Language Guardrails
Document ID: SCBE-LANG-2026-02-18 Scope: All project documentation, patent text, README, marketing copy Status: ACTIVE – Applies to all docs in this repo
Purpose
Prevent hype, unsupported claims, and promotional language from entering patent filings, technical documentation, or public-facing materials. Every statement must be traceable to evidence or clearly marked as speculative.
The Three Registers
1. Patent Register (PATENT_DETAILED_DESCRIPTION.md, claims text)
Rules:
- Every claim must map to a row in CLAIMS_EVIDENCE_LEDGER.md
- Use “configured to” and “comprising” per USPTO convention
- State what the system DOES, not what it COULD do
- Quantitative claims require experiment ID and AUC/metric
- No superlatives (“best”, “unbreakable”, “revolutionary”)
- No future tense (“will achieve”, “is expected to”)
Allowed: “The system achieves 0.9942 AUC across six attack types (Experiment 4).” Forbidden: “The system is virtually unbreakable.” / “This will revolutionize AI safety.”
2. Technical Register (CLAUDE.md, SPEC.md, architecture docs, code comments)
Rules:
- State what code does, not what it aspires to
- Distinguish CODE_EXISTS_UNTESTED from PROVEN
- Reference specific files and line numbers when describing behavior
- Use present tense for implemented features, conditional for planned
- No marketing language in technical docs
Allowed: “Layer 12 computes H(d,R) = R^(d^2). Production uses the bounded variant 1/(1+d+2*pd).” Forbidden: “Our groundbreaking harmonic wall provides military-grade security.”
3. Experimental Register (CLAIMS_AUDIT_V4.md, experiment logs)
Rules:
- State pass criterion BEFORE running experiment
- Report exact numbers (AUC to 4 decimal places)
- Report failures and disproven claims with equal prominence
- No cherry-picking: report all conditions, not just favorable ones
- Include trial count, seed, and pipeline version
Allowed: “GeoSeal v2: AUC = 0.543. Swarm dynamics destroy discriminative signal.” Forbidden: “GeoSeal shows promising early results.” (when AUC = 0.543)
Forbidden Phrases
These phrases are banned from all registers:
| Phrase | Why | Replace With |
|---|---|---|
| “unbreakable” | No system is; patent-toxic | “achieves N-bit security at distance d_crit = X” |
| “revolutionary” | Promotional; no technical meaning | [describe what it does] |
| “military-grade” | Meaningless marketing term | “NIST Level 3 (ML-KEM-768 + ML-DSA-65)” |
| “quantum-proof” | Nothing is proven quantum-proof | “post-quantum: based on MLWE + MSIS (NIST FIPS 203/204)” |
| “unhackable” | Invites contradiction | “100% detection rate across N tested attack types” |
| “AI-powered” (as differentiator) | Everything is; adds no information | [describe the specific mechanism] |
| “patent-pending” (in technical docs) | Legal status, not technical property | Move to README/index only |
| “will be” / “is expected to” | Future tense = unproven | “is” (if proven) or “requires validation” (if not) |
| “state-of-the-art” | Requires comparative benchmark you may not have | “achieves AUC = X on benchmark Y” or omit |
| “novel” (without citation) | Must cite what it’s novel relative to | “novel relative to [prior art]: [specific distinction]” |
| “impossible to circumvent” | Mathematically false for any finite system | “cost scales as R^(d^2), requiring d_crit = 9.42 for 128-bit security” |
| “zero-day proof” | Not a real security property | [describe specific attack classes tested] |
Status Label Rules
When describing a feature’s readiness, use exactly one of these labels:
| Label | Meaning | Allowed Claims |
|---|---|---|
| PROVEN | Experiment exists with AUC >= 0.95 or 100% gate accuracy | “The system detects X with AUC = Y” |
| PROVEN_PARTIAL | Proven under restricted conditions | “Under synthetic conditions, the system achieves…” |
| CODE_EXISTS_UNTESTED | Code runs, no comparative experiment | “The implementation computes X. Comparative validation pending.” |
| THEORETICAL_ONLY | Math is correct, no code | “The mathematical framework defines X. Implementation pending.” |
| DISPROVEN | Experiment contradicts claim | “Experiment N shows X does not hold: [data]” |
| REFRAMED | Original claim disproven, new framing works | “Originally claimed as X; reframed as Y based on [data]” |
Rule: The label must match the row in CLAIMS_EVIDENCE_LEDGER.md. If you describe a feature and there’s no ledger row, add one first.
Citation Format
When referencing experimental evidence:
(Exp N: AUC = X.XXXX, N trials, pipeline = [synthetic|real])
When referencing prior art distinction:
Novel relative to [Author, Year]: [1-sentence distinction]
When referencing code:
See [filename]:[line_range] or [module_name]
Review Checklist
Before merging any documentation change:
- No forbidden phrases present
- Every quantitative claim has an experiment ID
- STATUS labels match CLAIMS_EVIDENCE_LEDGER.md
- No future-tense promises for unvalidated features
- Disproven claims are reported, not hidden
- Patent text uses USPTO register conventions
- “Novel” claims cite what they’re novel relative to
These guardrails apply to all human and AI-generated text in this repository. Violations should be flagged in PR review.