references__voice-safety-gates

Synced from references/voice-safety-gates.md in script-pool skill on 2026-05-18. Edit upstream in the skill; this file is overwritten on next sync.

Voice-Safety Gates Spec (10 gates + Gate 0 cognitive-depth meta-check)

Run in modes/draft-script.md Step 5 after draft composition, before showing to Daniel. Audit trail written into voice_safety frontmatter block of script.md.

Canon sources

Voice-safety reads from TWO canonical locations:

Source	Owner	Used by
/Users/mikroverse/Projects/afriend-business/content-extraction/_brand/voice-preservation.md	A Friend brand	Gates 1-7 (brand-level voice rules: verbatim, two-beat, promise-close, ban-list, tool-naming-Pillar-7, anti-brain-rot, credibility-skin-in-game)
/Users/mikroverse/Projects/onari-intelligence/_canon/voice-deseltrus/	Daniel-personal cognitive depth + style	Gate 0 + Gates 9-10 (cognitive-depth, AI-style ANTI-PATTERNS L1 BLOCKING, ANTI-PATTERNS L2 WARNING)

The compose-deseltrus skill (also reads from _canon/voice-deseltrus/) is the canon-owner; afriend-script-pool consumes the same canon for cross-skill voice consistency.

Execution split

• Gate 0 (meta — Cognitive Depth) is LLM-judgment. Pre-flight check before numbered gates. Read VOICE.md Level 0 spec. If draft fails Level 0 (no multi-layer thinking, oversimplified, glossing complexity), skill regenerates BEFORE running numbered gates.
• Deterministic gates (1, 2, 4, 5, 9) run via scripts/voice-safety-check-deterministic.mjs. Pure JS regex + string-match. Fast. Reproducible.
• LLM-judgment gates (3, 6, 7, 8, 10) run as a structured-prompt step inside Mode C. The skill prompts itself to evaluate the draft against the gate’s source-of-truth and emit JSON. Slower but handles semantic checks regex can’t.

Gate 0 — Cognitive Depth (meta-check, pre-flight, added 2026-05-06)

Type: LLM-judgment. Runs BEFORE numbered gates. If fails, draft regenerated with explicit Level-0 guidance — numbered gates not even attempted on a Level-0-failed draft.

Rule: Voice canon VOICE.md Level 0 specifies four cognitive-depth markers (structural, ABOVE all stylistic markers):

1. Multi-Layer Thinking: practical observation → metakognitive insight → visionary implication, all in same stream
2. Willingness to Sit in Difficulty: if the input thought is hard to articulate, the draft must also be hard. Not flattened to easy understandability
3. Beyond-clip thesis (where applicable): use the cite as evidence for Daniel’s-deeper-claim, not as standalone summary
4. Register-preservation: practical / meta-cognitive / visionary registers all preserved (NOT all flattened to one register)

LLM prompt:

Read the script body. Verify against /Users/mikroverse/Projects/onari-intelligence/_canon/voice-deseltrus/VOICE.md Level 0 markers (Multi-Layer Thinking, Willingness to Sit in Difficulty, Bridge-not-substitution, Register-preservation). Return JSON: {“multi_layer”: bool, “sits_in_difficulty”: bool, “registers_preserved”: bool, “evidence”: "", “pass”: (all true), “if_fail_advice”: ""}.

Failure surface: “Gate 0 (cognitive depth) failed: . Draft is too flat / too clean / single-layer. Regenerating with Level-0 prompt augmentation.”

Behavior on failure: unlike other gates, Gate 0 failure auto-triggers regeneration (not Daniel-confirm). The skill silently re-prompts itself with explicit Level-0 guidance and re-runs Gate 0. Max 2 retries before surfacing to Daniel for manual override.

Gate 1 — Banned vocab scan

Type: Deterministic regex.

