What changed: the cortex speaks, rough

Theory update. 2026-06-26 · the cortex starts to speak · the acquisition-queue round

Three assumptions shifted this round. The cortex stopped being only a reader and became a rough speaker, and "production is comprehension read the hard way" got its first evidence. A reply that answers you teaches more than the same words overheard cold: the reactive bet has first support, and it now survives a live model. And the no-curriculum rule generalized again: a count learner needs no difficulty schedule and no write-throttle. Here is what changed and the exact line it adds to the standing theory.

What this round was

The whole acquisition build queue, run at once: sixteen experiments from AU to BJ, plus three follow-ups. The full scorecard and every number lives in the round narrative. This post is the synthesis: what the round changed in the assumptions, and where it sends the next swing.

(a) The cortex is a speaker now: rough, but real

The theory said production is comprehension read the hard way: the same counts a reader uses to predict the next word, read backwards as a many-to-one selection to say one. This round gave that claim its first evidence. The chunk lexicon (Exp AU) mints whole committed units (the within-word transition decays to 0.0003 where pure transitional probability holds it at 1.000) and those units become the vocabulary the generation turn emits. The producer (Exp BD) says them +18.5 points more well-formed and 28.5% less over-generating than a flat sampler, and the follow-up (Exp BL) pushed frame-survival to 87.5%. Every emitted word is spelled out of committed chunks. The chain composes end-to-end: chunk, construct, say.

The honest boundary: generation is rough, not fluent, and the chunk lexicon earns its place as a segmenter and emission vocabulary, not as a predictor. Exp BK settled that cleanly. Voting the chunk distribution into the prediction pool makes next-letter cost monotonically worse at every positive weight; a calibrated backoff already owns the chunk's confident completions. So the lexicon's value is the unit, not the guess. That sharpens the thesis rather than denting it: production reads the counts the hard way, but the organ that makes production possible is a vocabulary, not a better reader.

(b) Contingency teaches: the reactive bet has first support

The reactive loop was the central wager: a token that closed a turn should teach more than the same token read passively. Exp BE confirmed it. With the same tokens and only their timing scrambled in the yoked control, real contingency wins by +0.45 bpc, and the gap vanishes exactly when timing stops predicting content, the signature of a real mechanism. A reply that answers you teaches more than one overheard cold.

And the live-model version ran. No API key was needed: the machine's logged-in claude CLI is itself an authenticated interlocutor, so the loop shells out to it for one headless turn per reply. Against claude-haiku-4-5, forty real replies captured live, the win survives: contingency-ON beats the scrambled-timing yoked control on the agent's surprise at the replies at 12 of 12 sweep settings, best margin +0.25, and on turn-overlap at 12 of 12. One honest caveat we keep: the agent's own utterances are still per-char gibberish, so the model often answered them as nonsense. What is confirmed is that a contingent reply teaches more than the same replies re-timed to ignore the agent, not that the cortex held a real conversation. The mechanism is live; the fluency is not there yet.

(c) The no-curriculum rule generalized: to writes and to structure

The standing rule from starting small was that a count learner can't get stuck, so a difficulty curriculum is a no-op: only the final memory matters. This round extended it twice. Exp BI tried throttling writes on a Goldilocks curve and lost to writing everything down, while hurting the rare tail it skipped: the write-side echo of the same rule. Exp BJ ordered the stream by embedding depth and tied full-input-from-start on center-embedded agreement: the rule now covers structural ordering too. No input curriculum, no write-throttle, no structural curriculum. A count learner does not need to be eased in along any axis.

The synthesis

The theory now reads with one thing settled that was open at the start of the round: the cortex can produce, not only read. Production is comprehension read the hard way, and we have the first measured evidence: a chunk lexicon that supplies whole units, a producer that says them more well-formed than gibberish, frames preserved 87.5% of the time. Contingency teaches, with first support from a reactive loop and now a live model on the other end of it. And the no-curriculum rule now stands across input order, write-gating, and structural ordering at once. The spine did not move; it grew a mouth.

The frontier

Two probes are named precisely. The live reactive loop is done: Exp BE's contingency gate now runs against a live Haiku partner through the logged-in CLI and the win holds; the next move is to put Exp BC's recovery loop on the same live channel and to let the agent's utterances grow past gibberish so the conversation is real, not just the contingency. And generation's ceiling (the ~53% well-formedness the construction structure alone reaches, lifted to 80% only by the frame-true representative) points at the same coherence frontier the reader hit: a message or referent cue, not just a lexical frame, is what would push production past it.

The proposed root changelog line

2026-06-26 · The cortex speaks (rough): production is comprehension read the hard way. A chunk lexicon (Exp AU) supplies whole committed units as emission vocabulary; a coverage-competition producer (Exp BD/Exp BL) says them +18.5 pts more well-formed than flat sampling and preserves frames 87.5% of the time: generation is real but rough. The chunk lexicon is a segmenter and emission vocabulary, not a predictor (Exp BK). A contingent reply teaches more than the same words overheard cold (Exp BE, +0.45 bpc over the scrambled-timing control), and the win survives a live model (Exp BM): run against claude-haiku-4-5 through the logged-in CLI with no API key, contingency-ON beats the yoked control at 12 of 12 sweep settings, first support for the reactive-loop bet, now live-confirmed (the agent's utterances are still gibberish, so it is the contingency that is confirmed, not fluent dialogue). The no-curriculum rule generalizes: a count learner needs no write-throttle (Exp BI) and no structural curriculum (Exp BJ), only enough final memory.

Lineage

Grew from the harness substrate (Exp AT) and the acquisition library it asked for. Folds into the round narrative, the acquisition queue, which carries every number.

Thread: acquisition and generation as one mechanism read in two directions. This is the round where the cortex first read it backwards and spoke.