clawback paired shadow-capture (turn-matched baseline + armed in one session)

Run .skills/scripts/run_paired.sh from the project root. It answers a specific problem with the plain .skills/ab block: two arms driven as two separate claude sessions diverge after the first sampled token (different turns, different counts — 342 vs 352 in the 75-min Haiku run), so turn k of A0 is not the counterfactual of turn k of A5. The paired tee removes that: one session, fanned to both arms, so every turn has an exact partner.

claude ──▶ tee ──┬─▶ clawback PRIMARY (armed, default A5) ─▶ Anthropic   (streamed back to claude)
                 └─▶ clawback SHADOW  (A0 passthrough)     ─▶ Anthropic   (usage tapped, discarded)

The SHADOW arm captures the no-clawback baseline while the PRIMARY arm applies (and serves claude) the armed knobs — that is the knob this skill adds: a baseline capture that does not forfeit the optimizations.

The ~2x cost, and the switch

Shadow mode sends every billable turn to both arms, so it bills your Anthropic quota ~2x for the run. That is the price of an exact paired measurement, and it is unavoidable by construction. So:

• .skills/scripts/run_paired.sh refuses to start without `--ack-2x` and prints

the warning. Passing --ack-2x is the deliberate opt-in.

• It forwards --ack-2x to benchmark/bin/tee.js, which has the same guard

(the tee refuses non-interactively without it, and prompts on a TTY).

Never run this for a long headline window without budgeting the doubled spend against your own limits — the very quota clawback exists to stretch.

Why the arms can't contaminate each other

• The two arms' forwarded bytes already differ — the PRIMARY (e.g. A5)

strips the per-request cch and/or rewrites the TTL, while the SHADOW A0 keeps the bytes pristine — so their ANTHROPIC KEYs differ and neither can warm the other's prompt cache. No salt needed, and none is used: production forwards no salt, so the harness must not either (a system-prepended salt forced whole-body re-serialization that zeroed within-arm cache warming — the confound we removed).

• The SHADOW is --passthrough, so A0 stays byte-transparent — the honest

baseline that measures exactly what production sends.

• Each instance gets its own `--state` file, so neither resumes the

other's persisted SESSION KEYs (the clawback-side carry-over guard).

Both instances are pre-seeded with one inert session so a fresh proxy does not auto-arm a baseline capture that would force the opening ~5 turns to passthrough (which would contaminate exactly the turns we want measured).

How the analyzer pairs the arms

The tee writes one NDJSON record per arm per turn, both stamped with the same internal pairSeq (a tee-only counter — it is never put on the wire: no header, query param, or body field is added to either upstream request). The outputs are named A0.ndjson (shadow) and <primary-arm>.ndjson (e.g. A5.ndjson) so analyze.js labels them by basename with no --label. The analyzer's paired path groups by pairSeq, computes mean(base − treat) over matched pairs, and reports a paired-bootstrap 95% CI that is tighter than the unpaired two-sample CI when the arms co-vary (they do — same fixture/cch/tools per pair). The reclaim total is the exact sum of per-pair deltas, not a projection.

The fragmentation table is empty on a tee run by design: the tee sees only the pre-clawback body and cannot know what each instance forwarded after its own rewrites, so it omits systemStableKey/sessionKey rather than fake them. Fragmentation is measured separately by the standalone A0-vs-A5 block (.skills/ab) and the static inspector (.skills/inspect).

Usage

# short L0 paired smoke on Haiku (8 turns, both arms => ~16 turns billed):
.skills/scripts/run_paired.sh --ack-2x --profile L0 --turns 8 \
  --model claude-haiku-4-5-20251001

# headline L0 paired run, 75-min wall clock, Haiku:
.skills/scripts/run_paired.sh --ack-2x --profile L0 --max-sec 4500 \
  --model claude-haiku-4-5-20251001 --out runs/paired-haiku-L0

# pair a different armed stack against the baseline (e.g. strip-ephemeral only):
.skills/scripts/run_paired.sh --ack-2x --primary-arm A4 --profile L0 --turns 50 \
  --model claude-haiku-4-5-20251001

Options

• --primary-arm A1..A5 — armed stack the PRIMARY applies (default A5).

A0 is rejected (both arms passthrough is meaningless).

• --profile L0|L1|L2|L3|L4 — inter-turn gap regime (default L0).
• --gap-sec N — fixed inter-turn gap; overrides --profile's range.
• --max-sec N — cap the run by wall clock (e.g. 4500 = 75 min); --turns

becomes an upper safety bound.

• --turns N — turns to drive (default 8; ≥200 for a headline; remember

each turn bills twice).

• --model ID — pin claude's model (cost control; e.g.

claude-haiku-4-5-20251001).

• --effort low|medium|high|xhigh|max — pin claude's reasoning level

(claude --effort). Resizes the cached thinking blocks, so it shifts the cache economics. Haiku has no effort control; xhigh is Opus 4.7/4.8 only.

• --prompts PATH — pty prompt file (default benchmark/prompts/coding.txt).
• --settle-sec N / --confirm-sec N — pty quiescence / submit-confirm

windows (bump --confirm-sec for slower models, e.g. Opus).

• --host / --listen-port / --primary-port / --shadow-port — bind

host and the three ports (default 127.0.0.1, 8788 / 8790 / 8791). Point claude (the driver does) at --listen-port.

• --out DIR — output dir (default runs/paired-<timestamp>).
• --ack-2x — acknowledge the ~2x token cost (required to run).
• --no-plot — skip chart rendering.

Outputs

In the --out directory:

• A0.ndjson — SHADOW (baseline) turn-log, the authoritative paired record.
• <primary-arm>.ndjson (e.g. A5.ndjson) — PRIMARY (armed) paired record.
• instance.A0.ndjson / instance.<arm>.ndjson — each clawback instance's

OWN --turn-log, a cross-check (their numbers should agree with the tee's).

• proxy.primary.log / proxy.shadow.log / tee.log — process logs.
• state.primary.json / state.shadow.json — each instance's seeded state.
• report.md, report.csv, summary.json, manifest.json — analyzer

outputs; the turn-matched reclaim headline is at the top of report.md.

• charts/ — the four §9 charts (.skills/plot).

Read the headline with sed -n '1,40p' <out>/report.md.

Where this sits

.skills/ab runs arms serially as separate sessions (good for many arms and fragmentation, with each arm isolated by its own --state file + natural ANTHROPIC-KEY divergence). This skill runs two arms concurrently off one session for an exact turn-matched pair (good for a tight reclaim CI), at ~2x cost. Both call .skills/{drive,bench,plot} for the load, analysis, and charts. Capture a fixture once with .skills/capture if you also want a replay arm (this skill is pty-only — it drives real claude).