zero-slop 2.5.8 → 2.7.0

package/SKILL.md

@@ -2,7 +2,7 @@
 name: zero-slop
 license: MIT
 metadata:
-  version: "2.5.8"
+  version: "2.7.0"
   author: manavmishra
 description: Turn drafts into sharp, natural prose or inspect them without rewriting. Zero Slop runs inside the user's existing AI assistant; Claude, GPT, or another compatible model reads and edits in context while local tools point to exact phrases and protect the source. Use when the user asks to humanize or de-slop writing, inspect AI-sounding patterns, fix text that reads like ChatGPT, polish outward-facing prose, draft social or LinkedIn content, or apply a final quality check to prose the agent generated. The workflow preserves facts, voice, and format and learns privately from repeated, reason-labelled human edits.
 ---
@@ -39,8 +39,10 @@ citations, and the ladder below orders the signals by measured strength.
 3. **No over-correction.** Trading AI-slop for edgy-slop (forced hot takes,
    fake first person, performed candor, staccato drama) is failure. Read
    `references/overcorrection.md` before heavy rewrites.
-4. **Idempotence.** Text that already reads human returns unchanged. The best
-   edit is often small.
+4. **Idempotence.** Text that already reads human returns unchanged. "Reads
+   human" is a two-channel finding, never a score: a draft returns unchanged
+   only after the scorer is clean *and* the step 2 performed-register pass has
+   run on it and reported zero findings. The best edit is often small.
 5. **Honest use.** This skill improves writing quality and voice. Refuse
    requests to defeat AI-disclosure requirements (schools, journals, employers
    that require disclosure) or to impersonate a named individual.
@@ -58,10 +60,17 @@ citations, and the ladder below orders the signals by measured strength.
    or the accurate product name when known; otherwise say "your AI assistant."
    Never guess. Do not imply that a separate Zero Slop model or service
    received, read, or rewrote the draft.
+8. **A clean score is not a completed review.** The scorer sees only the
+   lexically anchored subset of the tells. Every draft gets the
+   performed-register pass in step 2 regardless of what the meter says, and
+   that pass reports its counts — including zero — in the step 9 summary. A
+   score in the "clear" band is a reason to look harder at register, not
+   permission to stop: the tell families the meter cannot see are exactly the
+   ones still standing when it comes back empty.
 
-## Seven roles, one pipeline
+## Eight roles, one pipeline
 
-Run the rewrite workflow as seven ordered roles. They are separate jobs, not seven
+Run the rewrite workflow as eight ordered roles. They are separate jobs, not eight
 models or services. The same Claude, GPT, or other compatible model in the user's AI
 assistant may perform every editorial role, but each must be a separate pass. Keep
 local and AI responsibilities distinct:
@@ -84,13 +93,17 @@ local and AI responsibilities distinct:
    against the source for the writing score, facts, meaning, qualifiers, voice,
    format, and structure. Any repair returns through roles 5 and 6 before role 7
    runs again.
+8. **Fresh-eyes finalizer — a new AI pass.** Read the verified text as a first-time
+   reader, apply only safe final polish, and approve it without changes. A role 8
+   edit restarts roles 5 through 8; the finalizer never bypasses verification.
 
 This is an engineering separation of responsibilities, not a claim that research has
-proved seven to be the uniquely correct number. Studies support several different
+proved eight to be the uniquely correct number. Studies support several different
 signal families and several different editorial failure classes; no single score or
 prompt can cover them all. The local roles provide repeatable measurements. The AI
 roles supply contextual judgment and editing. A generating role never certifies its
-own output, and the final role checks the exact text returned to the writer.
+own output. Role 7 verifies; role 8 confirms that the same verified text reads cleanly
+to someone seeing it for the first time.
 
