A conversation with Haiku

2026-06-28 · the cortex holds a live conversation · experiment HAIKU-LOOP

The generation track had a producer and an offline answer key. This round puts the producer in a live conversation. A real Haiku partner speaks, the cortex comprehends the turn through its frozen comprehension monitor, the cortex speaks back by emitting the real words it can voice for the meanings Haiku activated, and the reply goes back to Haiku. Fourteen exchanges, live, seed 0. The cortex emits 42 of 42 real dialog words, shares an exact word with Haiku in 10 of 14 turns, and Haiku stays on the cortex's topic in 12 of 13. And a few times the loop does something more: a topic word threads across turns, the cortex says it, Haiku reflects it, the cortex says it again. This is the honest starting line for live dialogue. It is word-salad with real glimmers of relevance, not a conversation that holds.

The setup

The last round built a producer that says real words on purpose. A separate learned inverse, trained on its own feedback against a dialog-trained ear, produces the words of real conversation and the frozen judge recovers the intended one 0.317 of the time. That round graded the producer offline, one meaning at a time, against an answer key.

This round is the same machinery in a live conversation. Nothing about the cortex changes. The comprehension monitor, the producer, and the emission vocabulary are imported unchanged from the offline round. What is new is the loop.

The partner is a real Haiku, run with no API key through the logged-in claude -p command in a child register, the same live partner the reactive-loop result used. Haiku opens with a short simple line. Then one exchange runs like this:

• Haiku speaks.
• The cortex comprehends the turn. It runs each content word through its frozen comprehension monitor, which returns a distribution over the 160 real-word meanings the cortex knows, and sums those distributions across the turn. The top of that sum is what the cortex understands Haiku's turn to be about, in its own small vocabulary.
• The cortex produces a reply. It takes those top meanings, the ones Haiku's turn most activated, and emits the real word it can voice for each, preferring the meanings it can actually pronounce. The reply is a three-word utterance, about what Haiku said.
• The reply goes back to Haiku, and Haiku answers. That answer is the next exchange.

Read and speak run through the one frozen pathway. The cortex understands a turn and answers it with the same organ, used forward to comprehend and backward to produce. That is the thesis of the whole generation track: production is comprehension inverted.

The conversation

Three exchanges from the session, verbatim.

HAIKU : That's a fun fan! Do you have one at school?
CORTEX: for fan school

HAIKU : you want to find a match for a place? That sounds fun!
CORTEX: for place match

HAIKU : Do you like to dance? That sounds like so much fun!
CORTEX: match dance for

The cortex hears fan and school, and says them back. It hears place and match, and says them back. The words are not glued into grammar, but they are real, and they are Haiku's.

What threads

The thing worth watching is not a single turn. It is what happens across several.

A topic word forms a call and response that holds for a few exchanges. Around the middle of the session the cortex emits match; Haiku picks it up ("you want to find a match for a place?"); the cortex comprehends Haiku's turn, re-activates match, and emits it again ("for place match"). Exchanges five through eight sustain that shared match thread. Later, exchanges eleven through fourteen lock onto dance the same way: the cortex says it, Haiku reflects it, the cortex says it again.

So the reactive read-speak-hear loop produces, a few times, a coherent multi-turn topic word, not just an isolated reply. The cortex and Haiku are on the same word, passing it back and forth. That is the "cool to see" part, and it is real.

The measures

Light and honest, on the fourteen exchanges.

Every word the cortex says is a real word, because the producer only voices the words it can pronounce. More often than not, one of those words is a word Haiku just used. And Haiku, almost every turn, reuses a word the cortex said, a soft sign that the partner understood the cortex enough to stay on it. The comprehension peak is weak in absolute terms (0.105 of the mass on the top meaning) but it is about seventeen times chance: the read is doing real work, not firing at random.

The honest assessment

This is word-salad with real glimmers of relevance, the honest baseline for live dialogue at this stage. The cortex emits real, grammatically loose three-word utterances that more often than not contain a word from Haiku's turn, and a few times sustain a coherent shared topic. It is not a conversation that holds. It is the first one the loop can have.

The glimmers are real and so is the gap. The clearest thing the session shows is the production gap the offline round predicted. The cortex comprehends a turn well enough to land on the right region of meaning, but the producer often articulates a neighbor of the meaning it reached for, not the exact word. Comprehension runs ahead of production, which is exactly the shape of the offline result, now visible live: the cortex understands more than it can say.

The honest caveats

• Word-salad with glimmers, not solved dialogue. The replies are real words, often relevant, sometimes threaded across turns. They are not grammatical sentences, and the conversation does not hold a thought for long. This is a starting line.
• The production gap is the live number to lift. The cortex reaches for a meaning and the producer voices a neighbor. Relevance (a shared word) runs ahead of recovery (the exact intended word), the same gap the offline round measured, now in a conversation.
• The keyless partner breaks character sometimes. The claude -p partner is the project-aware Claude Code agent, not a bare model, so about three turns in fourteen it steps out of the child-register pretend chat and answers as an assistant. A short role reminder prepended to each turn the partner hears (which reframes the partner's turn only, never the cortex's words) helped, and the cortex's child-like replies pull it back into character. A bare-model key would be cleaner.
• One live session, seed 0. A live partner is not deterministic across runs, so this is this session, captured verbatim, not an average over many.

The frontier

The loop is live, and the cortex stays on Haiku's topic more often than not. The number to lift is the production gap: a sharper ear, a less coarse meaning region, would turn the neighbor the producer voices into the word the cortex meant, and turn relevant words into relevant sentences. And the loop still only agrees with itself and a partner that mirrors it: it has no shared convention anchored against private drift. The external half, a referential game where a listener has to recover the meaning and the convention is held against drift, is the named next move, now after a live read-speak-hear loop. The cortex can hold a conversation. Giving it a sharper ear, and a real stake in being understood, is the next run.

Lineage

Grew from producing real words, which built the producer this round speaks with and graded it offline against an answer key, leaving exactly this question: can the same organ hold a live conversation. And from the acquisition queue, which first wired the keyless Haiku partner and showed a contingent reply teaches more than the same words overheard cold, the reactive bet this loop runs on.

Thread: generation, and the comprehension-production gap. The mechanism is the separate learned inverse trained feedback-first, now reading and answering a live partner through its one frozen comprehension pathway. The producer is real, the loop is live, and the glimmers of relevance are the honest first sign of dialogue. The frontier is a sharper ear (close the production gap) and the external half (a listener that makes the convention shared).