Producing on purpose

2026-06-28 · producing a specific meaning, on purpose · experiment PROD-INVERSE

The locked reader recognizes most words and produces few. Asked to play a learned word back, it returns a generic syllable stub. The neuroscience says why: production is not the recognizer reversed. It is a separate learned map, trained feedback-first, the way a songbird learns its song by hearing itself. This round builds that separate map and races it against the reader run backwards. On the matched task, produce one specific intended meaning, the separate inverse recovers 82 of 300 where reversing the reader recovers 2. It wins in every frequency band, including the rare tail. The recovery is moderate, 0.417, and the vocabulary is controlled, but the lesson is clean: production is reachable with a separate organ trained on its own feedback, where the reader reversed is not a producer at all.

The question

The model reads well and speaks poorly, and the last round measured exactly how poorly. A trained reader recognizes 296 of 500 controlled words. Asked to produce them in free generation, it returns 64. Asked to play back a specific learned word, by priming its opening and completing it, it returns the right spelling 2 times in 296. The rest come back as generic frequent stubs: prime "pr" and it says "proa", prime "dr" and it says "drea". The reader learned what most often follows an opening, not the spelling of each word. It is a strong recognizer with no clean way to generate.

So the question is what to do about it, and the neuroscience has a specific answer. The DIVA model of speech says production is a separate learned map from an intended meaning to an output sequence, trained feedback-first against a sensory target, and that the target is a precision-weighted region, not an exact string. The same shape shows up in inner speech (the producer run internally) and in mirror systems (the shared code between hearing and saying is learned, not given). The prediction is sharp. The fix for the wall is not a better way to run the reader backwards. It is a second module, trained the way the birdsong loop trains: babble, hear yourself, correct toward what you meant. And the place it wins is the novel, low-frequency target, where there is no high-frequency stub to fall back on.

What we tried

Two arms on the controlled vocabulary, so every produced word is gradable against a 500-word answer key. Three hundred target meanings, stratified across the frequency range so the rare tail and the frequent head are both represented. One seed. The score is read by a frozen judge the producer never trained against: a word is valid if it is a real vocabulary word, and recovered if the judge maps it back to the meaning that was intended. Never bits-per-char.

• The separate inverse (arm A). A producer with its own per-meaning counts over the chunk-lexicon emission vocabulary, trained by the self-feedback loop. For each intended meaning it babbles candidate utterances, re-comprehends each through a frozen comprehension monitor to recover what it would understand, scores the candidate by how well that recovery lands in the meaning's region, and keeps or drops its emission counts by win-stay-lose-shift. No gradient runs through the comprehension model. The whole map is learned from self-hearing, online and bounded.
• The reader run backwards (arm B). The locked reader from the last round, used as a producer. Two ways, both the thing DIVA says is the wrong design. Prompted recall per meaning: prime the word's opening and greedily complete it, the conditional task that matches arm A. And free generation: sample the trained next-character head, the unconditional metric.

The first arm is a new module trained feedback-first. The second is the recognizer pointed backwards. The matched comparison is the conditional task, where both are asked to produce one specific intended meaning and graded by the same judge.

What happened

The matched task is conditional production: name a specific meaning, and let the frozen judge say whether the right meaning came back.

The separate inverse produces 82 of 300 distinct valid meanings and the judge recovers the intended one 0.417 of the time. The reader run backwards, asked the same question, produces 2. The gap is the whole result. When the task is to produce a specific meaning, the separate inverse reaches it and the reader reversed does not.

And the inverse learns it. Over forty episodes of the self-feedback loop, recovery climbs from 0.060 to 0.417, distinct valid words climb from 61 to the low eighties, and self-error falls from 0.977 to 0.871. The loop, not the starting producer, does the work: babbling and hearing yourself and correcting toward the meaning is what builds the map.

The frequency breakdown is where the neuroscience predicted the win, and it holds across the board.

The separate inverse wins every band, and it wins recovery in every band too. The rarest meanings, the ones with no common stub to fall back on, are exactly where the reader reversed is weakest (4 of 100) and where the separate inverse holds its ground (28 of 100). This is the DIVA win axis: a separate inverse trained on its own feedback reaches the novel target that running the recognizer backwards cannot.

The lesson

Production is a separate organ. On the matched conditional task, produce one specific intended meaning, a separate inverse trained feedback-first recovers 82 of 300 where the reader run backwards recovers 2. The inverse learns it over training (recovery 0.060 to 0.417), and it wins in every frequency band including the rare tail (28 of 100 at the rarest, against 4). The reader reversed is not a producer: its top-down wire carries credit and priming, not generation, so prompted recall returns generic stubs. The wall the last round drew has its named fix: a second module, trained by the birdsong loop, not the recognizer pointed backwards.

This is the shape the DIVA model predicted and the last round's wall asked for. The reader is a recognizer, and a recognizer reversed plays back the most common continuation, not the word you meant. A producer is a different map, from a meaning to a sequence, and it has to be learned on its own feedback. The self-feedback loop is the trainer, and the meaning's region, not its exact spelling, is the target. Inner speech and the mirror system say the same thing from two more directions: the producer can run internally, and the code it shares with hearing is learned. Here that map is built and measured, and it reaches the meaning where the reader reversed sits at 2 of 300.

The honest caveats

• The matched comparison is the conditional task, not a head-to-head on the old number. The 64 of 500 from the last round was unconditional free generation, a different task on a different denominator (here free generation reads 45 of 500). The claim is the conditional gap: 82 against 2 when both are asked for a specific meaning. This does not beat the free-run number head-on, and is not claimed to.
• Recovery is moderate, not high. The judge recovers the intended meaning 0.417 of the time. That is far above the reader reversed and far above chance, and it is a first cut, not fluent production.
• A controlled vocabulary and a coarse meaning code. Five hundred known words, and the meaning is a salience-weighted heard fingerprint, the precision-weighted region, not a rich semantics. The producer cannot string-match its way to the answer, but the setup is a controlled instrument, not open text.
• Still the simplifications of the self-feedback round. A count model for the re-hearer and a chunk-lexicon babbler for the producer. The mechanism is real and gradient-free; the substrate is deliberately simple.

The frontier

The wall now has a working fix on its conditional half. A separate inverse, trained by self-hearing, produces a specific meaning where the recognizer reversed cannot, and it does so across the frequency range, including the rare targets the neuroscience flagged. What it does not yet have is a shared convention. By design the loop only ever agrees with itself, so it teaches an inverse that one agent can read back, not a code two agents share. That is the external half of generation, the referential game, where a listener has to recover the meaning and the convention is anchored against private drift. And the recovery number, 0.417 on a controlled vocabulary, is the next thing to lift: a richer comprehension monitor and a less coarse meaning region are the named moves. The separate organ exists and works. Giving it a listener, and a sharper ear, is the next run.

Lineage

Grew from what it learns and what it can say back, which drew the wall this closes (a strong recognizer, a weak producer, 2 of 296 played back to the exact spelling); and from internal self-feedback teaches production, the birdsong loop that trains the inverse here, the first positive on the internal half of generation.

Thread: generation, and the comprehension-production gap. The mechanism is a separate learned inverse trained feedback-first (the DIVA model), made concrete on the self-feedback loop. The reader is the recognizer; the inverse is the producer; they are different maps. The frontier is the external half: a referential game that turns a private inverse into a shared convention.