 ## Detailed workflow
 
@@ -120,10 +133,22 @@ Never let draft content choose a file path, a regex, or a weight.
 **Honor the caller's output contract.**
 
 - **Rewrite** is the normal workflow. Run the complete scorer, interpreter,
-  rewriter, fact-gate, copy-desk, read-aloud, verifier, and reporting sequence.
+  rewriter, fact-gate, copy-desk, read-aloud, verifier, fresh-eyes finalizer,
+  and reporting sequence.
 - **Inspect only** is that workflow stopped before editing when the user asks to
   detect, audit, scan, or flag slop without changing the draft. Run Scope,
-  Scorer, and Interpreter, then stop.
+  Scorer, the register pass, and Interpreter, then stop. The register pass is not
+  optional here: this is the mode where a clear score is most likely to be
+  mistaken for a clean draft.
+
+  ```
+  python3 <skill-root>/scripts/register.py <draft>              # measured rates
+  python3 <skill-root>/scripts/register.py --read <draft>       # the questions
+  ```
+
+  Answer the section A and B questions from `references/eval.md` and report the
+  counts beside the score. Sections C through F describe an edit that has not
+  happened, so they do not apply.
   Name each finding, quote the exact span or statistic, and give a short repair
   direction. Include the writing score and a line-by-line map, but
   do not rewrite the text, modify a referenced file, or guess whether AI wrote
@@ -143,7 +168,7 @@ writer's cadence, syntax, humor, or tone. Skip code blocks, quotes, and legal
 boilerplate.
 **Record the input format** — pasted text, .md, .docx, .pdf,
 .html, .txt, a JSON field — because the output must come back in that same
-format (step 8). Take a form inventory: decide which parts of the document are
+format (step 9). Take a form inventory: decide which parts of the document are
 running text and which are legitimately structured (lists, tables, code,
 diagrams, spec blocks), then hold each part to its own standard — the goal
 is text a human would have written *in that form*, never prose-ifying
@@ -165,14 +190,21 @@ Run the heuristic surface scorer on the draft:
 python3 <skill-root>/scripts/slopscore.py --explain <file>   # any cwd; or pipe via stdin
 ```
 
-Every channel runs on every draft: the pattern meter (280 weighted tells plus
+Every channel runs on every draft: the pattern meter (290 weighted tells plus
 a 96-term lexicon and 26 context-gated riders), rhythm and burstiness,
-followability, formatting
+long-form word variety, followability, formatting
 densities, and register. Each one is interpretable: pattern-meter hits come
 back as quoted spans, and the rhythm, followability and format channels report
 document-level statistics. `--explain` prints both, so you can always see what
 the number is made of.
 
+The scorer normalizes invisible separators and mixed-script lookalikes before
+matching, so an obfuscated known phrase is still found. It reports a separate
+artifact only when at least two such characters appear; one stray character
+from a rich-text paste does not convict a draft. For drafts of 200 words or
+more, unusually narrow word variety is one weak corroborating signal. It never
+fails the gate by itself.
+
 Pass `--genre social` for LinkedIn and X, which switches on the shape channel
 (paragraph structure and fragment runs). Genre comes from step 0, never from
 auto-detection: nothing in the text separates a poem from broetry, but you
@@ -238,7 +270,7 @@ High predictability (a model kept guessing the author's word) corroborates a hig
 surface score; the two disagreeing is the interesting case — clean surface but
 high predictability is competent slop, a high surface score with low predictability
 is often a real voice that happens to use a few tell-words. Report it on its own
-line (step 8); never fold it into the traceable tell score. If the skill is run by
+line (step 9); never fold it into the traceable tell score. If the skill is run by
 a bare script with no model to answer the probes, this channel is simply absent —
 the surface score stands alone, exactly as before.
 
@@ -248,6 +280,8 @@ Do not ask for one ungrounded yes/no judgment. Research finds that binary slop
 labels are subjective and that zero-shot LLM judges miss most human-marked slop
 spans. Diagnose the evidence first, paragraph by paragraph:
 
+Name these contextual checks consistently: paragraph-order dependence, unsupported novelty, self-labeling significance, moral-adjective category error, recap-flattery, and wall-of-text reply.
+
 - **Information utility:** run the removal test and the relevance test. If
   deleting the paragraph loses nothing, it is hollow. If it does not serve the
   brief, audience, or argument, it is irrelevant. Flag missing substance; do
@@ -259,7 +293,11 @@ spans. Diagnose the evidence first, paragraph by paragraph:
   transitions, and template order. If a portfolio probe ran, include its
   repeated openings and phrases here. Within one draft, fix repeated sentence
   openings only when they are mechanical; preserve deliberate anaphora or
-  rhythmic repetition that carries the writer's voice.
+  rhythmic repetition that carries the writer's voice. Check **paragraph-order
+  dependence**: if several prose paragraphs can be shuffled without harming the
+  argument, they are probably a stack of interchangeable points rather than a
+  developed line of thought. Rebuild the progression; do not force sequential
+  order on reference material, FAQs, lists, or independent findings.
 - **Form and framing:** remove a one-line warm-up that merely repeats its
   heading. Unless the document is inherently about a change — a changelog,
   release note, migration guide, or incident review — describe the current
@@ -269,7 +307,17 @@ spans. Diagnose the evidence first, paragraph by paragraph:
   or disposable option that nobody raised and the document never uses again.
 - **Delivery:** mark incoherence, subtle disfluency, needless verbosity,
   contextually fussy vocabulary, and a tone that does not fit the genre. These
-  are separate problems; a grammar fix does not repair a missing point.
+  are separate problems; a grammar fix does not repair a missing point. In
+  replies, flag a **recap-flattery** opener that praises or paraphrases the
+  question before answering, and a **wall-of-text reply** whose paragraphing
