# Reduce

> Audit over-engineered codebases by factoring layers into live obligations, quotienting redundant distinctions, ablating unearned surface, and normalizing the survivors while preserving required behavior. Use when change latency or agent difficulty comes from frameworks, plugins, DI, codegen, task runners, config indirection, ORMs, GraphQL, monorepo or infrastructure tooling, web stacks, or requests to remove layers. In Actuating composition, classify enforcement sites before challenging candidates that preserve the same invalid region while adding or relocating compensating detection; never remove legitimate derived trust-boundary guards as residue.

- Skill: `tkersey/reduce` (Agent Skill, multi-file: 19 files)
- Install (CLI): `npx skillmds@latest add tkersey/reduce`
- Raw SKILL.md: https://api.skillmd.com/api/skills/tkersey/reduce/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: tkersey (https://skillmd.com/u/tkersey)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/tkersey/reduce

---


# Reduce

## Purpose

Act as the architecture **WINNOWING** reviewer. Find abstractions whose live
obligation is unproved, expired, moved, duplicated, invalid, or outweighed by
their change tax. Recommend a smaller normal form while preserving essential
truth.

Inside Actuating, return only a material challenge to the current construction;
no challenge needs no report. Actuating adjudicates in its active Working Set.
Reduce never selects architecture or grants mutation.

## Doctrine

```text
WINNOWING
  = FACTORING
  -> QUOTIENTING
  -> ABLATING
  -> NORMALIZING

guard:
  REFINEMENT-PRESERVING
```

- **Factoring** decomposes a layer into obligations, authorities, inputs,
  outputs, dependencies, observations, invalid states, escape paths,
  enforcement roles, and recomposition roles.
- **Quotienting** collapses distinctions no required observation can
  distinguish after congruence checks.
- **Ablating** removes, privatizes, collapses, or decommissions factors without a
  distinct live obligation.
- **Normalizing** recomposes the survivors around canonical semantic authority
  and lower primitives.
- **Refinement-preserving** retains required valid behavior while allowing
  obsolete, duplicated, invalid, or unrequired behavior to disappear.

## Abstraction elevator

`$reduce` descends. `$universalist` climbs. They share an altitude map without
sharing selection authority.

```text
descend   lower primitive preserves the live contract
climb     essential shape is missing; report the gap to Actuating
hold      a live obligation or proof weakness justifies the layer
split     remove incidental wrapper while preserving the essential invariant
quotient  collapse observationally indistinguishable factors
ablate    remove a discharged factor
normalize recompose around one semantic authority
```

In Actuating composition, `climb` reports an `essential-shape-gap`. It does not
call Universalist or recursively reopen architecture. Actuating decides whether
to request another nomination or block.

## Operating rules

1. Preserve required behavior unless current authority changes it.
2. Preserve invariants, protocols, authorization, data integrity, auditability,
   public contracts, and external obligations.
3. Use repository evidence first.
4. Treat absent evidence as uncertainty, not deletion authority.
5. Prefer reversible cuts and staged migration.
6. Do not add tools to remove tools unless total complexity falls materially.
7. Every removed factor needs obligation discharge.
8. Every target normal form needs recomposition proof.
9. Keep value and obligation risk separate.
10. Classify enforcement sites before calling them duplication or residue.
11. A candidate that preserves the same invalid region and merely adds, moves,
    renames, or duplicates compensating detection is dominated when an
    admissible owner- or cut-level candidate excludes the same family.
12. A derived trust-boundary guard is not dominated merely because primary
    enforcement exists elsewhere.
13. Do not turn the audit into another durable workflow unless independent
    durability is explicitly required outside the active Actuating run.

## Enforcement-site classification

Classify each implicated check or layer as exactly one primary role:

```text
primary-enforcement
  semantic authority for the law

derived-boundary-guard
  defense in depth or trust-boundary enforcement derived from that authority

compensating-guard
  downstream detection or repair required by permissive admission

compatibility-adapter
  temporary or externally required transition behavior

observability-guard
  monitoring, audit, or diagnosis without semantic admission authority

redundant-semantic-owner
  independently authored implementation of the same law

unknown
  current evidence cannot establish the role
```

Ablate or quotient only after the role and obligation are known. Prefer removing
compensating guards and redundant semantic owners after stronger enforcement is
proved. Preserve required derived, compatibility, and observability behavior.

