skills/cw-ai-check/references/structural-overcompletion.md

# Structural Overcompletion

Use this reference to detect AI residue that survives a lexical cleanup. It applies most strongly to essays, criticism, reported analysis, memoir, thought leadership, and other prose in which authorship depends on judgment rather than information delivery alone.

## Contents

- Governing principle
- Source-aware scan: provenance, certainty, causality, structure, voice, metaphor, portability, reader posture, research, and endings
- Source-blind scan
- Repair order
- Boundaries with adjacent skills

## Governing Principle

Generated prose tends to complete the source into the most legible possible artifact. It supplies missing causes, regularizes examples, performs the requested voice, closes open loops, and converts uncertainty into a finished lesson.

Prefer a specific unfinished truth to an elegant false one.

## Source-Aware Scan

When notes, transcripts, research, or earlier human drafts are available, compare the prose against them before making stylistic judgments.

### 1. Unsupported Connective Tissue

Flag any detail that makes the narrative cohere but cannot be traced to the source:

- motives, private thoughts, emotions, memories, or body sensations;
- scenes, times, frequencies, sequences, or quoted language;
- causal explanations for observed behavior;
- redesigns, tests, outcomes, or changes the source does not record; and
- relationship dynamics or biographical context inferred from tone.

Plausibility is not provenance. Remove the detail, ask the writer, or mark it as an inference.

### 2. Epistemic Inflation

Compare the source's degree of certainty with the draft's:

- observation becomes intention;
- suspicion becomes diagnosis;
- possibility becomes mechanism;
- a useful lens becomes proof;
- a partial test becomes a general finding; or
- an unresolved experiment becomes a transformation.

Restore the original epistemic temperature. Keep precise uncertainty; cut only empty hedging.

### 3. Causal Closure

Flag explanations that make mixed evidence unnaturally clean:

- one cause explains every example;
- every anecdote supports the thesis;
- a taxonomy absorbs exceptions without residue;
- chronology is rewritten as a deliberate decision; or
- a concluding equation compresses a tradeoff the evidence has not settled.

Separate what happened from what the writer thinks it may mean. Let examples disagree when they genuinely disagree.

### 4. Framework Compliance

Flag structure that appears optimized independently of the material:

- sections have suspiciously equal length or identical internal logic;
- every example receives the same diagnosis, lesson, or decision;
- headings form a complete conceptual system the body did not need;
- a list has been padded or compressed to a satisfying number; or
- the draft chooses a governing framework before the live question is clear.

Do not add disorder for camouflage. Reorganize around the actual sequence of discovery, evidence, decision, or contradiction.

### 5. Performed Voice

Models can imitate a writer by oversupplying recognizable moves. Flag density, not mere presence:

- several jokes, metaphors, coinages, asides, fragments, or register collisions in one paragraph;
- a short release line at the end of most paragraphs;
- repeated signature syntax after the thought no longer calls for it;
- quirky specificity with no source; or
- memorable language whose primary job is to signal authorship.

Preserve sourced oddity. Remove simulated personality. A paragraph usually needs no conspicuous voice move; one is often enough.

### 6. Metaphor Systems

Flag a metaphor when it becomes organizing infrastructure without continuing analytical work:

- the image spreads into headings and transitions;
- callbacks exist mainly to make the ending feel designed;
- adjacent figurative and literal sentences make the same point; or
- the image remains memorable after its explanatory mechanism has disappeared.

Keep each return only when it produces a new inference. Otherwise end the metaphor and resume literal language.

### 7. Portable Profundity

Apply the portability test: could the line move into an unrelated essay, keynote, or social post with only the topic nouns changed?

Common forms include:

- polished causal equations;
- universal declarations after one personal example;
- detachable short diagnoses;
- inversions written to be quoted; and
- final lines that restate the paragraph at a higher register.

Reattach the sentence to the named person, artifact, choice, or consequence. Cut it when specificity destroys its force; the shape was doing the work.

### 8. Reader Assignment

Flag movement from personal evidence to claims about the reader:

- "you" states what the reader has done, felt, feared, or built;
- "we" converts one person's behavior into a collective condition; or
- a broad lesson appears without corroboration or a clearly offered invitation.

Stay with "I" while the evidence is personal. Offer recognition rather than assigning it.

### 9. Decorative Research

Flag research added to certify a conclusion rather than answer or complicate a live question:

- studies arrive after the interpretation is already settled;
- multiple sections repeat a study, caveat, takeaway rhythm;
- evidence supplies borrowed authority but changes no judgment;
- the study's scope is stretched to fit the personal example; or
- anonymous research language hides provenance.