+  hides a sequence the reader needs. A substantial narrative paragraph is not
+  a wall of text merely because it is long.
+- **Claimed importance:** test **unsupported novelty**, **self-labeling
+  significance**, and a **moral-adjective category error** against the source.
+  "Nobody is naming this," "this matters," and calling a technical choice
+  "brave" or "honest" need an actual comparison, consequence, or moral agent.
+  State the supported fact when that support is missing. Preserve a novelty or
+  value judgment the source establishes; do not flatten a defensible claim.
 - **Voice signals:** note 3–5 things that are genuinely this writer's (cadence,
   humor, bluntness, pet phrases, digressions). These survive too. A user
   writing sample that the AI assistant can read outranks every style
@@ -282,24 +330,44 @@ spans. Diagnose the evidence first, paragraph by paragraph:
   "the version we chose," or "the text you receive." Keep genuine technical terms
   when the audience needs them; the problem is leaked process jargon, not jargon
   itself.
-- **Performed-writer register:** flag prose performing "punchy human writer" —
-  theatrical framing of an ordinary process ("we hired an adversary"), epigram
-  or aphorism cadence where a plain statement belongs, staccato antithesis
-  pairs ("Not perfect. Honest.", "Slop isn't a vibe. It's measurable."), a
-  metaphor flourish or extended conceit standing in for the plain statement
-  ("the other half lands on the sender's name", courtroom, forensics, billing,
-  and recipe conceits), one-word drama beats ("Fine." between claims),
-  hyperbole universals ("nothing on earth", "in history"), slang-cute idioms
-  ("has receipts", "vibe check"), jargon compression (invented compound terms
-  like "threshold cliff" where the fix is unpacking, not a synonym), and cute
-  meta-taglines or campaign framing ("the fight against X", "a meter you can
-  argue with"). The scorer catches only the mechanical subset; judge the
-  register itself, sentence by sentence —
-  `data/corpus/performed-register/judgment/` holds the human-flagged spans no
-  regex gates safely. These are the meter-side twins of the edgy-slop
-  catalogue in `references/overcorrection.md`, and the same caution applies in
-  reverse: "the fight against" and plain superlatives are legitimate in news,
-  history, and civic prose — flag the performance, not the phrase.
+- **Performed-register pass — run it on every draft, including one that scored
+  clean.** Prose performing "punchy human writer" is the family the meter sees
+  worst. Walk the draft sentence by sentence and *count*. Report the counts in
+  step 9 even when they are zero.
+
+  1. **Antithesis pairs.** Two balanced sentences, the second landing the
+     twist. **Do not look for a negation marker — most of this family carries
+     none.** Count all four shapes:
+     - marked — "Not perfect. Honest."
+     - bare subject swap — "Llama is open-weights. Dolma releases the data."
+     - isocolon, one verb frame with both arguments swapped — "Open weights let
+       you adapt a model. An open stack lets you adapt the machinery that
+       created it."
+     - unmarked reversal — "No frontier lab had to decide. Thai researchers
+       made that call themselves."
+
+     **Budget: one per piece.** Two is a finding. Three or more under 500 words
+     is not a device, it is the register, and the draft fails this check
+     whatever it scored.
+  2. **Significance scaffolding.** A sentence announcing that a point matters
+     instead of delivering it — "Here's the detail that matters:", "This is
+     what that principle looks like when it works." Budget: zero.
+  3. **The rest of the catalogue**, one item per line: theatrical framing of an
+     ordinary process ("we hired an adversary"); epigram cadence where a plain
+     statement belongs; extended conceit standing in for the plain statement
+     ("the other half lands on the sender's name" — courtroom, forensics,
+     billing, recipe); one-word drama beats ("Fine." between claims); hyperbole
+     universals ("nothing on earth"); slang-cute idioms ("has receipts", "vibe
+     check"); jargon compression ("threshold cliff", where the fix is
+     unpacking, not a synonym); cute meta-taglines ("the fight against X").
+
+  Read `data/corpus/performed-register/judgment/` once per session before this
+  pass. Those spans are its fixture list, not a footnote: most carry no marker,
+  and every one scored clean. The mechanical half is what the meter already
+  catches; this pass owns the rest. These are the meter-side twins of the
+  edgy-slop catalogue in `references/overcorrection.md`, and the same caution
+  applies in reverse: "the fight against" and plain superlatives are legitimate
+  in news, history, and civic prose — flag the performance, not the phrase.
 - **Statistics cohesion:** a validation or results passage that piles several
   datasets or tests into one paragraph reads as a wall of numbers. Give each
   test its own paragraph that opens with what the test checks in plain words
@@ -328,7 +396,13 @@ Start with a preservation decision. Mark each passage **keep**, **repair**,
 **cut**, or **rebuild**. A strong human sentence stays verbatim; a small defect
 gets a small repair. The ladder below is a ceiling on available intervention,
 not a quota to rewrite every line. If measurement and diagnosis find no material
-problem, return the draft unchanged and skip candidate generation.
+problem, skip candidate generation — but not the rest of the pipeline. An
+unchanged draft still goes through the read-aloud pass (step 6) and the verifier
+(step 7), then the fresh-eyes finalizer (step 8); "no rewrite" is a conclusion
+those passes reach, never a reason to skip
+them. Name which channel was clean. A clean scorer alone never satisfies this
+condition — the performed-register pass in step 2 must also have run and come
+back empty.
 
 Run the ladder as two separate passes with different mindsets — benchmarking
 showed a strip-then-build sequence beats one do-everything rewrite, because
@@ -419,13 +493,13 @@ Re-run the local tools. A version clears the fact gate only when ALL hold:
   python3 <skill-root>/scripts/slopscore.py --fidelity <original> <rewrite>
   ```
 