## Workflow

For standalone audits. Inside Actuating use the composition below directly,
applying relevant rules without a duplicate audit table or verdict template.

1. Map the relevant layers, lower primitives, public/wire/storage boundaries,
   proof surfaces, invariants, invalid states, escape paths, and owner set.
2. Trace at least one real request, change, or command through each major
   abstraction.
3. Factor each candidate by live obligation:

| factor | obligation | authority | enforcement role | inputs/outputs | observations | invalidity excluded | external commitment | recomposition role |
|---|---|---|---|---|---|---|---|---|

4. Classify each obligation `live`, `moved`, `expired`, `duplicated`, `invalid`,
   or `unknown`.
5. Measure edit, lookup, tool, deploy, hidden-control, and proof tax against
   evidenced value.
6. Test quotient candidates against an explicit observation set and congruence
   under accepted operations.
7. For bug-driven work, compare predicate-defined invalid regions, sanctioned
   admission coverage, frontier or cut coverage, owner sets, escape paths,
   guard roles, residual checks, and evidence strength.
8. Reject compensating guard movement presented as correctness improvement when
   the same invalid family still crosses the same sanctioned admission paths.
9. Preserve derived trust-boundary, compatibility, and observability guards
   unless their distinct obligations are discharged.
10. Check whether apparently removable shape is an essential product,
    refinement, agreement boundary, free construction, protocol, or external
    obligation.
11. Classify dominance as `dominant`, `dominated`, `incomparable`, or `unknown`.
12. Return one operator-level verdict:
    `keep`, `hold`, `factor`, `quotient`, `wrap`, `split`, `collapse`, `ablate`,
    `privatize`, `decommission`, `normalize`, `replace`, `validate-first`, or
    `climb`.

## Actuating composition

Reuse Actuating's exact candidate, source evidence, and accepted obligations.
Return only a factor whose obligation is discharged, an independently maintained
truth that can become derived, or a concrete smaller adequate construction that
could change selection. Name the affected code and enforcement role, the proposed
delete/derive/replace operation, preserved obligations, and the decisive evidence
or falsifier. Report material incomparability, an essential-shape gap, or missing
proof rather than manufacturing a smaller winner. No material challenge means
silence, not a `minimal` report or an additional audit.

Merely moving detection while leaving the same invalid state, transition, or
escape admissible is not constructional improvement. Preserve derived trust-boundary,
compatibility, and observability guards unless their distinct obligations are
discharged. Unknown ownership is not deletion authority. Never remove an independent
oracle merely because it checks the same law as the producer.

This is supporting analysis, not a selection or durable artifact. Use
[reduction-certificate.md](references/reduction-certificate.md) only for a standalone
audit, explicitly requested certificate, or independently durable handoff outside
Actuating. Do not call Universalist or reopen architecture; return a material gap
to Actuating.

## Implementation mode

When explicitly asked to implement:

1. Hand the material construction challenge to Actuating.
2. Actuating updates its ephemeral Working Set and binds it to the
   exact current Git head.
3. Implement one coherent reduction seam.
4. Preserve the old surface until the selected proof relation passes unless
   direct deletion is already proved safe.
5. Run recomposition, invalid-region, admission-coverage, bypass, guard-role, and
   residue checks.
6. Stop when new evidence changes an obligation, guard role, family predicate,
   owner set, or target.

No Ledger operation, Construction artifact, hotspot register, or one-operation
protocol is required.

## Standalone output

For a bounded seam, report the affected abstraction, live obligation, proposed
operation, preservation evidence, and first safe change. Include material risks
and uncertainty. A broad audit may expand the layer map, factorization, candidate
comparison, migration, and rollback where they change a decision. Emit RC-v1 only
when independently required. Do not repeat the findings in a second summary form.

