Skip to content

Latest commit

 

History

History
200 lines (174 loc) · 9.8 KB

File metadata and controls

200 lines (174 loc) · 9.8 KB

Changelog

V9 (2026-04-15 -- present) -- Experimental Structural Decomposition

  • Structural pendulum: decomposes sentence into equation (SUBJECT+EVENT+CONTEXT) before force accumulation
  • Lemma root system: collapses 4,544 words to ~500 emotional roots
  • Molecular bonding layer: 8 bond types, reaction table, two-layer molecules
  • Star-to-star gravity: heavier emotional atoms pull lighter ones
  • Phase-aware alloy composition: LIQUID words default to presumed charge
  • Ported all V8 systems: structures, force flow, zones, crisis, anomaly, solver, battleship
  • W→V coupling: low self-worth amplifies negatives, suppresses positives
  • Absence scope + forced choice cancellation in interpret_context
  • Genetic optimizer: champion knobs tuned across 50+ generations
  • Perspective dampening: OTHER_REF emotions dampened unless directed at SELF
  • 4-model council consensus (Claude, Gemini, GPT, Grok) on 500+ sentences
  • Confidence scoring: certain/majority/ambiguous truth split
  • Status: experimental, ~10pts behind V8 on stress tests

V8.3 (2026-04-10)

  • Conversational accuracy: 73.1% → 98.5% on conversational sentences
  • Crisis recall: 80.4%
  • 3,646 verified sentences accumulated

V8.2 (2026-04-08) -- Council Consensus

  • Full 4-LLM council consensus applied
  • Complexity dampening for literary register
  • Real-text spot-check: ~59% on 167 manually verified sentences across 5 corpora

V8.1 (2026-04-06) -- Interpret Context Layer

  • Discourse markers: "actually", "honestly", "I mean" as register signals
  • Negator inversion in context
  • Register dampening for casual text
  • Counterfactual detection

V8 (2026-04-04) -- Mass Zero + Real-Data Audit

  • Mass-zeroed 646 GAS atoms (inflated neutral words neutralized)
  • 275-sentence stress test: 73.1% (201/275 across 11 categories)
  • 126-sentence crisis benchmark: 70.6% recall, 0% false positive
  • Real-data audit: 59% accuracy on 167 sentences from 5 corpora (novels, Twitch, Reddit, philosophy, game dialogue)
  • 6 physics problems identified and documented
  • Vocabulary: 4,544 curated words (up from 4,108)
  • Structural patterns: 45+ (up from 26)
  • New patterns: MUNDANE_HYPERBOLE, BOUNDARY_VIOLATION, SELF_ERASURE, DIVESTITURE, METHOD_FIXATION, RARITY_MARKER, ABANDONMENT, LIFE_ACHIEVEMENT
  • SOLVENT dissolution: casual register flips LIQUID negative to positive
  • Mundane dampening: inert gas absorption of crisis energy
  • interpret_context layer: discourse markers, register detection, counterfactual inversion
  • Tests: 207 (up from 167)

V7 (2026-04-02) -- SOLVENT Physics

  • SOLVENT word role: REGISTER_CASUAL dissolves LIQUID atoms via phase physics
  • Pure physics solutions for context-dependent meaning

V6 (2026-04-01) -- Physics Upgrade

  • Contradiction sarcasm detection
  • Atmospheric grief handling
  • Adaptive momentum
  • Pipeline refactor: pendulum.py split into 8 pluggable stages
  • Port V2 anomaly detector: trajectory analysis for conversations
  • Pipeline trace module added

V5.5 (2026-03-31) -- 7D VADUGWI, Force Flow, Phi-4 LoRA, Bayesian Corrections

Engine

  • Full 7D VADUGWI coordinate system: V, A, D, U, G, W (Self-Worth), I (Intent)
  • Self-Worth (W) dimension: tracks user self-evaluation thread (shattered -> stable -> strong)
  • Intent (I) dimension: withdraw / deflect / neutral / connect / control
  • Force flow resolver (engine/force_flow.py): WHO does WHAT to WHOM directional analysis
  • Absence scope: "havent had X" dampens absent events instead of scoring them positively
  • Compound phrase resolution: "no one" -> nobody, "everyone" -> universal scope
  • Bayesian vocabulary corrections: over-weighted words identified and neutralized across 11 cycles
  • Forced choice cancellation: "A or B" does not double-count both options
  • RELIEF_ABSENCE, SELF_EXCLUDED, WITHHELD_POSITIVE structural patterns added
  • Confidence gate: NULL / LOW / MODERATE / HIGH output modes
  • 26 structural patterns total (up from 22 in V3.2)

Vocabulary

  • 4,108 curated words with 7D force vectors (up from ~2,400 in V3.2)
  • Cycle-by-cycle Bayesian neutralization: highway (+29->0), relationship (+37->+10), and 30+ others
  • 5-way AI consensus validation on all vocabulary cards (Gemini, Claude Opus, GPT-4, Grok, engine)

Tests

  • 167 tests passing across 8 test files (up from 156)
  • Coverage: word classification, structures, proximity, pendulum, solver, battleship, scaffolding, novel sentences

Performance

  • Engine size: ~452KB
  • Speed: 0.15ms/sentence, ~6,500-13,000 sentences/sec
  • SST-2: 69.6% (up from 51% in V3 era)
  • GoEmotions: 75.3%
  • 4-AI consensus benchmark: 76.3% (vs Gemini, Claude Opus, GPT-4, Grok on 131 sentences)
  • Novel sentences: 100% on 630 sentences
  • Crisis detection: 97.3%
  • Sarcasm: 90%
  • Safe text false positives: 0%

Training

  • Phi-4 LoRA training in progress: 52,642 entries, 10 epochs, teaches VADUGWI math
  • Training objective: model learns the force equations, not memorized outputs