Rule: No instance of these words in Daniel-voice spans (verbatim-quote spans are exempt):

• revolutionize / revolutionary / revolutionizing
• seamless / seamlessly
• AI-powered
• game-changer / game-changing
• leverage / leveraging / leveraged (verb form, not noun “leverage” in finance sense)
• disruptive / disrupt (in marketing-speak sense)
• ! (exclamation marks anywhere — inline punctuation OR end-of-sentence)

Implementation:

const banned = /\b(revolutioniz(e|ing|ed|er|ation)|seamless(ly)?|AI-powered|game-?changer?|game-changing|leverag(e|ing|ed)|disrupt(ive|ing|ed)?)\b/gi;
const exclamations = /!/g;
// Run only on Daniel-voice spans. Verbatim-quote spans (marked by italic *"..."* with citation context) are skipped.

Source-of-truth: _brand/voice-preservation.md §5.

Failure surface: “Gate 1 failed: '' found at line . Suggested fix: .”

Gate 2 — Verbatim-quote preservation

Type: Deterministic string-match.

Rule: Every sources[].text_excerpt from frontmatter MUST appear verbatim (byte-for-byte after whitespace normalization) in the script body.

Implementation:

for (const src of frontmatter.sources) {
  const quote = src.text_excerpt.trim();
  const bodyNormalized = body.replace(/\s+/g, ' ').trim();
  const quoteNormalized = quote.replace(/\s+/g, ' ');
  if (!bodyNormalized.includes(quoteNormalized)) {
    return { pass: false, details: `Source ${src.readwise_id} not found verbatim in body. Drift detected.` };
  }
}

Source-of-truth: _brand/voice-preservation.md §1 (verbatim contract).

Failure surface: “Gate 2 failed: source <readwise_id> '' was not found verbatim. Likely paraphrased — restore exact text.”

Gate 3 — Two-beat cadence

Type: LLM-judgment.

Rule: Script body contains at least 1 period/en-dash separated 2-clause structure with cadence balance.

Pattern examples:

• “Tools end conversations. Companions hold them.”
• “Private by design — your data, your rules.”
• “State and relationship. Same thing.”

LLM prompt (skill-internal):

Read the script body below. Identify all instances of period-or-en-dash-separated 2-clause structures where the two clauses have rhythmic parity (similar syllable count or grammatical mirror). Return JSON: {“two_beat_count”: N, “examples”: […], “pass”: (N >= 1)}.

Source-of-truth: _brand/voice-preservation.md §3 (two-beat cadence).

Failure surface: “Gate 3 failed: no two-beat cadence detected. Voice loses signature rhythm. Suggest adding one in the close beat.”

Gate 4 — Promise-not-sell close

Type: Deterministic regex.

Rule: The last beat (final ~5-15s of script body) MUST NOT match sell vocab.

Banned in close:

• preorder / pre-order
• buy now / purchase
• link in bio / bio link
• indiegogo (the platform name in close)
• sign up / sign-up / signup
• join the waitlist / waitlist
• click (as CTA)
• discount / early-bird (in close context)

Implementation:

const lastBeat = extractLastBeat(body);  // last ## Beat block
const sellVocab = /\b(pre-?order|buy now|purchase|link in bio|bio link|indiegogo|sign-?up|waitlist|click|discount|early-?bird)\b/gi;
return { pass: !sellVocab.test(lastBeat), details: ... };

Source-of-truth: _brand/voice-preservation.md §4 (promise-close).

Failure surface: “Gate 4 failed: '' in close beat. Brand voice ends on promise, not sell. Suggested rewrite: .”

Gate 5 — Pillar declared

Type: Deterministic enum check.

Rule: Frontmatter pillar_primary is in {1, 2, 3, 4, 5, 6, 7}.

Implementation:

const valid = [1, 2, 3, 4, 5, 6, 7];
return { pass: valid.includes(frontmatter.pillar_primary), details: ... };

Source-of-truth: _brand/pillars.md defines pillars 1-7.

