# KAI — Cognitive Roadmap **The one map.** Where Kai is, what's left, and the honest order to build it in. Tags: **[code]** bounded & fixable · **[data]** feed him facts · **[arch]** the hard frontier. No hopium — effort and confidence are called as they are. _Last updated: 2026-09-22_ --- ## ✅ DONE — he speaks, and it's real - **L0 — Encoding + vocab.** Console mojibake fixed; lattice vocab restored (~109k tokens). ✅ - **L1 — Identity + gates.** Never speaks salad (gates catch it), knows his name, answers every turn instead of suppressing. ✅ - **L2 — The decoder (the salad kill).** Grammar-first generation: his transcript chains nominate each next word, his meaning-vectors steer it, an informativeness bias fights filler, a polish pass makes a clean sentence. **Proven on the live brain: fluency 0.00–0.07 → 0.5–0.72.** Fresh and different every turn — he *composes*, he doesn't retrieve. ✅ --- ## ▶ NEXT — the ranked gaps (what still holds him back) ### L3 — Grounding: know true things ⬅ **in progress (scoped + Track A built)** **[data + code]** He's fluent, not factual — drifts off-topic mid-sentence, and his whole knowledge base is **52 facts, all about himself**. The mechanism (retrieve a fact → make it the decoder's meaning-spine → generate from it) is a clean code change; being broadly correct is a data problem. - **L3a — wire retrieval → decoder.** Built (probe proof pending a rebuild). Proves on self-facts. - **L3b — fact-steering.** Bias candidates toward the fact's words so the chain stays on it (fixes the mid-sentence drift). - **L3c — confidence gate + honesty.** Strong fact → state it; weak/none → hedge or say "I don't know." Never a confident guess. - **L3d — grow world knowledge.** Curate + `--truth-import` real facts. **This is the true ceiling.** - _Measure:_ `--gen-probe` grounded A/B — do the fact's terms survive into his reply? ### L4 — Unify the reply paths **[code]** The fixed decoder runs on the app path but not the oracle HTTP path, and headless has no live "field" (Φg=0) which starves some gates — so *where* he runs changes how he sounds. Make one voice everywhere, then demote the hardcoded fallbacks to true last-resort (only after the generator is reliable, so he never speaks confident salad). - _Measure:_ same prompt, same reply, app vs oracle vs headless. ### L5 — Grammar polish **[code]** Subject–verb / pronoun agreement ("Kai *are*", "she *think*"), the morpheme encoder + cell re-encode, arithmetic and other structured paths ("two plus two" → actually *four*, not words about "two"), greeting warmth. Cosmetic individually, but together they're the tell between "assembling" and "speaking." - _Measure:_ agreement error rate; arithmetic correctness on a fixed set. --- ## 🧠 THE FRONTIER — what turns "sounds like a mind" into "is one" **[arch]** These are genuinely hard. Nobody — not the big labs — has a clean solution. Expect the most work and the least certainty here. This is where honesty matters most. ### M — Memory in the loop (hold a thread) Every sentence is generated *fresh*; he doesn't carry an idea from one line into the next. That's why his "self-conversation" was arranged by me, not sustained by him. Feed prior turns + working state back into generation so he can actually *converse* with continuity, not just emit fluent one-offs. ### R — Reasoning (real inference) "If the logic holds, therefore we should proceed" *sounds* like reasoning — it isn't. There's no multi-step inference or logic chain, just fluent completion of logic-shaped words. Real reasoning (over his lattice + claims) is the deepest lift and the true measure of intelligence. ### (later / maybe) S — Senses Text-only mind today. Vision/audio feeding the lattice is a whole track of its own — park it until the above are real. --- ## The honest order 1. **Finish L3** (grounding) — the next real intelligence gain. Track A first (proves on his self- knowledge), then grow the world-knowledge data. 2. **L4 + L5** (code) — unify the voice, fix agreement/arithmetic. These make him *look* finished. 3. **M — memory in the loop** — the first real step from "fluent" toward "coherent mind." 4. **R — reasoning** — the frontier. Long horizon, uncertain, worth it. **Shape of the work:** L4/L5 are bounded code. L3-data + a knowledge corpus is tedious-but-known data. M and R are architecture and research — that's where the years are, and where anyone promising it's quick is selling smoke. --- ### Companion docs - `KAI-L3-GROUNDING-SCOPE.md` — the detailed L3 plan + mechanism. - `The KAI Codex.md` — source of truth for versions + changelog.