The orchestrator MUST pass output_dir=<run_dir>/iter<N> when delegating to /sop-rca in Step 8b (see "RCA delegation contract" below). Never let /sop-rca default to <cwd>/rca_reports/ — that splits artifacts.
• The orchestrator MUST NOT pass output_dir to /sop-by-action-eval or /sop-e2e-inference. Those skills do not accept an output-dir override — eval-ms decides where the job writes. The orchestrator captures the returned host_output_dir from each skill's JSON envelope and (a) persists it to run_state.eval_outputs.<phase>.host_output_dir, then (b) snapshots that directory into <run_dir>/iter<N>/<phase>/ (Step 7c).
• The eval snapshot is a full recursive copy of the eval-ms job dir (e.g. cp -a <host_output_dir>/. <run_dir>/iter<N>/by_action/). Copy everything the job wrote; do not hand-pick files.
• All ad-hoc analyses produced by the orchestrator (helper-script outputs, debug dumps, scratch yamls) MUST live under <run_dir>/ — never under <cwd>/ and never under any plugin path.
• One run = one <run_dir>. Resumes write into the same <run_dir> discovered by the Resume Protocol.
• Eval-ms-owned directories under <host_results_root>/ are not garbage-collected by the orchestrator.

RCA delegation contract (Step 8b call site)

/sop-rca reads paths it was told about — it does NOT construct paths from conventions. The orchestrator builds the delegation payload from run_state.eval_outputs (populated by Step 7a/7b) plus the snapshots it placed under <run_dir>/iter<N>/:

/sop-rca \
  e2e_outputs_dir=<run_dir>/iter<N>/e2e/outputs_action_recognition \
  ddm_outputs_dir=<run_dir>/iter<N>/e2e/outputs_temporal_segmentation \
  by_action_results=<run_dir>/iter<N>/by_action/inference_results.json \   # JSON is preferred; log.txt also accepted
  actions_json=<dataset_path>/actions.json \
  augment_config=<run_dir>/iter<N>/augment_config.yaml \
  ddm_train_config=<run_dir>/iter<N>/ddm_train_config.yaml \
  vlm_train_config=<run_dir>/iter<N>/train_config.toml \
  vlm_train_log=<run_dir>/iter<N>/training.log \
  output_dir=<run_dir>/iter<N> \
  iter=<N> \
  success_criteria="e2e_seq_acc>=1.0,e2e_action_acc>=0.95,by_action_acc>=0.95,ddm_f1>=0.6"

Dispatch this payload via a sub-agent (Agent tool) — see Step 8b for why this is mandatory (keeps the heavy RCA out of the orchestrator's context so RCA is always affordable). The sub-agent writes analysis JSONs to <output_dir>/rca_analysis/ and the report to <output_dir>/rca_report.md, and returns ONLY the compact RCA_RESULT: block (its skill's return contract). The orchestrator consumes that block — verifying RCA_RESULT.report_path exists and copying RCA_RESULT.typed_actions into rca_reports[] — and never needs to load the report prose.

Dataset Split Auto-Detection

If the user provides a single dataset_path (no explicit eval_dataset_path), run:

python3 ${SKILL_DIR}/scripts/auto_detect_splits.py <dataset_path>
# Success: prints TRAIN=<abs_path> and EVAL=<abs_path>
# Failure: prints diagnostic and exits 1 — set paths explicitly in inputs.yaml

Matches <stem>_train ↔ <stem>_test pairs; fails on zero or ambiguous matches.

Step 1: Prerequisites

Run all checks from references/prerequisites.md. Auto-fix soft issues (start services, install packages). Block on hard failures. Do not proceed until all pass.