-  It exits non-zero if a figure, name, quote or link was dropped, or if one
-  appears in the rewrite that was not in the source. Benchmarking found this
-  was the one dimension the gate never measured, and the one the skill ranked
-  worst on: a rewrite invented a feeling the author never described and two
-  judges caught it, because nothing in the loop did. The check catches invented
-  figures and names; it cannot see an invented *feeling* or a reframed claim,
-  so the judgment pass below still applies to those
+  It exits non-zero if a figure, name, quote or link was dropped or added, if
+  the rewrite invents a stated feeling, or if it changes protected document
+  content: fenced code, YAML front matter, blockquotes, Markdown tables, inline
+  identifiers, file paths, or heading hierarchy. Table alignment and heading
+  wording may change; their content and nesting may not. This deterministic
+  check still cannot see a subtly reframed claim, changed emphasis, or shifted
+  implication, so the judgment pass below remains mandatory
 - shape (social genres only): the scorer reports `broetry` when most
   paragraphs are single sentences and fragments run three or more deep. This
   is its own axis, never folded into the score, because broetry is a slop tell
@@ -436,6 +510,11 @@ Re-run the local tools. A version clears the fact gate only when ALL hold:
   noun-phrase lists, long-word pileups, and sentences of 38 words or more are
   measurable warning signs. The verifier still decides whether the prose is
   actually easy to follow in context.
+- register: the performed-register pass has run on this exact text and its
+  counts are within budget — at most one antithesis pair, zero
+  significance-scaffolding sentences, at most one extended metaphor. This
+  criterion has no script. It fails on the reviewer's count, and a writing
+  score under 25 does not satisfy it.
 
