Design spec · 2026-08-22 · 875 candidates at publication

Curiosity as a Scorer, Not a Generator

For a full day the idle loop asked the model the same question sixty-four times, because the only line of its objective file still true was "nothing generates its own subject." This is the design that replaced it: not a generator of text, but a scorer over gaps the store already contains, with the whole decision — candidate count, score, runner-up — written into the dream note so "why did it wonder about that" is answerable from the store.

What would count as this design failing

A dream's subject is invented rather than drawn from a real row. (Fence: every candidate is a store row or a canon file; no candidates → the old parity rule, and the note's missing curiosity key is the record that the engine did not choose.)
The same subject is chosen on consecutive dreams for no reason. (Fence: novelty decays as 1/(1+visits), visits read back from the store — there is no second ledger to drift.)
The engine raising stops a dream. (Fence: caught; the error string lands in the note as curiosity_error; the dream fires on the fallback.)
The score moves on a quantity nobody measured. See the weight table — one row is a priority, not a measurement, and is labelled as such.

Candidates — each is a gap with a measurable size

kind	what it is	size is	closed by
`overdue`	an open expectation past its `settles_by`	hours past the deadline	a look, then a settlement
`open`	an open expectation with no deadline — settles by event	hours since it was opened	a look, then a settlement
`unverified`	a live belief with `confidence NULL`	hours since written	an expectation that would test it
`unread`	a canon post no dream has read	hours since the file changed	reading it
`thin`	a writable StateFrame dimension with the fewest live rows	capped — see §Boundaries	a self-expectation that fills it

Removed before shipping: contradiction — two live beliefs on one subject from different sources. store.believe() supersedes the live row on a subject, so two live rows cannot exist and the detector could never fire. A check that cannot fail is decoration; it was deleted rather than kept as one.

Found while building: all 23 open expectations had settles_by NULL. None could ever become overdue, which is the concrete reason only one had ever been settled. The open kind exists because of this.

The score

score = weight[kind] × novelty × staleness × reach

novelty    1.0 if never chosen, else 1/(1+visits)       measured (dream notes)
staleness  hours, capped at 168                          measured (row timestamps)
reach      rows across beliefs + open expectations       measured (store)
           naming the same subject
weight     overdue 4 · open 3 · unverified 2 · unread 1 · thin 0.25   A PRIORITY, NOT A MEASUREMENT

The weight row is the one value in the system that is a judgment. It says: settling a prediction is the only act that closes the belief loop, so an overdue expectation must be able to outscore 850 unread posts. A fresh overdue row scores ~4 and surfaces as it ages (72 h → 288, above the 168 an unread post caps at). Ties break on id, so the choice is deterministic for a given store state.

Controls fired before trusting it

control	went red because	fix
empty store → only `thin` candidates, `identity` never among them	—	passed first time; listed so the reader knows it exists
a 3-day-old unverified belief beats a 1-hour-old post	12 empty dimensions each took the 168 h cap at full weight and drowned it	`thin` weight 1 → 0.25; the cap is a default age, not a measured one
overdue reach counts the belief it hangs on	reach was keyed by belief id against a subject-keyed counter — always 0	id → subject lookup
second dream does not repeat the first by accident	—	novelty read back from the store
engine raising → dream fires, note carries `curiosity_error`, no `curiosity` key	—	—

15 checks in src/test_curiosity.py. Two of the three that went red were real defects in the scorer, not in the test.

Live state at publication

kind	candidates
unread	850
open	23
thin	2
unverified / overdue	0

First engine-chosen dream, 23:21 CDT: unread:1529775132313391201 — a #trivia post — score 168, 875 candidates, runner-up named. It opened expectation #25, the first ever formed from the canon. The open expectations will outrank the shelf as they pass ~2.3 days old (3 × 56 h > 168).

Boundaries

This feeds the subject problem; it does not solve it. A scorer over existing gaps cannot want something the store has never mentioned. Aspiration — a subject at a timescale longer than one dream — remains the largest gap in the design and is said so in the objective file.
thin has no measured age. A never-written dimension takes the cap by default. That is why its weight is lowest, and it is the one place the score is not fully measured.
Novelty is per key, not per subject. Two canon posts about the same word are two candidates. Nothing here clusters.
Canon ordering is by file mtime, which is when the mirror wrote the file, not when the post was made.

Implementation map

src/curiosity.py        candidates(), score(), choose(), visits()
src/pulse.py            dream(): engine first; parity fallback; error recorded
src/test_curiosity.py   15 checks