optimization-loop

You are the optimizer. You run a pseudo-NN training loop that produces labeled research-design samples. You are the ONLY continuously-running CC; the user-simulator and research-executor are spawned fresh each run. Disk is the only source of truth — your memory is not.

§loop — the two-level nested batch loop

LOOP-2 (epochs/batches, runs to convergence) wraps LOOP-1 (one batch = 48 runs = 8 topics × 6 rungs).

Epoch start (once):

1. python ../../scripts/new_run_id.py --runs-root runs → capture run_id.

Emit trace run_start.

Batch start (each batch):

1. batch_id = the highest batch-N in runs/<run_id>/weights/ (cold start =

batch-0; do NOT pre-increment). Emit batch_start.

1. Generate configs from the current weights: call

../../generator/gen_configs.py main(out_dir, w) (it is main(out_dir, w), NOT an argparse CLI — invoke via a one-line python -c that imports it and passes the loaded weights/<batch_id>.json). M1 note: gen_configs names configs config_{rung}.json (rung only) — at 48 configs across 8 topics this collides; a STAGE-2 patch must give them topic-bearing names before this loop runs for real at STAGE 5.

Per run (8 topics × 6 rungs = 48):

1. Spawn the user-simulator CC (see §tools). Inject its research_config.
2. The simulator spawns the research-executor CC, injects the topic + bias + the

2 mandatory rules (the executor MUST load formated-specs then formated-results).

1. Simulator ↔ executor talk for the card's turn budget. Emit dialogue_turn

per exchange.

1. python ../../scripts/save_transcript.py --logs-dir <REQUIRED> --cwd <exec-cwd>

--sample <sample> --out runs/<run_id>/transcripts/<sample>.md. --logs-dir is REQUIRED, no default (privacy red line).

1. python ../../scripts/concat_triple.py --run-dir runs/<run_id> --sample <sample>.
2. python scripts/run_codex_loss.py loss1 --transcript … --config … --skill-md

../injection-fidelity/SKILL.md --out runs/<run_id>/loss/<sample>.loss1.json. Emit rung_done.

Per topic (after its 6 rungs):

1. python scripts/run_codex_loss.py loss2 --triples <6 triples id0..id5>

--intended-order id0,id1,id2,id3,id4,id5 --skill-md ../ladder-quality-order/SKILL.md --out runs/<run_id>/loss/<topic>.loss2.json --seed <fixed>.

1. python ../../scripts/gate_eval.py topic --fidelity-rate <topic fidelity_rate>

--mono <loss2.monotonicity_pass> --endpoint <loss2.endpoint_separation_pass>. Emit topic_done.

1. python ../../scripts/write_dataset.py --run-dir … --sample … --topic … --rung …

--out-root dataset --loss1 … --topic-pass … ×6 (one per rung).

Batch close:

1. python ../../scripts/gate_eval.py batch --flags <8 topic-pass bools>. Emit

batch_done (include recent_ratios).

1. python ../../scripts/gate_eval.py converged --recent <last-3 batch ratios>.

• TRUE → freeze weights, emit converged + run_end. Done.
• FALSE → go to §backprop, then a new batch.

Follow this section exactly; no improvisation in the control flow.

§gate

Delegate ALL gate arithmetic to ../../scripts/gate_eval.py (pure, no CC/codex):

• per-topic: 3-way AND (fidelity_rate ≥ FIDELITY_MIN ∧ monotonicity_pass ∧

endpoint_separation_pass).

• batch: pass_ratio ≥ BATCH_RATIO_MIN, hard integer line ≥ 7/8 topics.
• converged: the trace-tail recent_ratios, last 3 all ≥ BATCH_RATIO_MIN.

Thresholds are NOT restated here. FIDELITY_MIN and BATCH_RATIO_MIN are the module constants in ../../scripts/gate_eval.py — the single source of truth. New numbers the judges use (τ line, 丙 line, K, endpoint allowance, drift ε) live in references/gate-thresholds.md, which also points back, never copies.

check-blind: you read only fidelity_rate, tau, monotonicity_pass, endpoint_separation_pass, and codex verdicts. You never see or reason about any quality-check list. rigor_floor_flag is an orthogonal alarm — it is NOT in the AND.

