5.3 KiB
5.3 KiB
Engineering Session — 2026-08-13 — Language Faculty & the Poem Home
Companion to the book entry the-minds-we-forge/sessions/2026-08-13-the-poem-comes-home.md. Factual log of what was built overnight. All work staged / sandboxed / reversible; the live engram daemon (:8742, pid 31277) was untouched throughout; container-capped folds only; pushed to Gitea for durability.
Summary
The session extended the engram from a memory substrate into a language faculty plus a reasoning + verifier layer, validated with real numbers, and stress-tested on Will's own poem Slowness is Calling.
Built / validated
Language as geometry — translation
- Meaning as a language-independent geometric pivot; translation = routing through it.
- EN→ES→PT→EN "telephone" chain: routed cosine ES 0.973 / PT 0.967 / EN-final 0.969; retrieval top-1 15/15 at every hop. Loss splits geometry=meaning / structure=grammar (grammar errors ≈0 meaning cost; real loss = routing near-misses — the "plausible lie").
- Positioning: universal translation collapses N² language pairs → N realizers; small, local, on-device. Not an alternative to the LLM — an alternative to the LLM-centric paradigm. Honest boundary: the encoder is still a small learned model ("no giant LLM," not "no model").
Fully-functional Spanish realizer (no toy)
- UniMorph Spanish, ~1.2M inflected forms; ~34 syntactic constructions.
- Honest fresh held-out coverage 77.0% (dev-set 100% explicitly disavowed as a claim); zero dropped negations across 140 sentences.
- Realizer-vs-router concerns separated; mechanical ELP (
.el) port plan (avocabulary-es.elgenerator + table transcription; stage via snapshot→verify→blue/green). Sandbox~/Desktop/lang-realizers/; Neuron artifact5d61e6cf.
Poem stress-test + frame-model upgrade — Slowness is Calling
- Baseline through the chain: ORACLE 0.706, ROUTED 0.591 (~⅔ structural / ⅓ geometric). Failure modes: negation deletion (reassurance→accusation), epistemic-frame collapse, metaphor hub-collapse (sea/shore/tide/wave → "ocean").
- Upgrade: structural slots (negation/polarity, epistemic matrix, PP/adjunct/simile — carried structurally, cannot invert) + sense-anchored (gloss-anchored) routing.
- Result: ORACLE 0.706 → 0.777; END-TO-END 0.591 → 0.770 (+0.179). NEGATION preserved 0/11 → 11/11 ("you never fought the ocean" 0.377→0.991; "I was never losing you" 0.501→1.000). sea≠shore 2/6 → 5/6 distinct. Routing slips 54 → 7; every one of 18 verses improved. Sandbox
~/Desktop/lang-chain-experiment/.
Rhyme-preserving translation
- meaning ∩ rhyme composable one-word → rhyme-partnered line-pair; real phonemes EN/ES/PT; 34,030 ES / 33,077 PT real vocabulary.
- Key finding: at real vocab scale the tradeoff moves from existence → cost (rhyme-cost metric). Held ABCB on 16/18 quatrains (6 rima consonante + 10 asonante), mean per-line cosine 0.830; kept meaning on the 2 it couldn't rhyme (incl. truth/roots — already slant in the English). PT mechanism built; PT verse composition pending. Sandbox
~/Desktop/lang-poetic-translation/.
Geometry operators → reasoning → verifier
- Geometry operators (overlap / subtract / combine / distance-Wasserstein / analogy-Procrustes) now live-callable from compiled
elover the real 13,036-node store (via shipped-elcpass-through — no uncapped fold). Commits5336cfe,85eee42. - Reasoning layer (analogy / induction / abduction / causal / planning) — all five done-with-proof, 33/33 closed-form checks, ASan/UBSan clean, 0 leaks. Commit
a3358df. - Verifier layer (grounding + consistency) — proven, 29/29 checks. Catches the plausible lie: a claim grounded in real vocabulary yet polarity-inverted passes grounding, caught only by consistency (complementary checks) — directly flags the reassurance→accusation inversion. Commit
ca13471. - el-exposure of the variadic/point-input reasoning + verifier modes deferred (would need ABI changes risking an uncapped fold); C layer complete + proven.
Whitepaper
engram-cognitive-architecture-whitepaper.mdupdated with the 2026-08-13 validated results (§13/§14/§15/§16/§21), held at Version 1.0 (no bump), ELP64/064,275cross-ref preserved. Commitadc8646, pushed to Gitea.
Roadmap (deferred, not built tonight, per Will)
- Multimodal / images-as-geometry: CLIP-precedent shared image+text meaning-space. Image→meaning near-term + local; meaning→image the hard, asymmetric side. Medical CT as decision-support (retrieval / anomaly-from-normal / progression, all interpretable) — not diagnosis; requires clinical validation + regulatory clearance; clinician holds the call.
Durability / safety
- Pushed to Gitea:
elengram-tiered-storage77a4bc9..ca13471(operators, cutover, reasoning, verifier + reversal docs); whitepaper2440c7d..adc8646; aneurondocs reversal branch. - Live
:8742never touched (pid 31277 unchanged). No deploy, no launch-agent, no~/.neuronwrites. Reversal docs underel docs/runbooks/. No AI-attribution footers.
Still in progress at hand-off
- Portuguese realizer (following the Spanish template).
- English realizer core + US/UK/AU dialects (queued behind PT).
- Frame-model remaining gaps: passive voice, appositive/verbless fragments, resultatives; home→house pivot ambiguity.