Failure surface: “Gate 5 failed: pillar_primary missing or out of range. Skill-suggest pillar from clip + reflection content.”

Gate 6 — Tool-naming combination (Pillar-7-conditional)

Type: LLM-judgment. Only runs if pillar_primary == 7 OR pillar_secondary == 7.

Rule: Script body must contain (a) named concrete tools, (b) a “combination doing one job” sentence, (c) low-stakes verbs.

LLM prompt:

The script’s pillar is 7 (Process Transparency). Verify the body contains:

1. At least 2 concrete tools named (e.g., “Claude”, “Figma”, “the skill”, “Higgsfield”, “ffmpeg”, “Whisper”, “the device”, “Nano Banana Pro”, “Cursor”, “Patchright”, etc. — not vague “AI” or “my workflow”).
2. A sentence describing a combination of tools doing one specific job (e.g., “Claude reasons, the skill orchestrates, the device holds the memory” = combination doing job “continuous companionship across sessions”).
3. Low-stakes verbs (open, set up, try, check, grab, paste, show, name) — NOT high-stakes performative ones (deploy, ship, launch, implement, architect, engineer, optimize).

Return JSON: {“tools_named”: […], “combination_present”: bool, “low_stakes_verbs”: […], “high_stakes_verbs_found”: […], “pass”: (all three conditions met)}.

Source-of-truth: _brand/voice-preservation.md §6 + ~/.claude/skills/afriend-content-extraction/priming/tool-naming-checklist.md (4-question drill).

Failure surface: “Gate 6 failed: . .”

If pillar 7 is NOT in primary/secondary: gate is skipped, frontmatter records tools_named_when_pillar_7: "n/a".

Gate 7 — Anti-brain-rot

Type: LLM-judgment. Always runs.

Rule: Script must NOT use:

• Cliffhanger-without-payoff (a hook that promises a reveal but never delivers — “wait until you see what happened next” without the payoff in the clip)
• Performative vulnerability (vulnerability-as-hook with no real stake or insight, just emotional bait)
• Engagement-bait patterns (over-promising, manufactured controversy, fake-urgency, “you won’t believe”, etc.)

LLM prompt:

Read the script body. Per ~/.claude/skills/afriend-content-extraction/references/attention-preservation.md lines 176-192, identify any:

1. Cliffhanger-without-payoff: does the hook promise a reveal/answer/conclusion that the clip body doesn’t deliver?
2. Performative vulnerability: does the vulnerability beat read as honest stake-revelation OR as emotion-bait performance?
3. Engagement-bait: any “you won’t believe”, “this changed everything”, “the secret to”, manufactured-urgency, or click-bait patterns?

Return JSON: {“violations”: [{“type”: …, “location”: “Beat N”, “evidence”: ”…”}], “pass”: (violations.length == 0)}.

Source-of-truth: ~/.claude/skills/afriend-content-extraction/references/attention-preservation.md lines 176-192.

Failure surface: “Gate 7 failed: at Beat N. . Brand explicitly does NOT do brain-rot — rewrite or discard.”

Gate 9 — AI-style ANTI-PATTERNS L1 BLOCKING (added 2026-05-06, compose canon integration)

Type: Deterministic regex. Runs over Daniel-voice spans (verbatim quotes excepted).

Rule: Script body must NOT contain any of the 13 L1 BLOCKING patterns from /Users/mikroverse/Projects/onari-intelligence/_canon/voice-deseltrus/ANTI-PATTERNS.md. These are AI-jargon markers that destroy authenticity.

