Add prose style contract for the writing pass

Adapt the flossing project's anti-slop checklist to facap: keep the
general layers (lexical bans, em-dash budget, density floor, Buschek
integrity mapping) and replace the project-specific honesty rails with
facap's (retired 0.977, no-fit definition, estimator scope, hard-k as
null model only, inline assumptions, MNIST MAE-first, teacher gap as
scoping evidence). Note 41 section 7 and the schedule now point at it.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

1 files changed, 73 insertions, 0 deletions

diff --git a/paper/style_contract.md b/paper/style_contract.md
new file mode 100644
index 0000000..5662e56
--- /dev/null
+++ b/paper/style_contract.md
@@ -0,0 +1,73 @@
+# Style contract (operative checklist for every section pass)
+
+Adapted from the flossing project's contract. Sources: Shaib et al. 2026 slop
+taxonomy (density/relevance/coherence are the load-bearing dimensions);
+Buschek's four reviewer complaints; 2026 banned-pattern lists. Applied as a
+per-section pass, not a vibe. The number/claim discipline this file enforces
+is anchored in `notes/36_evidence_ledger.md` (§9 booby traps) and
+`notes/41_paper_plan.md` (§3 structure, §7 rules).
+
+## Hard bans (lexical)
+delve, tapestry, landscape (figurative), testament, pivotal, showcase, intricate, vibrant,
+seamless, elegant, dramatically, novel (self-describing), comprehensive (self-describing),
+notably, crucially, interestingly, importantly, "It is worth noting", "Moreover," as sentence
+opener, "In recent years", "has garnered attention", "paradigm shift".
+Technical terms that overlap ban lists (e.g., "soft" in "soft ramp", "robust" conditioning)
+stay.
+
+## Punctuation & rhythm
+- Em dash: ≤2 per section, never two in one sentence, never as paired parenthetical.
+- No exclamation marks. Semicolons fine.
+- Every paragraph contains at least one sentence under ~12 words.
+- No uniform paragraph shapes: don't end three consecutive paragraphs with a summary clause.
+
+## Structure
+- The contrast template ("X, not Y" / "not X, but Y") has a total budget of ONE device:
+  the title ("What Does Random Feedback Cost?") resolves to one echo at the end of paragraph 1
+  ("not free; irreducible but soft"). Zero elsewhere. Express other contrasts by stating the
+  positive finding and letting the numbers carry the negation.
+- Rule-of-three closers only when the three items are an exhaustive measured list
+  (the three contributions qualify; nothing else does).
+- No chronological narration of the investigation. No "we first tried hard-k, then...".
+  Structure follows note 41 §3.
+- Results topic sentences contain a finding or a number, never a plan ("We then examine...").
+
+## Integrity (Buschek's four, mapped to our risks)
+- Marketing language: zero unevidenced evaluative adjectives. "essentially exact" only with
+  the MAE or correlation in the same sentence.
+- Performative related work: every citation paired with the precise claim it supports; check
+  each against `notes/*.md` and the source paper before the citation lands (note 41 §6 lists
+  the per-cite rationale).
+- Misrepresentation: prior-work characterizations quote verbatim with section; the
+  Song-Xu-Lafferty complementarity is stated as `E_B[H_FA]=0` in the residual direction, not
+  as a correction of their convergence result.
+- Stretched summaries: no interpretive metaphors for results. One flourish per section maximum,
+  and it must be cuttable without losing content.
+
+## Density (the actual anti-slop weapon)
+- Every sentence must add a number, a definition, a caveat, or a logical step. Otherwise cut.
+- Modal verbs (could/might/may) confined to Limitations and future work.
+- Hedges appear only where the evidence concedes; one hedge per concession, not a seesaw.
+
+## Honesty rails (project-specific)
+- Closed-form ramp correlation: cite 0.933 (matched init) and 0.982 (5-init ensemble). The
+  retired 0.977 never appears. State which is which in the same sentence.
+- "No fit" is defined once at first use: no post-hoc scalar tuned to the reported data;
+  predictions come from initialization or two early kernel snapshots. Never drift from it.
+- The estimator is "snapshot-conditional, no fitted parameters" at first use, never
+  "architecture-only"; its lazy-regime scope (the MNIST eta-sensitivity) is stated where the
+  estimator is introduced.
+- Hard-k is a null model and a reported failure (2x-250x overprediction), never a gap
+  predictor. `E[e]=k/P` (null model) and `rho=1-E_B[e_0]` (measured share) are distinct
+  objects; never conflate them.
+- Every theorem states its assumptions inline: feedback independence and zero mean, the
+  output-layer-exact fact, A1 (lazy), A2 (scalar erosion, exact only in the residual direction).
+- "Soft ramp / no phase transition" refutes one named hypothesis (redundancy-exhaustion
+  threshold); it is the analyticity of `gap_T` in the spectrum, stated as such, not a universal
+  claim about FA.
+- MNIST estimator: report MAE first (0.00196, 3.9% relative); the correlation 0.911 appears
+  only with the gap range (0.040-0.063) that makes it the weaker statistic.
+- The teacher test-gap sign flip is scoping evidence for the optimization-gap claim, not a
+  contribution.
+- No promises (code release, future experiments) the authors have not decided. Code+config
+  release is planned (manifest note 40); state it only in the reproducibility section.