Eval-first expansion: 435 deterministic cases, 23 gates, teach/mimic voice system, measured model tiering

.gitignore

@@ -24,9 +24,14 @@ test_*.txt
 # Wiki sync state
 scripts/.wiki_sync_state.json
 
+# Taught voices: profiles, cards, and provenance are user data, not repo content
+.unslop/
+
 # Packaged skill
 *.skill
 
 # Behavioral eval run artifacts (prepared tasks, model outputs, judge results)
 evals/runs/
 runs/
+.claude/
+.codex/

.unslop/contrib/let-that-sink-in/manifest.json

@@ -0,0 +1,10 @@
+{
+  "category": "emphasis_crutch",
+  "date": "2026-07-06",
+  "pattern_name": "let-that-sink-in",
+  "post_redaction_sha256": "6e70f687ebfa4157f2173c00e589e37444bc1f272a4295e6919a991b3f56c7c4",
+  "pre_redaction_sha256": "6e70f687ebfa4157f2173c00e589e37444bc1f272a4295e6919a991b3f56c7c4",
+  "redactions": [],
+  "tell": "Let that sink in",
+  "word_count": 19
+}

.unslop/contrib/let-that-sink-in/report.md

@@ -0,0 +1,42 @@
+# Add emphasis_crutch pattern: Let that sink in
+
+## The specimen
+
+> The launch memo ends with a warning instead of evidence. Let that sink in before you ask what changed.
+
+- Source genre: TODO: source genre
+- Date: 2026-07-06
+- Redaction note: none
+
+## Why it's an AI-ism
+
+TODO: explain why this phrase is a reusable AI-writing tell in 2-4 sentences.
+
+## Detection
+
+- Pattern added: TODO: regex or phrase for `Let that sink in`
+- Severity: TODO: hard or soft
+- Gating rationale: TODO: explain literal-use boundary
+- Catalog entry location: references/taboo-phrases.md
+
+## Evals
+
+| row id | kind | what it pins |
+|---|---|---|
+| CONTRIB-FN-let-that-sink-in | FN | exact specimen flags `Let that sink in` |
+| CONTRIB-FP-let-that-sink-in | FP | literal-use protection for `Let that sink in` |
+
+The FN stdin is the unmodified specimen after approved redaction.
+
+## Gate results
+
+red-first: already green; proposed pattern appears active
+
+## Checklist
+
+- [ ] eval-first (row was red before the pattern)
+- [ ] literal-use FP row included
+- [ ] REC row if an existing word was gated
+- [ ] catalog + scanner parity green
+- [ ] coverage gate green (pattern exercised)
+- [ ] snippet publication approved by the user

.unslop/contrib/let-that-sink-in/row_fn.json

@@ -0,0 +1,24 @@
+{
+  "assertions": [
+    {
+      "gte": 1,
+      "path": "total_violations",
+      "type": "json"
+    },
+    {
+      "type": "violation_phrase_contains",
+      "value": "let that sink in"
+    }
+  ],
+  "category": "scanner_false_negative",
+  "command": [
+    "python3",
+    "scripts/banned_phrase_scan.py"
+  ],
+  "correct_behavior": "The scanner reports at least one violation containing the contributed tell.",
+  "failure_mode": "The scanner misses the contributed tell from the exact specimen.",
+  "id": "CONTRIB-FN-let-that-sink-in",
+  "stdin": "The launch memo ends with a warning instead of evidence. Let that sink in before you ask what changed.\n",
+  "target": "script",
+  "title": "New AI-ism fixture: Let that sink in"
+}

.unslop/contrib/let-that-sink-in/row_fp_TEMPLATE.json

@@ -0,0 +1,21 @@
+{
+  "assertions": [
+    {
+      "equals": 0,
+      "path": "total_violations",
+      "type": "json"
+    }
+  ],
+  "category": "scanner_false_positive",
+  "command": [
+    "python3",
+    "scripts/banned_phrase_scan.py"
+  ],
+  "correct_behavior": "No violations for the protected literal/domain use.",
+  "failure_mode": "A broad pattern could flag legitimate prose.",
+  "id": "CONTRIB-FP-let-that-sink-in",
+  "protects": "emphasis_crutch",
+  "stdin": "TODO: add a literal or domain-specific use that should remain clean.",
+  "target": "script",
+  "title": "Literal-use protection for Let that sink in"
+}

.unslop/contrib/let-that-sink-in/snippet.txt

@@ -0,0 +1 @@
+The launch memo ends with a warning instead of evidence. Let that sink in before you ask what changed.