V3.2 (2026-03-30) -- Structural Pattern Expansion + SmolLM2 Integration

New structural patterns (8 added, now 22+ total)

  • BETRAYAL: relationship trust weaponized ("wife cheated with best friend")
  • BRAVADO: overcompensation mask ("haha yeah im totally okay")
  • VICTIMIZATION: directional damage, who did what to whom ("she left me" vs "I left the room")
  • CALLING_OUT: complaint disguised as question ("why do you always do that")
  • DIRECTED_POSITIVE: positive aimed at other as dismissal ("good for you", "must be nice")
  • MINIMIZER: shrinking real impact ("it was just a joke", "youre too sensitive")
  • EXCLUDED_POSITIVE: self excluded from positive ("do you even love me", "my parents love my brother more")

Engine improvements

  • Smart CHOPPER: analyzes second-half content before overriding
  • POSSESSION words keep gravity but strip emotional force (objects have weight, not feelings)
  • Strong negative words resist negation (expletives cannot be logically negated)
  • Stemmer fix: -s tried before -es ("bites" -> "bite" not "bit")
  • Contractions recognized: youre, hes, shes, theyre as OTHER_REF
  • "too" added to AMPLIFIER
  • "someone/everybody/anyone" as OTHER_REF
  • SUSPICIOUS_CALM strengthened and excludes achievement contexts
  • FAREWELL excludes "back" (reclamation is not farewell)

Vocabulary expansion (2,400+ words)

  • Violence: stabbed, punched, slapped, choked, attacked, assaulted
  • Mockery: mocked, ridiculed, taunted, harassed
  • Resignation: whatever, k, cool, sure, nvm, idc
  • Achievement: worked, succeeded, graduated, hired, fired
  • Violation: deleted, changed, took, spent, sold, stole, ruined, destroyed
  • Invalidation: overreacting, dramatic, crazy, paranoid, delusional
  • Threat: swear, warn, threatening
  • Medical: herpes, cancer, sick, infected, pregnant
  • Judgment: compare, judge, criticize, blame, fault
  • Temporal intensity: always, constantly, every, forever
  • Exclusion: except, instead, more, prettier, smarter
  • Doubt: even, actually, anymore, supposed
  • Upbringing: foster, adopted, orphan, abused, neglected, molested
  • Milestone: million, verified, published, accepted, hero, dream
  • Resolution: made, well, anyway, survived, overcame

Liquid word fixes (same word, structure determines meaning)

  • "left" V=-45 -> V=-8 (agency vs abandonment, VICTIMIZATION resolves)
  • "give" V=+20 -> V=-3 (generous vs demanding, context resolves)
  • "hit" V=+28 -> V=-45 (violence vs achievement, VICTIMIZATION resolves)
  • "hope" V=+127 -> V=+45 (hope contains uncertainty, not opposite of despair)
  • "finally" V=+29 -> V=+5 (temporal marker, not positive)
  • "calm" V=+39 -> V=+20 (state vs command)
  • "today" V=+37 -> V=0 (time marker, not emotional)
  • "negative" V=-112 -> V=-25 (medical context = good, emotional = bad)
  • "surgery" V=-77 -> V=-25 (past surgery with "well" = relief)
  • "care" G=8 -> G=35 (care = embrace, high gravity)
  • "foster" V=-20 -> V=-3 (neutralizer/dampener, not negative)
  • "fuck" V=-40 -> V=-70 (resists negation)

Accuracy

  • 92% on unambiguous sentences (excluding context-dependent)
  • 85% crisis recall (was ~80%)
  • 100% on genuine positive (zero false positives)
  • 90% on internet speak
  • 80% on body language
  • 90% on conversation fight patterns

Model training

  • V3 model retrained: 7.7M params, 141K examples, 22 patterns
  • Role accuracy: 59.7%, Pattern accuracy: 97.9%, VADUG MAE: 2.8

SmolLM2 / Llama integration

  • Conversation loop: two characters with personalities argue
  • Living conversation: endless interaction until breaking point
  • 6 personalities: hothead, peacekeeper, ice, empath, joker, narcissist
  • LoRA training data: 47K VADUG-conditioned pairs formatted
  • HuggingFace Space live with Llama-3.2-1B via Inference API

Demo

  • Two-character browser demo with persistent emotional memory
  • 5 selectable characters with distinct appearances (skin, hair, clothes)
  • Speech bubbles, conversation log with per-message VADUG scores
  • Trauma tracking with time-based decay

V3.1 (2026-03-30)

  • All 2,315 vocabulary words apply force (not just "emotional" role)
  • Periodic table classification: 1,291 solids, 970 liquids, 54 gases
  • 78K sentence transition map (empirical word-to-word intervals)
  • Ripped out DEATH_WISH hardcoded pattern (physics handles it)
  • Sarcasm false positive fix (requires opener + mundane, not just positive + anything)
  • Pull verb family (chase/pursue/flee/stalk/escape)
  • Power verb family (use/control/command vs submission vs inversion)
  • Surprise as pattern interrupt (A-spike, not V-direction)
  • Shape traces for every sentence (V-line fingerprinting)
  • Browser engine at docs/index.html (78KB data + JS, zero server)
  • Trained V3 model: role 80.2%, patterns 99.0%, VADUG MAE 3.2

V3.0 (2026-03-30)

  • Complete rewrite from previous idiom matching to structural pattern recognition
  • 6-layer architecture: word roles, proximity, structures, physics, solver, battleship
  • 91% on novel sentences (previous was 39%)
  • 86% crisis detection on never-seen sentences
  • 90% sarcasm detection via structural inversion
  • 156 engine tests passing
  • Clean repo (clean history)