Step 2: Initialize Run State

Create <output_dir>/run_<YYYYMMDD_HHMMSS>/run_state.yaml. Print path as first output line: === RUN STATE: <full/path/to/run_state.yaml> ===

State + progress files (MANDATORY — generated, never hand-edited)

run_state.yaml is edited ONLY via scripts/rs_update.py — never with sed, str.replace(), or a from-memory Edit block. Hand-editing a growing YAML file caused two real production bugs (field cross-contamination from copy-pasted entry templates, and silent no-ops when a replace anchor didn't byte-match). rs_update.py does load→mutate→dump with round-trip validation:

python3 ${SKILL_DIR}/scripts/rs_update.py <run_dir> set iteration=N phase_status.rca=done
python3 ${SKILL_DIR}/scripts/rs_update.py <run_dir> append eval_history '<json row>'
python3 ${SKILL_DIR}/scripts/rs_update.py <run_dir> append rca_reports  '<json entry>'
python3 ${SKILL_DIR}/scripts/rs_update.py <run_dir> budget --substantive +1 --rca +1

progress.md and progress.html are GENERATED from run_state.yaml — never authored by hand. They are a pure projection of eval_history + phase_status, so they cannot drift behind the real state. Regenerate BOTH after every eval (Step 7c) and at Step 2 init:

python3 ${SKILL_DIR}/scripts/gen_progress.py <run_dir>
# prints: === PROGRESS (local file): <run_dir>/progress.md ===  (and progress.html)

(For richer chart annotations, eval_history rows may carry an optional chart sub-dict: {ph, t, lr, samp, qas} — gen_progress.py reads it; absence is fine.)

Both files + the canonical run_state are what the Step 8a.0 gate asserts are current — not optional. gen_progress.py/rs_update.py also refresh ~/.cache/sop-ft-orchestrate/active_run so the harness gate can locate this run_dir even though it lives outside the harness cwd.

Initialize from ${SKILL_DIR}/references/run-state-schema.yaml. Invariants:

• eval_vision_params — re-populate after every VLM training run; must match train_config.toml [custom.vision].
• phase_status — in_progress BEFORE delegating; done AFTER.
• rca_reports — every failed iteration must have an entry.

Step 3: Phase 1 — Import Dataset (run once)

Check: curl -s http://localhost:8100/api/v1/datasets | python3 -c "import json,sys; d=json.load(sys.stdin); print([x['id'] for x in d.get('datasets',[])])"

If not imported: copy dataset to <training_bp_root>/assets/data/, then bash ${SKILL_DIR}/scripts/import_dataset.sh <dataset_id>. Verify counts.

• Do NOT re-import on subsequent iterations.
• Do import every dataset provided, including training and validation

Step 4: Phase 2 — Data Augmentation (run iter 1 + on augment-config-change)

Increment augmented_dataset_id suffix; apply RCA augmentation config changes.

GQAs backend (priority order):

Check for ANTHROPIC_API_KEY (HIGHEST PRIORITY): if set → use Claude (preferred: no GPU, ~10× faster). Set in augment_config.yaml:

gqas:
  enable: true
  llm_type: "local"         # argparse only accepts "local" or "nvidia"; routing is by model name
  llm: claude-haiku-4-5-20251001   # detected as Claude by the override; use haiku for speed/cost
  local_llm_url: ""         # not used for Claude routing
  enable_thinking: "false"
  num_qa_llm: 8
  num_qa_per_chunk: 2

Setup: ${SKILL_DIR}/references/claude-gqas-setup.md.

Fallback: NIM: valid NGC_API_KEY → llm_type: "nvidia". After 429 for >5 min → local.

Fallback: local: vLLM container vllm/vllm-openai@sha256:2e08b462bb444a6da8a84a533f09024c61617574e67386efe4a723a0633fcc6a with Qwen/Qwen3-8B (no --reasoning-parser qwen3). Stop server after augment. Alt: Qwen/Qwen3.5-27B (requires --reasoning-parser qwen3).

• Never disable GQAs. Fix start failures; do not set enable: false.

GQAs Pre-flight (local vLLM only)

Complete all three steps in ${SKILL_DIR}/references/gqas-preflight.md before augmenting: launch, probe-validate, post-augmentation output check.

Write phase_status.augment: in_progress to run_state.yaml, then delegate to /sop-data-augmentation.

On success: run Step 3 output check from ${SKILL_DIR}/references/gqas-preflight.md; write augmented_dataset_id and phase_status.augment: done.