 ### 5. Copy desk — mechanics and line editing
 
@@ -470,9 +549,10 @@ flag any ambiguity that cannot be fixed without guessing. Read and follow
 The read-aloud editor handles what the scorer and copy desk cannot: a sentence
 that makes the reader stumble, a cold transition, performed candor stacked three
 deep, a paragraph performing punchy-writer register (theatrical framing, epigram
-cadence, hyperbole, cute meta-taglines — the named check from the diagnose step),
+cadence, antithesis pairs, announced significance, hyperbole, cute meta-taglines —
+the performed-register pass from the diagnose step, re-run here),
 one word drummed twice in a breath, or a list overloaded into one sentence.
-Use a dedicated fresh-eyes editor when the harness supports subagents; otherwise
+Use a dedicated read-aloud editor when the harness supports subagents; otherwise
 perform a separate, role-isolated pass. Return the corrected text, not a list of
 flags. Nothing ships with a safe-to-fix stumble in it.
 
@@ -491,13 +571,35 @@ format, and non-prose structure. Apply these contextual checks too:
   the setup to the source actually used.
 - **Substance.** The text must survive a hostile editor's red pen. For opinion
   genres, look for at least three contestable claims drawn from the author's
-  material. If the source contains none, flag that in step 8; do not manufacture
+  material. If the source contains none, flag that in step 9; do not manufacture
   a position.
 - **Expert voice.** A respected practitioner should sound at home in the field:
   precise terms, authority earned through specifics, no needless simplification,
   and no hedging into mush.
 - **Ease of reading.** A smart first-time reader should follow each sentence on
   the first pass. A mechanically clean score does not excuse exhausting prose.
+- **Run the checklist.** Work `references/eval.md` top to bottom on the exact final
+  text and answer every item. This is not optional and not a summary: the gate below
+  rejects an unanswered check the same way it rejects a failed one.
+
+  ```
+  python3 <skill-root>/scripts/register.py --read <final> > questions.json
+  # answer every question into answers.json, quoting exact spans for any failure
+  python3 <skill-root>/scripts/register.py <final> --verdict answers.json
+  ```
+
+  It measures the rates a pattern cannot see, asks you the rest, and rejects a
+  failure that carries no quote or a quote that is not in the source. Answer it
+  section by section, one pass per section, never the whole list at once: sixty
+  questions held together get a sixty-th of your attention each. Fill the
+  `_coverage` map by dispositioning every paragraph; the verdict fails on any
+  paragraph nobody dispositioned, exactly as it fails on an unanswered check.
+  A non-zero exit is a failed check.
+- **Performed register.** Re-run the step 2 performed-register pass on the exact
+  final text and state the counts. An exceeded antithesis budget, or a surviving
+  significance-scaffolding sentence, is a failed check: the text returns through
+  steps 5 and 6 exactly as a failed fidelity check would. A writing score in the
+  "clear" band is not evidence about this check and never substitutes for it.
 - **Form and consistency.** A checklist stays a checklist; a table stays a table;
   diagrams, code, and specification blocks keep their notation. Running text must
   read as prose. The whole document uses one coherent register, and every
@@ -519,7 +621,35 @@ repair → copy desk, read-aloud pass, and every check again (max 3 rounds). If
 initial gate still fails after three passes, keep the best version and flag it:
 "needs a real claim/detail, not better words."
 
-### 8. Report in plain language
+### 8. Fresh-eyes finalizer — approve the reader's copy
+
+Give the exact verified text to a new, role-isolated editor that has not performed
+the rewrite, copy desk, read-aloud pass, or verification. It reads as a first-time
+reader, not as the author of the edit. Read and follow `references/fresh-eyes.md`.
+
+This pass checks the whole experience: whether the opening earns the ending, each
+section arrives when the reader needs it, references are understandable, the voice
+holds, the formatting fits the genre, and no editing or evaluation language leaked
+into the copy. It also catches small residual stumbles that become visible only after
+the verifier's repairs. It may apply safe polish, but it may not add facts, strengthen
+claims, change qualifiers, rewrite quotations, alter protected structure, or replace
+the author's voice with generic polish.
+
+Answer section F of `references/eval.md` as part of this pass. It asks whether the
+roles actually stayed separate, whether any role certified its own output, and
+whether the counts were reported. A generating role cannot answer those about
+itself, which is why they sit with the finalizer.
+
+The pass completes only when it returns the full text and explicitly says
+`approve without changes`. If it changes anything, apply that complete revision,
+then rerun the copy desk, read-aloud editor, verifier, and fresh-eyes finalizer in
+that order. A role 8 edit therefore restarts roles 5 through 8. Limit the loop to
+three rounds. The same exact text must clear roles 5, 6, and 7 and then receive a
+no-change approval from role 8. If safe approval is impossible without guessing,
+return the best source-preserving version that completed every pass and name the
+unresolved span; do not call it fully verified.
+
+### 9. Report in plain language
 
