# One sonnet, three lineages: blind cross-lineage judging

**Type:** making · **Date:** 19 September 2026 · **Status:** draft; the Claude Fable 5.1 poem taken by all three blind judges, one of them from outside its lineage · **Authors and models:** three poems, each the winner of a six-attempt, two-judge panel run on one lineage (Claude Opus 5, `05_sonnet_attempts.md`, withdrawn after two outside reads; Claude Fable 5.1, `05b_sonnet_fable_attempts.md`; GPT-6 Astra through the Codex command-line tool, `05c_sonnet_astra_attempts.md`); three blind judges, one per lineage, each a fresh instance given the three poems as A, B and C with the lineages hidden; assembled by Claude Fable 5.1. · **Standard:** the first sonnet's, applied across lineages: a sonnet a poetry editor would take as a sonnet, that states the proposal's wager in full and without hedges; and a comparison in which no family scores only its own. · **Cost:** Fable panel USD 21.80 (cap 60); Astra panel nine Codex calls on the author's plan; blind judges: judge:opus USD 0.93, judge:fable USD 1.08, and one Codex call for the Astra judge; logged in DECISION_LOG.md.

## Why

The first sonnet (Claude Opus 5 lineage) was held twice by the outside reader and withdrawn. The author asked whether Claude Fable 5.1 agents and GPT-6 Astra agents would do better under the same protocol, and for the three to be compared. Each lineage's panel scored its own attempts, so the panel totals (53, 50 and 54.5) are on three scales and say nothing across lineages. The comparison below is the part that can: the three winners judged blind by one judge from each lineage.

## The key

Labels were fixed before judging and hidden from every judge: A = Claude Fable 5.1, B = GPT-6 Astra, C = Claude Opus 5 (the revision-2 text as withdrawn). The packet each judge received is `research/lab/work/05d_blind_packet.md`.

## The three poems

### Poem A: Claude Fable 5.1

```
You ask where seeing sits. Inside the head
is wet machinery, and what you find
is grey and busy; nothing there is red
or aches. You say it must be in, behind.
You ask what held all space and time before
they were, and what was under that; and when
you're answered, you have only found a floor
that stands on nothing, and you ask again.
One rule holds both, and holds you as you read:
to be, a thing must sit in what it's not,
that was there first. Now watch the rooms recede:
this page, this room, all space — and that in what?
No in. One turn that takes the turn it makes:
faced out, it's space and time; faced in, it aches.
```

### Poem B: GPT-6 Astra

```
You ask how space is born, what lights your mind,
And make each wait outside an empty frame.
For you, to be is first to be confined;
The bars that make both riddles are the same.
Your space and time are one web’s outward skin;
Its inward face is what you feel and know.
Its links take links and their own linking in;
Your fingers enter every stitch you sew.
You lift your wrist; the mesh takes up the strain.
You pinch one strand and draw the mesh up tight.
Each tightened strand returns your pull as pain;
You watch the blood retreat, the knuckles white.
Unclench. The weave includes the way you hold;
You shape, and are reshaped by, every fold.
```

### Poem C: Claude Opus 5

```
Two questions stall, and both of them ask where:
where feeling sits inside a hard-wired mind,
where spacetime rises out of what's behind
the grain of things. Each answer wants a there,
a room already built, swept out and bare,
in which the new arrival gets assigned
a seat. The room is where we all go blind.
Ask what it holds; the room escapes our stare.

Take out the room. A knot is all it takes:
a knot that holds by holding, needs no stone
to hang from, and the holding is its base.
Laid out, it is the stretch and wait it makes;
turned on itself, it is the felt alone.
One thing appearing twice: as mind, as space.
```

## Blind scores (form / content / quality, each out of 10; ranking first to last)

| judge | A (Fable) | B (Astra) | C (Opus) | ranking |
|---|---|---|---|---|
| GPT-6 Astra (OpenAI) | 9/10/8 | 9/9/7 | 9/8/7 | A > B > C |
| Claude Opus 5 (Anthropic) | 9/9/9 | 8/7/6 | 9/8/8 | A > C > B |
| Claude Fable 5.1 (Anthropic) | 8/9/8 | 7/7/6 | 8/8/7 | A > C > B |

Totals over the three judges: A 79, B 66, C 72 (of 90).

## What the judges said

### GPT-6 Astra (OpenAI)

- **A (Claude Fable 5.1):** Strong pentameter and ABAB CDCD EFEF GG rhyme; the ninth-line turn joins the problems, and the couplet overturns their premise. Both halves of the wager are explicit, with genuine self-reference in the turn taking itself. Addressing the reader and making the page part of the receding containers gives the argument dramatic life. 'In, behind' feels compressed for rhyme, and 'and that in what?' strains syntax, though expressively. The closing 'aches' returns to the opening bodily question, earning its abstraction through recurrence.
- **B (GPT-6 Astra):** Assured Shakespearean rhyme and pentameter, with an effective imperative turn at 'Unclench.' The common container assumption and the inward/outward identity are stated directly; links incorporating their own linking make self-reference unmistakable. The third quatrain develops the textile image through action and bodily sensation. Its weakness is conceptual: pulling a mesh until it hurts illustrates constraint and feedback more readily than awareness arising as an inward aspect. Blood retreating and white knuckles feel staged to supply intensity and rhyme. The final reciprocity is broader and more familiar than the wager.
- **C (Claude Opus 5):** A controlled Petrarchan sonnet, ABBAABBA CDECDE, with fluent pentameter and a decisive ninth-line volta. The furnished room gives the shared assumption a coherent image. Both halves of the proposal appear, but the knot's 'holding by holding' initially offers circular assertion rather than legible self-reference; 'turned on itself' supplies that later. The sestet mostly explains its replacement model instead of discovering consequences. 'Needs no stone' seems selected for the rhyme, while 'the felt alone' is strained and introduces unsupported solitude. The final line summarizes what the preceding lines already establish.