On failure: diagnose and fix; Do not skip GQAs unless genuinely unresolvable.

Keep 2 newest <label_data_id>_augmented_*/ dirs. Via docker exec <data_gen_container>:

docker exec <data_gen_container> bash -c \
  "ls -dt /workspace/assets/data/<label_data_id>_augmented_*/ 2>/dev/null | tail -n +3 | xargs -r rm -rf"

Step 5: Phase 3 — DDM Training (run iter 1 + on DDM training-config-change)

Set: num_gpus = GPUs detected; batch_size: < 25 GB → 4, < 50 GB → 16, else → 32.

Re-run: apply RCA changes; increment ddm_config_version.

DDM resize augmentation policy (overrides /sop-rca-plugin:sop-rca Pattern 10):

When RandomResize is recommended: never use [bilinear, bicubic, nearest] on datasets < 20 videos. Use one option at a time:

• interpolation: [nearest] — try first; strongest encoding-robustness signal
• interpolation: [bilinear] — fallback if convergence too slow (loss not decreasing after 10 epochs)

If neither improves E2E F1, disable RandomResize.

Write phase_status.ddm_train: in_progress, then delegate to /sop-ddm-finetuning with <dataset_id> and <validation_dataset_id> (if available).

On completion: record ddm_job_id, ddm_checkpoint, ddm_best_f1, ddm_best_loss. Write phase_status.ddm_train: done.

Keep top-3 epoch ckpts by F1; delete rest + last.ckpt; delete stale ddm_inference*.ckpt. Via docker exec <ddm_container>:

docker exec <ddm_container> bash -c "
  ls /workspace/sop-ddm-ftms/assets/results/<job_id>/train/<job_id>/epoch_epoch=*-val*.ckpt 2>/dev/null \
    | sort -t= -k2 -rn | tail -n +4 | xargs -r rm -f
  ls /workspace/sop-ddm-ftms/assets/results/<job_id>/train/<job_id>/epoch_epoch=*-val*.ckpt 2>/dev/null \
    | grep -q . && rm -f /workspace/sop-ddm-ftms/assets/results/<job_id>/train/<job_id>/last.ckpt
  find /workspace/sop-ddm-ftms/assets/results/ -name 'ddm_inference*.ckpt' | sort | head -n -1 | xargs -r rm -f
"

Note: If best val/F1 < 0.5, log a manual note. Do NOT stop — continue evaluating.

DDM Training Watchdog

bash ${SKILL_DIR}/scripts/watch_api_job.sh \
  DDM <DDM_BASE_URL> <JOB_ID> <LOG_PATH> <TIMEOUT> 60 \
  >> <run_dir>/watchdog_ddm.log 2>&1

tail -f <run_dir>/watchdog_ddm.log | grep --line-buffered -E "DDM_DONE|DDM_FAILED|DDM_HANG|DDM_TIMEOUT"
tail -f <run_dir>/watchdog_ddm.log | grep --line-buffered "F1_UPDATE"

Monitor 1: DDM_DONE → mark done; DDM_FAILED → read log tail, fix, restart; DDM_HANG/DDM_TIMEOUT → kill container, fix, restart. Monitor 2: F1_UPDATE → log F1 and loss.

If Monitor silent >5 min: check ps aux | grep watch_ddm; restart watchdog if dead.

Write <run_dir>/progress.md snapshot.

Step 6: Phase 4 — VLM Training (run iter 1 + on VLM training-config-change or augment-config-change)

Batch Size and Learning Rate

Set train_batch_per_replica in train_config.toml:

• Sufficient VRAM → 4; Low VRAM → 1. If OOM at batch=4, retry with batch=1.

Learning rate — MODE-AWARE (LoRA vs full fine-tune). Check whether [policy.lora] is present in train_config.toml:

• LoRA run ([policy.lora] present): small datasets (< 20 videos) use optm_lr = [1.5e-5, 1.5e-5, 1.5e-5].
• Full fine-tune run (no [policy.lora]): small datasets (< 20 videos) use optm_lr = [5e-6, 5e-6, 5e-6]
• Larger datasets (≥ 50 videos): service default may be acceptable; validate after iteration 1.
• Repeated action pair confusion despite correct DMCQ → check LR first.