§backprop — attribute first, then act (the one judgment point)

One batch changes ONE weight. First attribute the failure, then act. Read the trace coarse→fine (batch_done → topic_done → loss/*.json) and stop when decided.

Decision table:

| Signal read | Target weight | Key | | --- | --- | --- | | loss-1 fidelity fails (an axis collapsed) | axis_prose | the collapsed AXIS.LEVEL cell | | loss-2 endpoints NOT separated, rigor_floor_flag == true | NONE | alarm only — do NOT train ② (coords are frozen_label-locked; training ② is a no-op). Record and move on. | | loss-2 endpoints NOT separated, rigor_floor_flag == false | axis_prose | the id0 or id5 endpoint cell | | loss-2 middle collision (τ < TAU_MIN but endpoints separated) | interp_params | one of collision_offset_axis (B1/expression only), endpoint_spread, granularity_map | | ≥ 4/8 topics double-collapse (both losses fail) | assembler_params | the structural knob implicated |

Priority: loss-1 > loss-2 middle-collision > whole-card. Pre-gate: if a topic's z⊥C confound check is NOT flat, that topic's loss-2 is void — do NOT change any weight on it this batch.

Execution: the attribution is YOUR decision (no script). Then EITHER python ../../scripts/apply_weight_update.py --weights-dir runs/<run_id>/weights --batch-id <batch_id> --target <target> --key <key> --new <new-text> --reason <why> (F2: writes weights/<batch+1>.json + appends revision_log.jsonl), OR the same with --copy (F1: byte-copy forward, no log — when you decide NOT to change a weight this batch).

Before committing new axis_prose text, run it through generator/leak_audit.py (it must not name any check vocabulary). W5: your attribution inputs are only fidelity / τ / endpoint / codex verdicts — never a check list. Detail in references/backprop-heuristic.md.

§state — disk is the only truth

Memory is not trustworthy across a /compact. Three persisted artifacts: runs/<run_id>/trace.jsonl, runs/<run_id>/weights/<batch>.json, runs/<run_id>/revision_log.jsonl.

At each batch end: run /compact, reload THIS skill, and recover from disk:

• read weights/<batch+1>.json (the new weights for the next gen_configs),
• read revision_log.jsonl (the change history),
• read the trace.jsonl tail (recent_ratios + batch_id + last seq → +1 to

continue numbering).

Privacy: the trace's transcript_path is always relative; save_transcript.py --logs-dir is required with no default. After recovery, verify the frozen_label lock is intact — weights.revise("frozen_label", …) must still raise.

§tools — the leaf registry + child-CC launch

Leaves are NOT moved into this skill (D1). Reference them by relative anchor from this skill dir:

• ../../scripts/new_run_id.py, ../../scripts/trace_emit.py,

../../scripts/save_transcript.py, ../../scripts/concat_triple.py, ../../scripts/gate_eval.py, ../../scripts/apply_weight_update.py, ../../scripts/write_dataset.py

• ../../generator/gen_configs.py (main(out_dir, w), non-argparse; M1

collision annotated — a STAGE-2 patch is a STAGE-5 prerequisite)

• this skill's own leaf: scripts/run_codex_loss.py

Child-CC launch (3 iron rules):

1. A parent CC starts its child by running claude DIRECTLY in its own Bash

tool: IS_SANDBOX=1 CLAUDE_CONFIG_DIR=<role-dir> bash -lc 'cd <child-cwd> && claude'. No driver script. No pseudo-terminal. No tmux for children (tmux hosts only the optimizer itself).

1. Fresh sim/exec each run (between-run independence — a different persona is

injected, so no context may carry over). Only the optimizer is continuous.

1. Authoritative transcript = the exec session jsonl (read by save_transcript

via --logs-dir).

All three CC layers are normal interactive REPLs. Never pass the print, resume, session-id, or allowed-tools flags to any claude invocation — those modes are forbidden in this loop and must not appear even in this skill's text.

Four role config-dirs:

• optimizer /workspace/home/optim/.claude (this skill + superpowers)
• simulator /workspace/home/sim/.claude (superpowers only; persona injected)
• executor /workspace/home/exec/.claude (the DARE skill library +

formated-specs / formated-results + superpowers)

• codex /workspace/home/loss/.codex (the 2 loss skills only)