RFD 1125 details: what already exists, and the three detectors that cannot fire

Every section below names its gate in an HTML comment under its heading. That mark is the mechanism this RFD adds, so this file is also its first example.

Why the rule was split

The blanket rule was not wrong when it was written. It was written for one kind of document, and the repository then grew a second kind without anyone noticing that the rule had stopped fitting.

ASD-STE100 exists so a maintenance technician reading in their second language can take a procedure and act on it correctly the first time. Every constraint it carries serves that: bounded sentences, one approved term per referent, no contractions, no ambiguity about what a pronoun points at. Applied to a schema table or an install step, those constraints are exactly right.

Applied to an argument they remove what makes it an argument. A decision persuades through how its reasons sit against each other, and a controlled vocabulary flattens that into a list of assertions. The reader loses the ability to tell a load-bearing claim from an incidental one.

The failure an essay actually has is one ASD-STE100 says nothing about. An essay fails by reading as though nobody wrote it. The tropes.fyi catalog names that failure, which is why a second gate buys something instead of duplicating the first.

The per-section mark follows from a fact about this repository rather than a preference. A DETAILS.md is where a measurement lands, and it is also where the reasoning that motivates the measurement lands. Those are two kinds of writing in one file, and any file-level rule has to be wrong about one of them.

What already exists

This RFD proposes less than it first appears. The tropes-removal-model project holds most of the parts:

part	state
`seeds/trope.parquet`	28 ASD-STE100 violation categories
`seeds/ai_trope.parquet`	20 AI-writing tropes, seeded from tropes.fyi
`runtime/prose_nodes.py`	CommonMark prose extraction, so a gate reads no syntax
`gate.py`	the ASD-STE100 gate; it runs the STE table only
a runner for the AI tropes	absent

scripts/seed_ai_tropes.py reached the same conclusion as this RFD before this RFD did. Its own docstring states that ASD-STE100 “gates the wrong things” in an essay. Two facts remain to settle, and this RFD settles them.

gate.py does not import prose_nodes, and nothing reads ai_trope.parquet. The catalog and the parser exist and are not connected.

runtime/prose_nodes.py imports markdown_it. pixi.toml does not declare it. The RFD gates pin markdown-it-py==3.0.0, so the two projects must agree on a version once both depend on it.

The measurement

scripts/measure_ai_tropes.py in tropes-removal-model reads seeds/ai_trope.parquet and runtime/prose_nodes.py. It owns no patterns and no parser, because a second copy of either would drift from the first.

It reads all 119 RFD README.md files. The counts are exact: the population is fixed and the script reads every member, which is what CLAUDE.md rule 5 asks for.

The table holds 20 tropes: 15 regex, 3 classifier, 2 manual. Three of the 15 find anything:

trope	documents	share	hits
Compulsive Counting	6	5%	7
Overused Adverb	2	2%	2
Em-Dash Addiction	2	2%	2

Three detectors cannot fire

Twelve detectors report zero. Nine of those zeros are real. Three are not.

--audit feeds each pattern its own recorded example_phrase, passes it through prose_nodes, and asserts the pattern still matches. A detector that cannot match its own example is BROKEN, and its zero on a corpus means nothing.

detector	fault
Bolded Lead Clause	matches `...`; `prose_nodes` removes the markers
Where/What/Why Header	matches `#{1,6}`; `prose_nodes` removes the hashes
Negative Parallelism	spells `it'?s`; its own example says “it is”

The first two are one fault twice. Both patterns were written against raw markdown, and the extractor hands over the text a reader sees. Measured on one file: Bolded Lead Clause finds 5 in the source and 0 after extraction.

Each repair is small. Move the two formatting rules to run before extraction, or rewrite them against the extracted text. Add it is to the third alternation.

A zero from a broken detector reads exactly like a clean corpus. That is the failure CLAUDE.md rule 3 names, and it is why --audit runs first.

Every hit is a false positive

Run --inspect to see each hit in context. The inspection changes the result, so the count alone is not the measurement.

trope	what the hits are
Compulsive Counting	exact counts of items the document then enumerates
Overused Adverb	“silently”, describing a check that skipped in silence
Em-Dash Addiction	two parenthetical dashes across 119 documents

seed_ai_tropes.py predicted this. Its docstring records that a regex is “a screen, not a verdict”, and that Compulsive Counting “fires on the phrase ‘three things’, which is sometimes just three things”. The corpus agrees: all seven hits count things the document goes on to list.

No hit in the corpus is a true positive. Against the mechanical subset, the corpus is clean, and every firing rule needs narrowing before it gates anything.

What no mechanical detector reaches

tropes.fyi lists 49 tropes. The table seeds 20, and records that limit per row in its coverage column rather than in a README.

Five of the 20 have no regex. Invented Concept Label, Premise Stacking and Enumeration As Prose need the classifier. Circular Fractal and Symmetric Paragraph Length have no automatic detector at all.

The ASD-STE100 half splits the same way: about 18 mechanical rules as regular expressions, about 8 semantic rules through a SetFit classifier, and 2 more through a whole-document pass in runtime/cross_sentence.py. An essay gate needs the same second half, and this RFD does not build it.

The conflict this RFD resolves

seed_ai_tropes.py already assigns work to each table, and it puts READMEs on the ASD-STE100 side:

trope.parquet      ASD-STE100        READMEs, procedures, reference, rules
ai_trope.parquet   AI-writing tells  essays, design notes, logbook entries

This RFD moves READMEs to the other column. The disagreement belongs in the open rather than under a silent overwrite.

That assignment is defensible on its own terms. A README in most repositories is reference material, and reference material belongs under a controlled language. It is wrong here because an RFD README is not that kind of README. RFD 1000 asks it to state a problem and a decision in forty lines, which makes it the most compressed argument in the repository rather than the most procedural document.

The measurement decides it rather than merely permitting it. Both tables were run over the same 119 READMEs, through the same extractor:

table	documents flagged	hits
`ai_trope.parquet`	10 of 119	11
`trope.parquet` patterns	119 of 119	4,397

Three orders of magnitude apart, on identical input. Three+ Nouns in a Row alone flags every document in the corpus. Contraction Used flags roughly two thirds of them, and Semicolon Used about a quarter.

Neither number is a verdict, and the trope side is not flattered by the comparison: its 11 hits are all false positives, as the section above records. These are weak-label screens from scripts/seed_labels.py rather than the shipped classifier, which applies thresholds these patterns do not. Read them as direction and magnitude, not as counts of real defects.

The direction is the whole argument. A gate that fails 119 documents out of 119 is not holding a standard, it is producing noise, and people stop running it. prose_nodes.py was written to avoid the same outcome, which arrives here through the choice of table rather than through the choice of parser.

How to re-run

cd .tropes_removal_model
python3 scripts/measure_ai_tropes.py --audit    # check each detector can fire
python3 scripts/measure_ai_tropes.py            # the AI-trope table
python3 scripts/measure_ai_tropes.py --inspect  # every hit, in context
python3 scripts/measure_ai_tropes.py --compare  # both tables, same extractor

--audit exits non-zero while any detector is broken. Run it before you trust a zero from the other modes.

The default path is ../.request_for_discussion, which holds when both projects sit at the workspace root. default.xml places them there.