VLM Training Watchdog

bash ${SKILL_DIR}/scripts/watch_api_job.sh \
  VLM <CR_BASE_URL> <JOB_ID> <LOG_PATH> <TIMEOUT> 120 \
  >> <run_dir>/watchdog_vlm.log 2>&1

tail -f <run_dir>/watchdog_vlm.log | grep --line-buffered -E "VLM_DONE|VLM_FAILED|VLM_HANG|VLM_TIMEOUT"
tail -f <run_dir>/watchdog_vlm.log | grep --line-buffered "status=running"

Monitor 1: VLM_DONE → proceed to 6a; VLM_FAILED → read log tail, fix; VLM_HANG/VLM_TIMEOUT → kill CR container, fix, restart.

If Monitor silent >10 min: check ps aux | grep watch_vlm; restart watchdog if dead.

When re-running: apply RCA config changes; increment vlm_config_version.

Write phase_status.vlm_train: in_progress, then delegate to /sop-cr-finetuning with augmented_dataset_id.

6a. Record training mode

The Step 6 delegation to /sop-cr-finetuning runs the whole training procedure end-to-end — including its own LoRA detection and merge. On return, <results_dir>/<timestamp>/safetensors/step_<N>/ is always a self-contained, inference-ready HF model.

Record training mode by checking whether cr-finetuning's merge step left a step_<N>_lora_adapter/ sibling next to step_<N>/:

• Sibling exists → LoRA run. cr-finetuning Phase 3b created the backup during the in-place merge.
• Sibling absent → Full fine-tune. No merge was needed; step_<N>/ was written as a full model directly by training.

Set in run_state.yaml:

vlm_inference_path is the same path in both modes. No LoRA-vs-full branching is needed in Step 7.

Write phase_status.vlm_train: done to run_state.yaml.

Keep top-3 job_ids by eval_history.e2e_seq_acc; delete *.safetensors and adapter_model.* from others. Never delete log.txt, configs, or tokenizer files. Via docker exec <cr2_container>:

KEEP="<job_id_1> <job_id_2> <job_id_3>"  # from run_state.yaml eval_history
docker exec <cr2_container> bash -c "
  for d in /workspace/sop-cr-ftms/assets/results/*/; do
    j=\$(basename \"\$d\")
    echo '$KEEP' | grep -q \"\$j\" && continue
    find \"\$d\" -name '*.safetensors' -o -name 'adapter_model.*' | xargs -r rm -f
  done
"

Step 7: Phase 5 — Evaluation

Run both evaluations in parallel if possible, otherwise sequentially.

Eval is API-driven. Both /sop-by-action-eval and /sop-e2e-inference POST to the BP evaluation-ms HTTP service (default localhost:32090) and return a structured JSON envelope on stdout. Eval-ms decides the output directory (uuid under its RESULTS_ROOT); the orchestrator only records the host-side paths from the envelope.

Pre-eval requirements:

• run_state.training_job_id and run_state.val_dataset_id must be set (populated when the training and dataset registration steps complete).
• For E2E: run_state.ddm_training_job_id must be set.
• Read fps, max_frames, total_pixels from train_config.toml [custom.vision] and write them into the eval inputs.yaml. Never use eval-script defaults.
• run_state.host_results_root must point at the host directory that docker-compose maps to /workspace/sop-eval-ms/assets/results. Default: <bp_deployment_root>/assets/results.

7-pre. eval-ms reachability check

curl -fsS http://${EVAL_HOST:-localhost}:${EVAL_PORT:-32090}/health

Non-2xx → stop and tell the user to bring eval-ms up (docker compose up evaluation-ms from the BP deployment root) before retrying.

7a. By-Action Evaluation

Generate inputs_by_action_iter<N>.yaml into <run_dir>/iter<N>/ from ${SKILL_DIR}/references/by-action-eval-template.yaml. Required fields: training_job_id (from run_state), val_dataset_id, host_results_root. Set backend: transformers for LoRA evaluations; vllm for full-FT. Pass fps from train_config.toml.