Patterns (regex/match):

#	Pattern	Detection
L1.1	”It’s not X. It’s Y.” family	`\b(?:it’?s
L1.2	”Not because X. But because Y.”	\bnot because\b.*?\.\s*but because\b
L1.3	”Not that X. But that Y.”	\bnot that\b.*?\.\s*but that\b
L1.4	”That’s [Name].” punchy intro	`(?:^
L1.5	”Here’s the thing:” / “Here’s how” filler	`\bhere’?s (?:the thing
L1.6	”Let me break this down”	`\blet me break (?:this
L1.7	Em-dash — or --	`(?:—
L1.8	”In der heutigen schnelllebigen Welt”	\bin der heutigen schnelllebigen\b
L1.9	”Game-changer” / “Next level”	`\b(?:game[- ]?changers?
L1.10	”Unlock your potential”	`\bunlock (?:your
L1.11	”5 Tipps für mehr…” listicle	`\b\d+\s+(?:tipps?
L1.12	”Kennst du das auch?” pseudo-question	`\bkennst du (?:das
L1.13	”Navigating our AI-infused…” buzzword	`\bnavigating\s+(?:our

Implementation: scripts/voice-safety-check-deterministic.mjs exports gate9AntiPatternsL1(body, sources). Excludes verbatim-quote spans (sources[].text_excerpt — those are sacred even if a foreign speaker uses banned vocab).

Source-of-truth: /Users/mikroverse/Projects/onari-intelligence/_canon/voice-deseltrus/ANTI-PATTERNS.md Level 1 BLOCKING table.

Failure surface: “Gate 9 failed: at . AI-jargon — rewrite. See _canon/voice-deseltrus/ANTI-PATTERNS.md for the principle.”

Critical implication for educational format: my own S02 smoke-test draft (“Conversation as Company Architecture”) FAILED 3 instances of these — “Conversation is not a feature. It’s the architecture.” (L1.1), “That’s Jack Dorsey…” (L1.4), em-dash usage throughout (L1.7). Gate 9 enforces these now.

Gate 10 — AI-style ANTI-PATTERNS L2 WARNING (added 2026-05-06, compose canon integration)

Type: LLM-judgment. Format-conditional (always runs for scripts).

Rule: Script body should not exhibit L2 WARNING patterns from _canon/voice-deseltrus/ANTI-PATTERNS.md. These are quality issues — corrigible, not blocking.

Patterns to detect:

#	Pattern	Description
L2.1	Difficulty Flattening	Heavy thought translated to easy language. “Compounds in ways I’m only starting to see” instead of executing the visionary thought-stream.
L2.2	Bridge Substitution	Daniel’s unique bridge replaced with external enrichment. Brooks’ Law displaces Daniel’s user-bridge.
L2.3	Vision Gesture	Pointing AT the vision instead of executing it. “That compounds in ways…” is a gesture, not a thought.
L2.4	Accessibility Override	Every paragraph optimized for max comprehensibility. ALL flattened to one register (practical-reflective LinkedIn). Visionary + meta-cognitive registers disappear.
L2.5	Announcer-Colon Setup	”What’s emerging is an evolution: [concept].” Build-up + colon + reveal = AI-typical pattern.
L2.6	Hook → Problem → Solution → CTA	Too formulaic. Organic thought-flow preferred.
L2.7	Each sentence as own paragraph	LinkedIn-cringe. Normal paragraphs where sensible.
L2.8	Übertriebene Begeisterung / Pseudo-Verletzlichkeit / Prediger-Ton / False Humility	Inauthentic emotional registers.

LLM prompt:

Read the script body. Per _canon/voice-deseltrus/ANTI-PATTERNS.md Level 2 WARNING patterns, identify any matches. Return JSON: {“violations”: [{“pattern”: “L2.N”, “evidence”: "", “fix_advice”: ""}], “pass”: (violations.length == 0)}.

Source-of-truth: _canon/voice-deseltrus/ANTI-PATTERNS.md Level 2 WARNING.

Failure surface (warning only): “Gate 10 (L2 WARNING): patterns flagged — review and adjust before commit. Drafts pass anyway, audit trail preserves the violations.”

Gate 8 — Credibility / skin-in-game (added 2026-05-06)

Type: LLM-judgment. Format-conditional — only runs if format ∈ {educational-talkinghead, educational-greenscreen}. Runs in advisory-only mode if reflection’s daniel_stake_optout: true.

