BFM-decomp/docs/wave-metrics.md

Crack-wave performance metrics

Evolvable reference (docs/ layer). Started 2026-08-04 (Phase 30, sessions S33–S37). Append one row per wave. The point of this file is that the wave harness has knobs — the prompt, the concurrency shape, the model tier — and their effects are only visible if the numbers are recorded in one place. Two of them were already measured to matter a lot (§ below).

Sources, all derivable — do not hand-transcribe (R33):

• counts/tokens/duration: the Workflow completion notification (agent_count, subagent_tokens, duration_ms)
• per-agent wall-clock + parallelism: min/max(timestamp) per agent-*.jsonl under ~/.claude/projects/.../subagents/workflows/<runId>/
• banked: the whole-binary gate log (.run/s*_gate.log) — the gate is the arbiter, never the agents' claim (G3/P9)

The table

wave	script	targets	templ ins	agents	tokens	wall	parallelism	median / slowest agent	claimed	banked	+members
1 (S33)	s33w.js	17	26,227	25	6.87M	169 min	4.8×	33 / 63 min	15/17	15 (13+2 rec)	65
2 (S34)	s34w.js	13	37,943	19	4.51M	208 min	2.7×	20 / 79 min	11/13	10 (9+1 rec)	18
3 (S35)	s35w.js	13	17,644	14	2.73M	140 min	2.1×	9 / 51 min	13/13	13 (12+1 rec)	21
4 (S36)	s36w.js	14	16,844	17	2.99M	136 min	2.5×	14 / 48 min	14/14	14 (13+1 rec)	26
5 (S37)	s37w.js	16	16,884	19	2.71M	82 min	3.8×	9 / 50 min	16/16	16 (16+0 rec)	26

All waves: Sonnet drafters with an Opus escalation rung (§136i), whole-binary byte-gate as the sole arbiter, R22 clean-fleet 140/140 after every banked batch.

Finding 1 — the PROMPT is the lever, and the agents write it

Bank rate 76% → 77% → 100% → 100% → 100%, with the models and the gate held constant. The only variable was the prompt's extra block.

The jump came from STEP 0: grep -rn "<MAGIC>" src/ using a distinctive literal from the target .s, placed ahead of engine_core.h in §136c's search order — because §136c's first two steps are same-TU/shared-header scoped and structurally cannot reach a banked twin in a different overlay's TU, which is where the big template classes live.

That step came from a wave-2 agent's index_gap report. The wave schema asks every agent for index_hit / index_gap; harvesting those into the next wave's prompt is what compounds. By wave 4 most agents cited step 0 by name and reported index_gap: none.

Honesty caveat: wave 3–5 targets also trended easier (smaller families, more banked twins available), so the prompt is not solely responsible for 100%. The claim that holds without qualification is the direction and the mechanism, not the exact percentage.

Finding 2 — pipeline() vs batched parallel(): 40% faster on 14% more targets

Waves 1–4 split each wave into two parallel() batches, which is a hard barrier — batch 2 cannot start until batch 1's slowest agent finishes. Visible directly in the launch timestamps as 37–50 minute dead gaps at each boundary.

	wave 4 parallel()	wave 5 pipeline()
targets	14	16
wall-clock	136 min	82 min
parallelism	2.5×	3.8×
median agent	14 min	9 min

The batching existed to dodge S10's 30-wide server throttle, but the harness already caps workflow agents at min(16, cores-2) and these waves are 13–16 targets — so it bought nothing and cost a barrier. Use .run/s37w.js's execution block for every future wave.

The remaining floor is real: the slowest single agent is still ~50 min at 200–300 turns. That is genuine match_one iteration (one agent tested 470+ statement orderings on a 793-instruction function). Wall-clock cannot go below the slowest chain, so the lever there is target selection, not concurrency.

Finding 3 — economics

Roughly 170k–300k tokens per banked head across waves 3–5 (the stable regime). But a head is not the unit of value: the unit is the head plus its propagated members, and the sweep rate is a property of the FAMILY, not the wave — wave 3 swept 21/21 while wave 2 swept 18/165, because wave 2's big families are per-location variants that do not template (settled by probe: remapped member is BUILD OK and byte-different ⇒ genuine per-member codegen). Rank targets by open templatable instructions derived from corpus.stubs, and expect the sweep yield to be bimodal, not average.

Finding 4 — a perfect gate is a signal that the prompt rules landed

Wave 5 is the first 16/16 banked with ZERO reconcile. Waves 1–4 each needed 1–2 declaration reconciles after the gate; wave 5 needed none. The difference is that by then the prompt carried both §138 rules (the (macro-shape, TU-shape) pair, and the reconcile-direction line-number check) plus the wave-4 lesson that two targets sharing a TU can create each other's conflicts. The reconcile lane is the fallback, not the plan — when it goes quiet, the prompt is doing its job. Lifetime reconcile record across the session: 21/22.

How to add a row

After a wave's gate + sweep land, before the commit:

# tokens / agents / duration: from the Workflow completion notification
# parallelism + per-agent spread:
python3 - <<'PY'
import json, glob, datetime
d = "<runId dir under ~/.claude/projects/.../subagents/workflows/>"
T = lambda s: datetime.datetime.fromisoformat(s.replace("Z","+00:00"))
ag=[]
for f in glob.glob(f"{d}/agent-*.jsonl"):
    ts=[T(r["timestamp"]) for r in map(json.loads, open(f)) if r.get("timestamp")]
    if ts: ag.append(((max(ts)-min(ts)).total_seconds()/60, min(ts), max(ts)))
wall=(max(a[2] for a in ag)-min(a[1] for a in ag)).total_seconds()/60
tot=sum(a[0] for a in ag); ag.sort()
print(f"{len(ag)} agents · wall {wall:.0f}m · agent-time {tot:.0f}m · {tot/wall:.1f}x · "
      f"median {ag[len(ag)//2][0]:.0f}m · slowest {ag[-1][0]:.0f}m")
PY

Related: cookbook §138 (the reconcile/search rules these waves produced), memory wave-prompt-seed-step0-and-gaps (what to seed the next prompt with), docs/effort-map.md (Ultracode/breadth policy), docs/calibration.md (templatability measurements).