Write phase_status.eval_by_action: in_progress, then delegate to /sop-by-action-eval with the generated inputs.yaml.

Capture the JSON envelope. The skill emits a single JSON line on its last stdout. Parse it and persist into run_state.eval_outputs.by_action:

eval_outputs:
  by_action:
    eval_job_id: <envelope.eval_job_id>
    host_output_dir: <envelope.host_output_dir>
    snapshot_dir: <run_dir>/iter<N>/by_action       # full copy made in Step 7c
    inference_results_json: <envelope.artifacts.inference_results_json>
    log: <envelope.artifacts.log>
    overall_accuracy_evalms: <envelope.headline_metrics.overall_accuracy>
    overall_accuracy_authoritative: null   # filled by analyze_by_action_confusion in Step 8b

7b. E2E Evaluation

Generate inputs_e2e_iter<N>.yaml into <run_dir>/iter<N>/ from ${SKILL_DIR}/references/e2e-eval-template.yaml. Required fields: training_job_id, val_dataset_id, ddm_training_job_id, host_results_root. Set score_threshold = run_state.ddm_threshold. Pass max_frames/total_pixels through resolution_config and fps at the top level.

Write phase_status.eval_e2e: in_progress, then delegate to /sop-e2e-inference.

Capture the JSON envelope and persist into run_state.eval_outputs.e2e:

eval_outputs:
  e2e:
    eval_job_id: <envelope.eval_job_id>
    host_output_dir: <envelope.host_output_dir>
    snapshot_dir: <run_dir>/iter<N>/e2e             # full copy made in Step 7c
    e2e_results_json: <envelope.artifacts.e2e_results_json>
    accuracy_json: <envelope.artifacts.accuracy_json>
    video_name_to_output_text_json: <envelope.artifacts.video_name_to_output_text_json>
    action_recognition_log: <envelope.artifacts.action_recognition_log>
    temporal_segmentation_dir: <envelope.artifacts.temporal_segmentation_dir>
    temporal_segmentation_log: <envelope.artifacts.temporal_segmentation_log>
    sop_e2e_eval_log: <envelope.artifacts.sop_e2e_eval_log>
    log: <envelope.artifacts.log>

No file-existence polling, no watchdog scripts. The eval skill blocks until terminal status and returns the envelope.

7c. Record Results

Snapshot the eval job output(s) into <run_dir>/iter<N>/. For each eval run this iteration, full-copy its job dir from the envelope's host_output_dir:

# by-action (if run this iter)
mkdir -p <run_dir>/iter<N>/by_action && cp -a <eval_outputs.by_action.host_output_dir>/. <run_dir>/iter<N>/by_action/
# e2e (if run this iter)
mkdir -p <run_dir>/iter<N>/e2e        && cp -a <eval_outputs.e2e.host_output_dir>/.        <run_dir>/iter<N>/e2e/

Append eval_history entry in run_state.yaml:

eval_history:
  - iteration: N
    by_action_acc: 0.XX
    e2e_action_acc: 0.XX
    e2e_seq_acc: 0.XX
    ddm_f1: 0.XX
    ddm_threshold: 0.XX
    changes_applied: ["description of what changed vs prev iteration"]

Write progress chart — every eval, generate <run_dir>/progress.html:

import pathlib, json
tmpl = pathlib.Path('${SKILL_DIR}/references/progress-chart-template.html').read_text()
data = 'const RUN=' + json.dumps({
  'id': '<run_id>', 'dataset': '<dataset_name>', 'target': <seq_target_as_pct_0to100>,
  'iters': [
    # one entry per completed eval — all accuracy fields are 0-100 (not 0-1)
    # {'n':'1','ph':'I','t':'first','ba':18.2,'ea':33.3,'sq':0,'f1':0.848,'th':0.60,'d':7,'w':11,'m':6,'lr':'1.5e-5','note':'Baseline'},
  ]
}) + ';'
pathlib.Path('<run_dir>/progress.html').write_text(tmpl.replace('/* ITER_DATA */', data))