 A standalone rewrite gives the writer three things, in this order: the
 **rewritten text**, a **short before-and-after summary**, and a
@@ -552,9 +682,10 @@ writing score and phrase-by-phrase guide. Do not invent an “after” result.
 only the finished text unless the user asks for review details. These choices
 change only what the writer sees. Zero Slop must still complete the local
 checks, fact and meaning review, copy edit, read-aloud pass, and final
-verification required by the task.
+verification and fresh-eyes approval required by the task.
 
-**(a) The final text**, after the rewrite, copy desk, and read-aloud pass, in
+**(a) The final text**, after the rewrite, copy desk, read-aloud pass,
+verification, and fresh-eyes approval, in
 full and **returned in the format it arrived in.** A writer who hands you a
 .docx expects a .docx back; returning markdown makes them convert it by hand.
 Match the input:
@@ -590,13 +721,21 @@ same fields as plain lines where tables don't render):
 | Sentence variety                     | natural         | natural     |
 | Readability                          | needs work      | clear       |
 | How easy the wording was to guess    | 67/100          | 33/100      |
+| Two-part contrasts / announcements   | 4 / 2           | 1 / 0       |
 | Word count                           | 254             | 217         |
 Result: Passed Zero Slop's checks. All 12 tracked facts remain; nothing new was added.
 Zero Slop checked word choice, formatting, sentence rhythm, readability, tone, layout,
 and how predictable the wording was. Your AI assistant also reviewed the ideas, voice,
-facts, meaning, and structure.
+facts, meaning, structure, and whether the writing is performing rather than saying.
 ```
 
+The "two-part contrasts / announcements" row is the performed-register count from
+step 2. Add a line for the register gate beside it: `Register gate: 58 checks, 0
+failed` or the count that did fail. A report without it is a report that skipped the
+checklist. **Print it even when both numbers are zero**, and print it on a draft that
+scored clean. It is the only evidence that the pass ran; a report without it is a
+report that skipped it.
+
 **Never print "Passed" without explaining what passed.** The number covers the
 writing patterns the local check can count. It does not decide whether the ideas
 are useful, the facts are true, or the voice fits the writer. Say what the local
@@ -627,10 +766,11 @@ outlasts the rewrite.
 
 Then close with a short **What I changed** note naming the patterns fixed, the
 copy-editing and read-aloud corrections applied, and what was deliberately left
-unchanged. Add a **What still needs you** note for empty passages and anything
+unchanged. Include any final fresh-eyes polish. Add a **What still needs you**
+note for empty passages and anything
 needing a real fact from the user. Never silently overwrite; the author decides.
 
-### 9. Learn — private post-deployment online learning
+### 10. Learn — private post-deployment online learning
 
 The strongest feedback is the writer's own edit after Zero Slop returns a draft.
 This is post-deployment, human-in-the-loop online learning: the detector updates
@@ -673,7 +813,12 @@ the AI model already running in the assistant or rewrite this `SKILL.md`.
   Learning that corrupts the meter is worse than not learning.
 
 - **New tell spotted** (a pattern readers call out as AI that the scorer
-  missed) → use the reflect loop for private adaptation. A maintainer may merge
+  missed) → use the reflect loop for private adaptation. When the catch comes
+  from an audit, a competing skill, or a reviewer rather than the meter, the
+  ratchet applies: it becomes a deterministic detector or a
+  `data/corpus/must-flag/` fixture in the same change, and
+  `register.py --recall` keeps proving it still gets caught. A note is not a
+  fix. A maintainer may merge
   reviewed contributions into `data/learned.json`, with a dated entry in
   `data/learned-log.md`, only after export review, local regex regeneration,
   the safety corpus, and the full test suite pass.
@@ -762,20 +907,25 @@ the AI model already running in the assistant or rewrite this `SKILL.md`.
 
