# inference-arbitrage

Audits a Claude Code plugin — its skill/agent definitions and its real usage
transcripts — to find steps that are being done by raw LLM inference but pass
every test of a deterministic script, and files the well-evidenced ones as
issues on the target plugin's own repo. Runs on demand; each run persists a
snapshot, so cost, candidate status, and this plugin's own calibration
judgment all accumulate over time instead of resetting on every audit.

The name echoes `builder-arbitrage`: route each unit of work to the cheapest
executor that can do it correctly — here, "script vs. model" instead of
"which build node."

## Use it

```text
/offload-audit <plugin name or path> [--days N]
```

or ask directly: *"audit `cluster` for token optimization"*, *"is this plugin
wasting inference?"*. Read-only with respect to whatever it audits — it never
edits the plugin it's looking at and never implements its own recommendations.

## What it actually does

1. **Static pass** (`bin/plugin-inventory`) — verb ratios, duplicated command
   blocks, rule tables, script coverage, per plugin definition.
2. **Dynamic pass** (`bin/offload-scan`) — real cost attribution from this
   environment's transcripts: Mechanical Turn Ratio, recurring tool-call
   n-grams, read amplification, retry density, judgment density, fan-out.
3. **Boundary classification** — every candidate gets graded against the
   rubric below (`design/rubric.md`) and gets a `boundary_confidence`.
4. **Two hard gates**, enforced mechanically, never by agent judgment:
   - **FR-4.2** — no falsifiable "case where a human would overrule the
     script" → downgraded to a boundary question, never filed.
   - **FR-4.5** — the evidence-strength label (`measured` / `thin` /
     `unmeasured`) or the filing-threshold side must be *stable* across
     measurement windows, checked against this plugin's own snapshot
     history. A `measured 12%` that becomes `thin 0%` on a different date
     range is a defect in the auditor, not a fact about the target, and it
     gets self-reported on *this* repo, not the target's.
5. **Filing** — a `high`-confidence candidate above the 2% audited-spend
   threshold is filed as a tracked issue on the **target's own repo**,
   idempotent via a marker comment so a re-run comments instead of
   duplicating. Everything below the bar is reported as a boundary question,
   never filed.
6. **Run summary** — a durable page on this plugin's own wiki, plus (FR-9)
   drawers and knowledge-graph facts in mempalace: one set scoped to the
   audited plugin's own wing (so a human working on that plugin later finds
   the findings while searching their own project's memory), one set in this
   plugin's own wing accumulating cross-target calibration judgment. Recalled
   memory is advisory input only — it can never override the two gates above.

## The rubric, in one table

Every candidate is exactly one of four positions:

| Position | What it means | Example |
| --- | --- | --- |
| 1 — `pure-script` | Deterministic end to end | parsing, diffing, formatting a report |
| 2 — `script-with-compiled-judgment` | Judgment made once at authoring time, frozen into a rule | a hunt's require/exclude rules |
| 3 — `llm-over-script-digest` | Script gathers/filters/ranks; model judges only the shortlist — **where most real value is** | `worktree-audit`'s provably-safe classification, model handles only the must-reconcile pile |
| 4 — `pure-inference` | Irreducible — say so explicitly | intent, naming, novel synthesis |

The gate to reach position 1 or 3 is five determinism tests (closed input
domain, a stable output oracle, invariance, no semantic gap, bounded branch
factor) — all five, no averaging — plus a **falsifiability triple**: a typed
function signature, three input→output pairs, and the case where a human
would overrule the script. Missing the third element is the hard gate above.
Full writeup: [`design/rubric.md`](design/rubric.md).

## Worked examples, from real audits

- **`oleks/claude-plugin-cluster#56`** — `cluster:gitea-agent`'s sequential
  issue-read batching, `llm-over-script-digest`, high confidence, measured at
  7.26% of audited spend over 887 invocations.
- **`oleks/claude-plugin-anxious#29`** — `pr-rot`'s script-verb-dominant body
  (verb ratio 1.000, zero backing script), filed on the static case alone per
  FR-2.4 since it had zero invocations in the audited window (the plugin
  disables its own model invocation).
- **`kotkan/claude-plugin-inference-arbitrage#15`** — a real catch by the
  FR-4.5 gate: `anxious/agent-wip/release-policy-derivation` measured
  `thin` at 0.00% of spend on one window and `measured` at 11.85% on
  another. Correctness case was complete (a duplicate hand-synced
  implementation already exists in `bin/wip`), but the window-dependent
  value claim meant it was refused rather than filed — self-reported here,
  not on `anxious`'s repo, because the instability is this plugin's own
  defect.

## Design docs

- [`design/spec.md`](design/spec.md) — scope, functional requirements (the
  numbered `FR-*` this README references)
- [`design/plan.md`](design/plan.md) — architecture and file layout
- [`design/rubric.md`](design/rubric.md) — the full script-vs-inference
  boundary rubric
- [`design/tasks.md`](design/tasks.md) — build order and phase history

Tracked as issues on this repo, milestone `v0.1.0`.