Name the source, explain the relevant finding plainly, state the material limit, and return to what it changes. Cut research that changes nothing.

### 10. Over-Resolved Endings

Models tend to recap, universalize, complete the metaphor, and manufacture a portable final line. Flag an ending that:

- claims a decision or transformation the body did not demonstrate;
- resolves the writer's uncertainty for emotional satisfaction;
- repeats the thesis without adding a consequence;
- turns a personal experience into a universal lesson; or
- gives every recurring image a ceremonial callback.

End one step before the model wants to. Use a concrete changed decision, cost, implication, or specific open question.

### 11. Argument Self-Commentary

Also called rhetorical scaffolding. The prose stops making its argument and starts describing it. The model narrates the piece to the reader — rating its own evidence, announcing where the reader stands in the structure, labeling a claim's epistemic status, or telling the reader what a quote proves before the quote speaks.

**This is the most durable tell in otherwise clean copy.** It uses plain vocabulary, short sentences, and no item from the lexicon, so a lexical scan returns nothing. It usually appears at paragraph boundaries, because that is where a model reaches for a transition and writes commentary instead.

Four subtypes:

**Evidence self-rating.** The writer grades their own material.

- "That reversal is the cleanest thing we found, and the most expensive."
- "The good part is measurable."
- "The benchmarks are less charming."
- "The artifact that convinced me is the one I asked to be scannable."

**Structural signposting.** The sentence announces a position rather than advancing one.

- "The other half of the read is less flattering."
- "Then there's the map."
- "Verbosity is only the surface of it."
- "LFGBench sets the boundary on the optimistic version of this."

**Epistemic labeling.** The draft names a claim's status in place of making the claim.

- "The open question is whether the capability arrives in a form people want."
- "Two things stay provisional."
- "One result complicates the picture, and it deserves a flag."

**Quote characterization.** The draft interprets a quote before delivering it.

- "Mike put it best:"
- "The complaint was unanimous and differently worded."

Two tests, either sufficient to flag:

1. **Deletion test.** Cut the sentence. If the argument stands and the only loss is reader guidance, it was scaffolding.
2. **Referent test.** Flag nouns pointing at the artifact instead of the subject — *the read, the section, the picture, the case, the evidence, the boundary, the takeaway, the story.* A sentence whose grammatical subject is the piece itself is nearly always commentary on the piece.

**Fix:** Delete, or convert the commentary into the thing it was describing. Uncertainty becomes a fact about what limits it — "Forty hours isn't long enough to tell" rather than "the open question is." A turn becomes one word — "also," "either" — rather than a signpost sentence. A quote gets the person and the task, never an interpretation.

**Do not flag adjudication.** Deciding which of two findings controls a decision is a claim about the world and often required. "Mike's construction check could change the hotel choice. The HTML page could not" stays. "The construction check is the stronger of the two findings" goes. Publication handoff conventions also stay.

## Source-Blind Scan

When only the finished prose is available, do not claim invention or source inflation as fact. Flag observable symptoms and ask targeted questions:

- "Is this motive established in the source, or inferred?"
- "Did the writer reach this conclusion, or did the draft complete it?"
- "Do these examples actually share one cause?"
- "Does this callback add an inference?"
- "Is the reader claim supported beyond the writer's experience?"
- "Does this sentence make the argument, or describe it?"

The self-commentary scan needs no source at all. Run it on finished prose with the same confidence as a lexical pass — the deletion and referent tests are decidable from the copy alone.

## Repair Order

1. Remove or mark unsupported material.
2. Restore scope and certainty.
3. Reopen false causal closure.
4. Let the material determine the structure.
5. Strip argument self-commentary.
6. Reduce performed voice and decorative callbacks.
7. Repair research posture and reader overreach.
8. Run the lexical and sentence-pattern scan last.

Step 5 goes before step 6 for a reason: scaffolding often masquerades as a voice move, and cutting it first stops you defending a sentence that was never carrying voice.

## Boundaries

- Use `cw-ai-check` for generic model-generation patterns independent of publication or writer.
- Use `cw-voice-check` to compare prose with an active writer's documented voice.
- Use `cw-dev-edit` when the underlying human argument, evidence, stakes, or structure needs development rather than AI-residue removal.
- Use publication-specific guardrails for house standards that would apply even to wholly human prose.