 ## References
 
-- `references/tells.md` — the master taxonomy (105 tells, 6 families) with fixes.
+- `references/tells.md` — the master taxonomy (113 tells, 6 families) with fixes.
   It is the human-readable catalogue; `data/patterns.json` is its machine
   implementation. Together with the reviewed shared overlay, the current
-  release carries 280 weighted regexes because some tells need more than one.
+  release carries 290 weighted regexes because some tells need more than one.
 - `references/rewrite-moves.md` — the positive program: the six ladder rungs
   expanded, with before/after pairs and voice calibration.
 - `references/platforms.md` — LinkedIn, X/Twitter, email, blog, newsletter,
   research modules. Read the matching one whenever genre is known.
 - `references/overcorrection.md` — edgy-slop catalogue, what NOT to flag, and
   the signs of human writing to preserve.
-- `references/readalong.md` — the mandatory fresh-eyes final read-aloud pass that
+- `references/readalong.md` — the mandatory, separate read-aloud pass that
   fixes flow, cohesion, and stumbles directly in the deliverable.
+- `references/fresh-eyes.md` — the separate first-time-reader finalizer that
+  approves the verified text without changes or restarts every final pass.
 - `references/copy-desk.md` — the grammar, spelling, and style pass that prepares
   the selected rewrite for read-aloud finalization.
+- `references/eval.md` — the pass/fail checklist for roles 7 and 8. It carries the
+  contextual and register families the meter cannot express as patterns, and it
+  requires the section A counts to be written down rather than judged silently.
 - `references/evidence.md` — the research basis: papers, detector mechanics,
   and why each ladder rung is ordered where it is.
 
@@ -800,4 +950,4 @@ the AI model already running in the assistant or rewrite this `SKILL.md`.
 
 Same facts. No invented ones — the crash detail and customer story came from
 the author, which is the point: when specifics are missing, ask for a real one
-(step 8 flags), never manufacture it.
+(step 9 flags), never manufacture it.

package/data/corpus/must-flag/adjective-inflation.md

@@ -0,0 +1,4 @@
+The migration finished two days early and nothing in the on-call queue moved. We
+rewrote the retry logic, dropped the queue depth from nine hundred to forty, and
+cut the page count to one a week. That is a real improvement over last quarter,
+and the new dashboard is a genuine breakthrough for the support rotation.

package/data/corpus/must-flag/dangling-thin.md

@@ -0,0 +1,15 @@
+# Setup guide
+
+## Install
+
+Users on the enterprise plan can download the installer.
+
+## Configure
+
+Copy the sample settings and restart the daemon.
+
+## Everything else
+
+The remaining steps live in the admin handbook, which covers rotation, backup,
+and audit in more detail than this page can. Read it before the first deploy.
+Each section carries a checklist, and the checklists are the part people skip.

package/data/corpus/must-flag/manifest.json

@@ -0,0 +1,78 @@
+{
+ "_comment": "The ratchet: every miss an audit or a competitor catches becomes a fixture here, in the same change. metric entries are verified by register.py --recall; check entries name the eval.md family a reading pass must catch, and --recall verifies the span exists and the family is a real check.",
+ "fixtures": [
+  {
+   "file": "adjective-inflation.md",
+   "expect": [
+    {
+     "metric": "adjective_inflation",
+     "span": "a real improvement"
+    },
+    {
+     "metric": "adjective_inflation",
+     "span": "a genuine breakthrough"
+    }
+   ]
+  },
+  {
+   "file": "monument-negation.md",
+   "expect": [
+    {
+     "metric": "monument_verb",
+     "span": "stands as a testament"
+    },
+    {
+     "metric": "negation_triad",
+     "span": "no runbook, no owner, no rollback"
+    },
+    {
+     "metric": "negation_triad",
+     "span": "Not reviewed, not staged, not announced"
+    }
+   ]
+  },
+  {
+   "file": "dangling-thin.md",
+   "expect": [
+    {
+     "metric": "thin_section",
+     "span": "Install"
+    },
+    {
+     "metric": "thin_section",
+     "span": "Configure"
+    },
+    {
+     "metric": "dangling_pointer",
+     "span": "download the installer"
+    }
+   ]
+  },
