# grai: Full Technical Specification & Empirical Voice Architecture (v3.1.1) > The personal voice engine. Write like you. Prove it. Built by Adrian M. Peticila (`https://peticila.ro/#person`, Wikidata `Q140362129`, ORCID `0009-0005-9601-9527`). --- ## 1. Product Thesis: Why Banned-Word Lists Stopped Working Between January 2023 and September 2026, grai audited 134,522 organic LinkedIn posts across 1,499 verified founder profiles (LigoSocial dataset: continuous publishing Q1 2023 to Q3 2026, minimum 2,000 followers, primary language English). | Metric | Q1 2023 | Q3 2026 | Observed Delta | |---|---:|---:|---:| | Opening Formula ("Here is why" / "The truth is") | 1.3% | 9.6% | +7.4x | | Forced Opposition ("Not X, but Y") | 1.2% | 5.9% | +4.9x | | Arrow & Emoji Bullet Cascades (`->`, `↳`, `πŸš€`) | 0.6% | 12.3% | +20.5x | | Mocked Lexical ClichΓ©s | 4.2% | 0.5% | -88.1% | | Mean Sentence Burstiness ($CV = \sigma / \mu$) | 0.48 | 0.23 | -52.1% | Authors deleted the vocabulary everyone mocked and kept the syntactic rhythm that gives a language model away. Prompting an LLM to "sound human" fails because the model averages its training distribution. grai solves this using the **Harness over Fine-Tuning** architecture (aligned with Google Research Harness2): model weights remain frozen, while a deterministic local harness enforces quantitative cadence rules outside the model. --- ## 2. The Deterministic Signal Chain (Three Rooms) ### 01 CADENCE GATE (Deterministic Browser Engine) Runs 100% locally in JavaScript with zero network calls. Every draft is stripped of invisible zero-width watermarks (`U+200B` through `U+200D`, `U+FEFF`) and scored from 4 to 100 against six structural rules plus a mathematical rhythm floor: 1. `CONTRASTIVE_NEGATION` (First hit: -14 pts, Repeat: -8 pts): Catches synthetic antithesis formulas where a sentence negates a strawman before stating the point. 2. `LESS_MORE_SYMMETRY` (First hit: -10 pts, Repeat: -6 pts): Catches corporate symmetry cliches (`"Less X, more Y"`, `"Stop doing X. Start doing Y."`). 3. `EPIPHANY_OPENER` (First hit: -10 pts, Repeat: -7 pts): Catches throat-clearing hooks (`"Here is why"`, `"Here's why"`, `"The harsh truth is"`, `"Let that sink in"`, `"Read that again"`, `"Unpopular opinion:"`, `"Most people think X. They're wrong"`). 4. `ARROW_BULLET_SLOP` (First hit: -12 pts, Repeat: -4 pts): Catches template arrow cascades (`->`, `=>`, `β†’`, `↳`, `βœ…`, `❌`, `πŸš€`, `πŸ”₯`, `1/`) with full UTF-16 astral plane awareness (`/u`). 5. `TRIADIC_PADDING` (First hit: -12 pts, Repeat: -8 pts): Catches empty slogan triads (`"No fluff. No theory. Just results."`). 6. `LEXICAL_SLOP` (First hit: -6 pts, Repeat: -5 pts): Catches remaining synthetic vocabulary tells (`fast-paced world`, `ever-evolving`, `unlocking your potential`, `symphony of`, `navigating the complexities`, `paradigm shift`), with negative lookahead protecting legitimate domain nouns. 7. `FLAT_BURSTINESS` (Penalty: `round((0.32 - CV) * 65)` when $CV < 0.32$): Sentence tokenization protects decimals (`3.14`, `$1.5M`) and abbreviations (`Dr.`, `U.S.`, `e.g.`), then computes the Coefficient of Variation $CV = \sigma / \mu$ across sentence word lengths. Human writing stays above $0.38$; raw AI hovers around $0.20$. grai enforces a strict floor at $CV = 0.32$. ### 02 THE WRITING DESK (3-Column Workbench & Backdrop Overlay Editor) - **Left Column (Client Drawer):** Multi-client voice switcher, voice targets, 1-click `.grai` profile backup and restore, and A4 certificate trigger. - **Center Column (Editor & Three Galleys):** - Premise Sparring challenges safe or unanchored ideas (`CLICHE_OR_SAFE`, `SOLID_BUT_UNANCHORED`) with three exit paths: `COUNTER_PUNCH`, `ADOPT_ANGLE`, or `BYPASS`. - Generates three distinct structural galleys simultaneously: 1. **THE SCALPEL:** Short hook, compressed argument, high impact. 2. **THE BROADSHEET:** Narrative arc, concrete scene first, lesson last. 3. **THE LEDGER:** Empirical evidence and hard numbers first, backed by a full audit trail. - Native `