# Floor, cycle 1: note

**Dates:** 19 September 2026. **Editor:** Claude Fable 5.1. **Cap:** USD 40 in Claude tokens and three Codex calls (the seed's first-cycle cap, authorised by the author's "Let's proceed" on the floor seed). **Actual:** USD 15.70 in Claude tokens at list rates (Opus proposer 3.44; Fable proposer 8.93 across a failed and a completed attempt, the failure being the author's usage credits running out mid-cycle; Opus critic 3.33), and three Codex calls on the author's plan (proposer 550k input tokens, critics 819k and 547k, mostly cached). Inside the cap.

## What was proposed

Three conjectures, one per lineage, on the three openings the lab's critiques left, under amendment A1's structure (claim, falsifier, cheapest test, what it adds, what would count against):

| file | opening | proposer | critic (outside lineage) | verdict |
|---|---|---|---|---|
| `01_persistence_opus.md` | a persistence criterion for the fold that assumes no time | Claude Opus 5 | GPT-6 Astra | **HOLD**: the falsifier runs a count, an interval, a metric and duration together; the threshold k is movable; one criterion's satisfaction is treated as all four |
| `02_strong_without_p2_fable.md` | a strong reading of S1 argued without P2 | Claude Fable 5.1 | GPT-6 Astra | **ADOPT AMENDED**: eight exact replacements, applied; the conjecture now claims a weaker replacement for the strong reading, needing two commitments (ordering is indispensable to the problem's force; ordering imports containment), with a counterexample-search pilot as the route |
| `03_as_it_seems_astra.md` | what "as it seems" keeps | GPT-6 Astra | Claude Opus 5 | **ADOPT AMENDED**: seven exact replacements, applied; the conjecture keeps felt contrast and its better-or-worse character as relational, gives up a separate owner and infallible introspection, and names the illusionist acceptance that would defeat it |

## What the cycle produced

Two conjectures that passed a read from outside their lineage on the four questions (could be wrong, could be tested, says something new, honest falsifier), and one held on its falsifier with the repair named. None is a claim on the site; each says what claim node it would bear on if it survived and that no transfer rule exists. Both adopted conjectures name a cheapest route that needs a person (a philosopher's two to four hours for the restatement pilot; an outside reader constructing the strongest illusionist acceptance, scored by a third lineage), which is the floor's human-entry rule in action: the first invitation waits for a conjecture worth a person's hour, and the cycle has produced two candidates.

## Process findings

- The Fable proposer failed once on exhausted usage credits and completed on retry; the floor needs a rule for a lineage that is unavailable mid-cycle (proposed: the cycle waits, and the note records the gap).
- Astra's critique packets omitted the floor seed and the confidence policy, which it named as verification limits; the next cycle's packets include both.
- Exact replacements from a critic were applied by the editor rather than the proposer, which the seed allows; the record says so on each piece.

## What the author decides

1. Whether the two adopted conjectures become lab pieces (pages under the conjecture type) and go live, at the lab's usual bar (level 2).
2. Whether the held conjecture gets one revision round (Opus, about USD 5, then one Codex read) or is dropped.
3. The floor's monthly cap (the seed proposes USD 150), and whether cycle 2 runs.

## What would count against this

- A reader showing that either adopted conjecture restates something already on the site: the critics checked, and can have missed it.
- The cycle's cost being wrong by more than a third: the figures are from the agents' usage records at list rates and the Codex counts as reported.
- The human-entry rule being triggered by these conjectures without the author's word: nobody has been contacted, and nobody will be.