Rule: Educational/thought-leadership scripts MUST demonstrate Daniel-skin-in-game, not just summarize the source clip. The brand-pillar promise of educational format = “credibility + social validation through the founder’s stake in the field.” Without this, the clip reads as expert-summary and loses A Friend’s unique-positioning value.

Three sub-checks:

1. Daniel-stake explicit: at least one beat names a specific stake/build/decision/number from Daniel’s own founder context (A Friend, Onari Intelligence Bridge, the Indiegogo campaign, named milestones, real numbers, real architectural choices). Vague “we’re building” without specifics doesn’t count.
2. Beyond-clip insight: the script contains a claim or framing that wouldn’t exist if Daniel weren’t shipping A Friend himself. The cite is evidence FOR Daniel’s thesis, not summary of the speaker’s thesis.
3. Cite-not-endorsement: no implicit “this expert validates A Friend.” The speaker spoke about something else (Block restructuring, AI memory, etc.); Daniel uses architectural framing only. Banned phrasings: “validates what we’re doing”, “proves that”, “confirms our approach.”

LLM prompt (skill-internal):

Read the script body and the daniel-reflection.md Daniel-Stake section. Verify all three sub-checks above. Return JSON: {“daniel_stake_present”: bool, “stake_evidence”: "", “beyond_clip_insight”: bool, “insight_evidence”: "", “cite_as_evidence_not_endorsement”: bool, “endorsement_violations”: […], “pass”: (all three true)}.

Source-of-truth: Daniel’s locked decision 2026-05-06 — “diese Credibility und Social Validation auch in einer Form abbilden… Skin in dem Game.” Plus brand pillar definitions in _brand/pillars.md Pillars 5 (Startup Realness), 6 (Co-Founder Bridge), 7 (Process Transparency).

Failure surface: “Gate 8 failed: . . Educational format requires Daniel-stake — rewrite the DANIEL-STAKE beat or discard.”

Advisory mode: if reflection has daniel_stake_optout: true, Gate 8 runs but only WARNS instead of failing. Frontmatter records credibility_skin_in_game: "advisory-warn". Daniel can ship a pure-summary clip if he explicitly chose that path; the advisory-warn value flags it for downstream auditing (e.g., M114 refinement may surface it for re-draft).

Skip mode: if format is NOT educational, Gate 8 doesn’t run at all. Frontmatter records credibility_skin_in_game: "n/a".

Frontmatter audit format

After all 10 gates + Gate 0 meta-check run, write into script.md frontmatter:

voice_safety:
  cognitive_depth: true | false           # Gate 0 (auto-regen if false; final value reflects last attempt)
  banned_vocab_check: clean | violations  # Gate 1
  verbatim_quotes_unchanged: true | false # Gate 2
  ends_on_promise: true | false           # Gate 4
  two_beat_present: true | false          # Gate 3
  pillar_declared: true | false           # Gate 5
  tools_named_when_pillar_7: true | false | "n/a"  # Gate 6
  credibility_skin_in_game: true | false | "n/a" | "advisory-warn"  # Gate 8
  anti_patterns_l1_blocking: clean | violations    # Gate 9 (deterministic, BLOCKING)
  anti_patterns_l2_warning: clean | violations     # Gate 10 (LLM-judgment, WARNING only)

If gate 7 fails, also surface in body as inline comment block above the affected beat. Gate 7 audit doesn’t have a frontmatter field by default — could add anti_brain_rot_check: clean | violations if Daniel wants persistent tracking. Default for now: surface inline only.

Failure handling

Per modes/commit-or-discard.md: any gate failure surfaces in Mode D as a ⚠ warning, but doesn’t block commit (Daniel can accept-anyway). Audit trail preserves the failure for downstream consumers (M114 refinement skill, eval pipeline, etc.).