Schema: n=label, ph=I/II/III (I=pre-key-fix, II=key-fix-iter, III=post-fix), t=first/augVLM/vlm/ddm/eval, ba=by-action%(null if not run), ea=E2E-act%, sq=seq%, f1=DDM-F1, th=DDM-thr, d/w/m=dup/wrong/miss(int), lr=LR-string, note=one-line change, samp=total QA samples (training iters only; omit or 0 for eval-only), qas=object with per-type counts {bcq,mcq,gqa,gqas,dmcq} (omit for eval-only iters).

This is a mandatory local-file write. One row per eval (eval-only sweeps included — same as eval_history). After writing, confirm with === PROGRESS (local file): <run_dir>/progress.html ===.

Step 8: Phase 6 — RCA + Decision Loop

IMPORTANT: RCA is mandatory after every failed iteration.

• Do Not skip RCA.
• Do NOT read summary files and write your own RCA.
• Do NOT use Step 8b-fallback's heuristics as a routine substitute — they exist ONLY for the case where /sop-rca delegation actually fails (skill missing, error, no report after retry).
• Always delegate to /sop-rca. The only sanctioned skip is run_rca: false in inputs.yaml. Step 8a's pre-flight gate (below) blocks iterations that try to skip RCA.

8a. Check Success Criteria

8a.0 — Bookkeeping gate (RUN FIRST, before anything else in Step 8)

This is the executable form of the invariant promised above. It is a hard STOP, not advice. You may not check success criteria, select an action, or declare SUCCESS/PARTIAL until ALL FOUR assertions below pass. They are cheap; run them on entry to Step 8 every single iteration.

bud = run_state.iteration_budget

# (1) Every eval that has been RUN must have an eval_history[] row. One row PER eval —
#     a 3-point threshold sweep is 3 rows (disambiguated by ddm_threshold / nms_sec / changes_applied),
#     not 1. This is what was missing when the run drifted: evals happened, rows didn't.
n_evals_run = count_evals_actually_run()   # = number of /sop-by-action-eval + /sop-e2e-inference completions this run
assert len(run_state.eval_history) == n_evals_run, (
    f"BOOKKEEPING GATE: {n_evals_run} evals run but eval_history has "
    f"{len(run_state.eval_history)} rows. Append the missing row(s) NOW — including for "
    f"eval-only threshold/NMS sweeps — before proceeding. Do not batch 'consolidate later'."
)

# (2) Every FAILED eval must have a completed /sop-rca run behind it (unless run_rca: false).
#     rca_runs_completed is incremented ONLY in Step 8b after rca_report.md is verified on disk —
#     never hand-set it. eval-only iterations are NOT exempt: a failed re-eval still needs RCA.
if inputs.run_rca:
    n_failed = len([e for e in run_state.eval_history if not meets_criteria(e, inputs.success_criteria)])
    assert bud.rca_runs_completed == n_failed, (
        f"RCA GATE: {n_failed} failed evals but only {bud.rca_runs_completed} RCA runs. "
        f"Run /sop-rca (Step 8b) on the un-analysed failed eval(s) before selecting any action. "
        f"Writing your own diagnosis instead of delegating to /sop-rca is the exact violation "
        f"this gate exists to catch."
    )

# (3) Progress files must exist and be current for THIS iteration (local artifacts, Step 2).
#     progress.md gets this iteration's row; progress.html is regenerated.
assert progress_md_exists() and last_logged_iter(progress_md) == run_state.iteration, (
    "PROGRESS GATE: progress.md missing or stale (no row for current iteration). "
    "Append this iteration's row NOW (Step 7c / Step 9)."
)
assert progress_html_exists(), (
    "PROGRESS GATE: progress.html missing. Regenerate it from the template NOW (Step 7c)."
)