AGENTS.md

@@ -15,10 +15,16 @@ start with evals, then update the skill or scripts until the suite passes.
      proving the jargon use still flags.
    - Agent behavior change: add a `skill` row only when the behavior is rewrite,
      preserve, decline, or route.
+   - Context-heavy or literal-sense pattern: prefer an `evals/fixtures/pairs/`
+     minimal pair plus PAIR rows, then run `python3 evals/check_pairs.py`.
 3. If a `skill` row is added, update `evals/build_shared_benchmark.py` with its split
    and domain, then regenerate `evals/shared-benchmark.json`.
 4. Update `scripts/banned_phrase_scan.py`, `SKILL.md`, `presets/`, or `references/`
    only after the eval captures the desired product behavior.
+5. `python3 evals/check_pattern_coverage.py` (gate DOC-09) fails unless every
+   scanner pattern is exercised by a row and every category has a `protects` FP
+   row; `python3 evals/kata_add_pattern.py --run` (DOC-10) rehearses this. No
+   grandfathering.
 
 Prefer contextual patterns over broad word bans. A row for `wedge` or `load-bearing`
 should also protect literal uses such as construction, mechanics, law, medicine, code,
@@ -43,6 +49,8 @@ skill-benchmark validate evals/shared-benchmark.json --strict-leakage
 If you touched `SKILL.md`, `presets/`, or `references/`, also run
 `evals/run_behavioral.sh tune`.
 
+If you touched the co-writer, mimic, or detector-pack model features, run `python3 evals/run_model_parity.py --dry-run --responses evals/fixtures/parity/canned_responses.json` and, before merge, the live matrix across the GPT and Anthropic spectrums (see `references/pipeline.md`, Model Parity).
+
 `evals/evals.json` is the legacy happy-path suite; never edit it. Use
 `evals/adversarial-evals.json` as the source of truth.
 

CLAUDE.md

@@ -15,10 +15,16 @@ start with evals, then update the skill or scripts until the suite passes.
      proving the jargon use still flags.
    - Agent behavior change: add a `skill` row only when the behavior is rewrite,
      preserve, decline, or route.
+   - Context-heavy or literal-sense pattern: prefer an `evals/fixtures/pairs/`
+     minimal pair plus PAIR rows, then run `python3 evals/check_pairs.py`.
 3. If a `skill` row is added, update `evals/build_shared_benchmark.py` with its split
    and domain, then regenerate `evals/shared-benchmark.json`.
 4. Update `scripts/banned_phrase_scan.py`, `SKILL.md`, `presets/`, or `references/`
    only after the eval captures the desired product behavior.
+5. `python3 evals/check_pattern_coverage.py` (gate DOC-09) fails unless every
+   scanner pattern is exercised by a row and every category has a `protects` FP
+   row; `python3 evals/kata_add_pattern.py --run` (DOC-10) rehearses this. No
+   grandfathering.
 
 Prefer contextual patterns over broad word bans. A row for `wedge` or `load-bearing`
 should also protect literal uses such as construction, mechanics, law, medicine, code,
@@ -43,6 +49,8 @@ skill-benchmark validate evals/shared-benchmark.json --strict-leakage
 If you touched `SKILL.md`, `presets/`, or `references/`, also run
 `evals/run_behavioral.sh tune`.
 
+If you touched the co-writer, mimic, or detector-pack model features, run `python3 evals/run_model_parity.py --dry-run --responses evals/fixtures/parity/canned_responses.json` and, before merge, the live matrix across the GPT and Anthropic spectrums (see `references/pipeline.md`, Model Parity).
+
 `evals/evals.json` is the legacy happy-path suite; never edit it. Use
 `evals/adversarial-evals.json` as the source of truth.