feat(optimize): add --subtract, the subtraction axis, behind a deterministic floor

Every command so far asked an addition question — what to add, what to move,
what it costs. Nothing asked what is no longer earning its always-loaded rent.
This adds that axis as a fourth lensCheck on the existing hybrid motor rather
than a new scanner or a 22nd command: the measured payoff (~18% of one file)
justifies a mode, not machinery.

It is the only lens that proposes REMOVING config, so it carries a guarantee
the others don't need: a load-bearing block is never a candidate. Precision is
asymmetric — a missed dead line costs a few tokens per turn, a deleted one
costs a wrong remote or a broken script — so the floor is decided in code
(lib/floor-exclusion.mjs) before the opus judge sees anything, never in prose.

Granularity is the leaf block, with two structural exceptions: a paragraph
ending in ':' merges with the list it introduces, and an ordered list is a
contract whose steps inherit floor from any sibling. Unordered lists
deliberately do not inherit — a load-bearing bullet and a disposable one
routinely share a list, and container-reasoning is the error the hand-built
ground truth exists to catch.

Verified against that ground truth (built before any classifier existed), with
the comparison machine-checked rather than read by eye: zero load-bearing
blocks proposed, 11/18 deletable groups surfaced, ~756 tok ~ 18% of a ~4300
token file — inside the pre-registered band. The first run found five floor
violations the synthesized fixture missed; each got a structural rule and a
fixture shape so it cannot regress.

Three real bugs the dogfood run exposed, all now covered:
- JS \b is ASCII-only, so /\bunngå\b/ never matches — every Norwegian keyword
  ending in æ/ø/å was silently dead.
- A bare word/word is not a path; "pros/cons" vetoed the largest deletable
  block until PATH_RE was tightened to rooted paths and globs.
- "Mid-sentence" must key on a preceding lowercase letter; the loose version
  read **bold labels:** and quoted openers as entities, costing 4 of 11 groups.

BP-SUB-001 is grounded entirely in the Anthropic steering blog already cited by
BP-MECH-001..004 and asserts nothing from the talk that motivated the feature —
no "80%", no ablation figure.

Suite 1365 -> 1382/0. Frozen v5.0.0 snapshots untouched; plain optimize output
byte-identical on identical input (--subtract adds keys only when passed).
knowledge-refresh-cli's reference date moved to 2026-08-01: its premise that
every seed entry was verified 2026-06-20 expired when BP-SUB-001 got a genuine
verification date, and backdating the entry to fit the test would have been a
lie about when its source was checked.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RW2haJXbxZpKivKHseSXNh
This commit is contained in:
Kjell Tore Guttormsen 2026-07-31 16:24:51 +02:00
commit e9921d3c9d
13 changed files with 1080 additions and 32 deletions

View file

@ -279,6 +279,7 @@ Your team configuration changes over time. Track it:
| `/config-audit manifest` | Ranked table of every token source (CLAUDE.md, rules, agents, skills, output styles, MCP, hooks) sorted by estimated tokens — each tagged with its **load pattern** (always-loaded / on-demand / external) plus an **always-loaded subtotal** ("≈X tokens enter context every turn before you type"). Component-level: no coarse plugin roll-up (it would double-count) |
| `/config-audit feature-gap` | Context-aware feature recommendations grouped by impact |
| `/config-audit optimize` | Optimization lens (mechanism-fit): config that works but fits a better mechanism — procedure→skill, lifecycle→hook, unscoped path→rule, "never"→permission. Hybrid motor (deterministic pre-filter + opus precision gate), every finding cites a best-practices-register rule |
| `/config-audit optimize --subtract` | **Subtraction lens** — the inverse question no other command asks: what no longer earns its always-loaded rent? Ranks CLAUDE.md blocks that correct general model *behaviour* rather than stating a local fact, split into **dead** (never missed) and **earned** (returns if the model stumbles), with the token payoff (`BP-SUB-001`). **Load-bearing local facts are excluded deterministically before the judge sees anything** — remotes, versions, paths, filenames, policy invariants and unresolvable entity names are never candidates, and an ordered list is treated as a contract. Opt-in, proposes only, never writes. Pair with `--global` to reach the user-level CLAUDE.md, where the always-loaded cost actually sits |
| `/config-audit fix` | Auto-fix deterministic issues with backup + verification |
| `/config-audit rollback` | Restore configuration from a previous backup |
| `/config-audit plan` | Generate prioritized action plan from audit findings |
@ -595,6 +596,31 @@ date, and a `confidence`. It is the source of truth for the optimization lens (O
`scanners/lib/best-practices-register.mjs` (zero-dependency, native JSON). See
`docs/v5.7-optimization-lens-plan.md`.
### The subtraction floor
`optimize --subtract` is the only lens that proposes *removing* configuration, so it carries a
guarantee the others do not need: **a load-bearing block is never a candidate.** Precision here
is asymmetric — a missed dead line costs a few tokens per turn, while a deleted load-bearing
line costs a wrong remote or a broken script — so the floor is decided in code
(`scanners/lib/floor-exclusion.mjs`), before the opus judge sees anything, rather than being
left to prose judgement.
A block is floored when it carries an underivable local literal (inline code span, rooted path,
domain, concrete filename, version pin), when it states a policy invariant (secrets,
credentials, production, prompt-injection boundaries — floor *by decision*, not by
classification), or when it names a capitalized entity the mechanism cannot resolve without a
dictionary. That last rule is a deliberate conservative default: it declines to decide and
keeps the block, paying in recall rather than risk.
Granularity is the **leaf block** — one list item including its wrapped continuation lines, or
one paragraph — with two structural exceptions: a paragraph ending in `:` merges with the list
it introduces, and an *ordered* list is treated as a contract whose steps inherit floor from
any sibling. Unordered lists deliberately do not inherit, so a load-bearing bullet and a
disposable one can coexist in the same list. Measured against a hand-built ground truth
(35 blocks, 13 genuinely ambiguous): zero load-bearing blocks proposed, 11 of 18 deletable
groups surfaced, ≈18 % of an always-loaded file. On a well-maintained config this axis is
mostly a no-op — which is itself the finding.
---
## Testing