# (4) Each eval run THIS iteration must be snapshotted into iter<N>/<phase>/ (Step 7c).
#     This is what gives eval-only iters a folder and pins which eval_job_id was this iter.
itdir = f"{run_dir}/iter{run_state.iteration}"
for phase in evals_run_this_iter():          # subset of {"by_action", "e2e"}
    assert snapshot_nonempty(f"{itdir}/{phase}"), (
        f"SNAPSHOT GATE: {phase} eval ran this iter but {itdir}/{phase}/ is missing or empty. "
        f"cp -a <host_output_dir>/. {itdir}/{phase}/ NOW (Step 7c) before proceeding."
    )

If any assertion fires, the ONLY valid next move is to do the missing work (snapshot eval output / append rows / run RCA / write progress files), not to continue. There is no path through Step 8 that leaves these unsatisfied.

8a.1 — Success check

r = latest_eval
c = inputs.success_criteria

all_met = all([
    c.ddm_f1 is None            or r.ddm_f1            >= c.ddm_f1,
    c.by_action_accuracy is None or r.by_action_acc     >= c.by_action_accuracy,
    c.e2e_action_accuracy is None or r.e2e_action_acc   >= c.e2e_action_accuracy,
    c.e2e_sequence_accuracy is None or r.e2e_seq_acc    >= c.e2e_sequence_accuracy,
])

If all_met → SUCCESS. Go to Step 9.

To trigger PARTIAL, ALL of these must be true:

budget = run_state.iteration_budget
attempts = budget.attempts

partial_gate = all([
    budget.iterations_substantive >= budget.max_pipeline_iterations,
    len(attempts.augment_config_change)    >= 2,    # at least 2 distinct augment strategies tried
    len(attempts.training_config_change)   >= 1,    # at least 1 training-config-change tried
    len(attempts.ddm_training_config_change) >= 1,  # at least 1 ddm-training-config-change tried
])

If partial_gate is True → PARTIAL. Run Step 8b (RCA report is mandatory handoff), then Step 9.

If partial_gate is False — KEEP ITERATING. Judgment calls ("more data needed", "won't work") are NOT grounds for PARTIAL without exhausting the substantive budget across all fix categories.

Common premature-stop antipatterns:

• DDM threshold sweeps at 0.3/0.4/0.5/0.6 → eval-only, budget unused.
• DMCQ retrain regressed → alternative DMCQ mode (adjacent vs confusion) not tried yet.
• LR not verified correct → fix LR first before concluding training is stalled.

You must run Step 8b before anything in Step 8c.

8b. Run RCA (required gate before action selection)

Write phase_status.rca: in_progress, then dispatch /sop-rca as a sub-agent (Agent tool, e.g. subagent_type: general-purpose) — instruct the sub-agent to invoke the sop-rca-plugin:sop-rca skill with the payload below and to return exactly what that skill's "Invocation & Return Contract" specifies.

Why a sub-agent (this is the rule, not an option): the /sop-rca skill body, its helper-script output, and per-video analysis are large. Running it in a sub-agent keeps ALL of that in the sub-agent's context; the orchestrator receives only the compact RCA_RESULT: block. This makes RCA always affordable no matter how many iterations run — so there is never a context/token reason to skip it or hand-author the RCA yourself. Authoring your own RCA, or reading the eval files and writing your own diagnosis, is a violation (Step 8a's gate catches it); the sub-agent exists precisely to remove the temptation. The /sop-rca skill is path-driven and self-contained, so the fresh sub-agent has everything it needs from the payload.

The sub-agent reads from the durable iter<N>/ snapshot (Step 7c) plus the augment/DDM/VLM config snapshots and training log — everything the payload references lives under <run_dir>/iter<N>/:

/sop-rca
  e2e_outputs_dir:       <run_dir>/iter<N>/e2e/outputs_action_recognition
  ddm_outputs_dir:       <run_dir>/iter<N>/e2e/outputs_temporal_segmentation
  by_action_results:     <run_dir>/iter<N>/by_action/inference_results.json   # JSON preferred; log.txt also accepted by analyze_by_action_confusion.py
  actions_json:          <dataset_path>/actions.json
  augment_config:  

…(truncated)