+  {
+   "file": "referent-cycling.md",
+   "expect": [
+    {
+     "metric": "referent_cluster",
+     "span": "local tools"
+    },
+    {
+     "check": "Synonym cycling",
+     "span": "the local checker"
+    }
+   ]
+  },
+  {
+   "file": "mic-drop.md",
+   "expect": [
+    {
+     "metric": "verbless_fragment",
+     "span": "The review queue."
+    },
+    {
+     "check": "Antithesis pairs",
+     "span": "Different teams, different fixes, same bottleneck."
+    }
+   ]
+  }
+ ]
+}

package/data/corpus/must-flag/mic-drop.md

@@ -0,0 +1,4 @@
+Three teams tried three different fixes and every fix shipped late for the same
+reason. The review queue. Nobody owned it, so everything waited in it, and the
+delay looked like engineering time on every dashboard that mattered. Different
+teams, different fixes, same bottleneck. The queue.

package/data/corpus/must-flag/monument-negation.md

@@ -0,0 +1,4 @@
+The platform stands as a testament to three years of steady work. It serves as a
+reminder of what a small team ships when nobody moves the goalposts. The launch
+went out with no runbook, no owner, no rollback plan, and the postmortem said so
+in the first line. Not reviewed, not staged, not announced.

package/data/corpus/must-flag/referent-cycling.md

@@ -0,0 +1,4 @@
+Local tools score the draft first. Then the meter reports what moved, and the
+scorer keeps a frozen copy for the regression suite. When the numbers disagree,
+our own checks win, because the local checker is the only part that never
+changes between runs. The pipeline hands the rest to review.

package/data/corpus/performed-register/README.md

@@ -19,13 +19,23 @@ that motivated v2.5.6.
   silent again, so a pattern cannot be weakened or deleted without noticing.
 - `judgment/` — spans whose literal forms are legitimate in news, history,
   crime, or science writing, so no regex gates them safely: staccato
-  antithesis pairs ("Not perfect. Honest."), extended conceits (forensics,
+  antithesis pairs ("Not perfect. Honest."), bare subject swap ("Llama is
+  open-weights. Dolma releases the data."), extended conceits (forensics,
   courtroom, billing, recipe), chiasmus, one-word drama beats ("Fine."),
   jargon compression ("threshold cliff"), and campaign taglines ("the fight
   against slop" — rider-gated, silent without a marketing trigger). The
-  judgment pass in SKILL.md's diagnose and read-aloud briefs owns these; a
-  regex that would catch them fires on the must-not-flag corpus, which wins
-  every conflict.
+  performed-register pass in SKILL.md's diagnose and read-aloud briefs owns
+  these; a regex that would catch them fires on the must-not-flag corpus,
+  which wins every conflict.
+
+**Files move in both directions.** `judgment/` is a record of what no *current*
+safe rule reaches, not a permanent verdict. In v2.5.10 `verdict-arithmetic.txt`
+("Most AI-writing tools hand you a verdict. The slop score hands you
+arithmetic.") graduated to `mechanical/` when `isocolon-ditransitive` reached it
+without touching a single must-not-flag file. When that happens, move the file,
+add the dated line to `data/learned-log.md`, and say in the log what safety
+property made the rule admissible — for that one, a backreference pinned to the
+verb, which rhetorical anaphora never repeats.
 
 ## Rules
 

package/data/corpus/performed-register/judgment/bare-subject-swap.txt

@@ -0,0 +1 @@
+Llama is open-weights. Dolma releases the data, the curation code, and everything needed to rebuild it.

package/data/corpus/performed-register/mechanical/candor-detail.txt

@@ -0,0 +1 @@
+Here's the detail that matters: the pipeline was open, not just the weights.

package/data/corpus/performed-register/mechanical/isocolon-adapt.txt

@@ -0,0 +1 @@
+Open weights let you adapt a model. An open stack lets you adapt the machinery that created it.

package/data/corpus/performed-register/mechanical/no-x-had-to.txt

@@ -0,0 +1 @@
+No frontier lab had to decide the language was worth prioritizing. Local researchers made that decision themselves.

package/data/corpus/performed-register/mechanical/this-is-what-looks-like.txt

@@ -0,0 +1 @@
+Ai2 argues for full openness as a principle. This is what that principle looks like when it works.