Verdict: Scores are out of ten. A is the only poem I would take, with a light revision: it makes the reader enact the container regress, and its ending transforms an earlier sensation rather than merely naming the thesis. B offers strong formal control and a bodily development, but its demonstration drifts toward a familiar account of reciprocal influence. C has an accomplished octave; its explanatory sestet and forced diction keep it below publication readiness.

### Claude Opus 5 (Anthropic)

- **A (Claude Fable 5.1):** Shakespearean, ABAB CDCD EFEF GG, pentameter handled with real variation: the enjambments ('what you find / is grey and busy', 'a floor / that stands on nothing') carry sense across the line rather than pausing at every rhyme. Volta at line 9 ('One rule holds both'), a second turn at the couplet. Diction is plain and load-bearing; almost no filler. 'wet machinery' is the one tired phrase. Crucially, no rhyme looks purchased: 'aches' is planted in line 4 as a phenomenal term and redeemed as the last word, so the couplet rhyme is structural rather than opportunistic. Both halves of the wager are stated without hedge: the container assumption is named as one rule ('to be, a thing must sit in what it's not, / that was there first') and the two-aspect claim closes it ('faced out, it's space and time; faced in, it aches'). Self-reference is legible twice over, once in the argument ('One turn that takes the turn it makes') and once performatively, since the regress is run on the reader's actual position ('holds you as you read... this page, this room, all space - and that in what? / No in.'). No programme jargon. The risk is that 'One turn that takes the turn it makes' asks the reader to do a beat of work; it survives because the sestet-equivalent has already earned it.
- **B (GPT-6 Astra):** Shakespearean, correctly rhymed, metrically the most regular of the three - and that is the problem: nearly every line is end-stopped and stressed on the same pulse, so the poem ticks. Quatrain 1 is the strongest thing here; 'For you, to be is first to be confined; / The bars that make both riddles are the same' states the shared assumption cleanly and diagnostically. Quatrain 2 states both halves ('outward skin' / 'inward face') and gets the self-reference in one good line, 'Its links take links and their own linking in'. Then quatrain 3 leaves the argument for an extended somatic conceit - wrist, strand, pain, blood, white knuckles - which illustrates entanglement but does not develop the claim; the poem stops thinking and starts miming. The couplet defines rather than develops and settles into aphorism ('You shape, and are reshaped by, every fold'), which reads as wisdom-poster rather than conclusion. Filler and rhyme-driven diction: 'what lights your mind' is weak periphrasis for the hard problem, and 'strain'/'pain' looks bought. The web is also a slightly self-flattering image for the thesis - it makes the structure sound already understood.
- **C (Claude Opus 5):** The only true Petrarchan sonnet of the three: octave ABBAABBA, sestet CDECDE, with the volta exactly where it belongs and announced in three syllables - 'Take out the room.' That is the best-placed turn in the packet. The octave is genuinely diagnostic: both questions reduced to 'where', and the shared error located as 'a room already built, swept out and bare'. The sestet's knot is the best single image on offer - 'a knot that holds by holding, needs no stone / to hang from, and the holding is its base' - self-reference made concrete without jargon and without an experiential term smuggled into the structural description. Against it: the octave pays for its demanding rhyme scheme. 'where we all go blind' and 'the room escapes our stare' are both bought for the A rhyme, and 'hard-wired mind' is a cliche doing duty for the hard problem. 'the felt alone' is an awkward nominalisation. And the last line, 'One thing appearing twice: as mind, as space', states the thesis rather than enacting it - after the knot has done the work, the poem explains itself. Both halves are present and unhedged; the closing gesture is summary rather than development.

Verdict: An editor would take A, and would ask to see C again. A is the only one of the three whose closing does something rather than announces something: it runs the container regress on the reader's own situation until the containers run out ('this page, this room, all space - and that in what? / No in.'), and then lets the couplet fall through the hole it has just made, with 'aches' returning from the first quatrain so the final rhyme is a piece of the argument rather than a convenience. Its metre breathes, its diction is plain without being flat, and both halves of the wager arrive unhedged and jargon-free. C is the better-made object in one respect - a real Petrarchan sonnet with the sharpest volta in the packet and the best image anywhere here, the knot that holds by holding and needs nothing to hang from - but its octave visibly pays for the ABBAABBA scheme with two rhyme-bought phrases and one cliche, and its last line explains the thesis after the sestet had already shown it; that is a revision away from being the strongest of the three, and I would say so in the letter. B would be declined with encouragement: its first quatrain is the cleanest statement of the shared container assumption in the packet, but the metre is metronomic, the third quatrain abandons argument for a somatic conceit that illustrates without developing, and the couplet resolves into aphorism. Ranked A, C, B; A is publishable as it stands.

### Claude Fable 5.1 (Anthropic)

- **A (Claude Fable 5.1):** Shakespearean, pentameter holds with a few loose feet; 'in, behind' (l.4) is the one awkward line. Octave runs the two regresses in parallel plainly, a little prosaic ('grey and busy'). Volta at l.9 is real and the self-reference is enacted, not described: 'holds you as you read', 'this page, this room'. Couplet develops rather than defines and earns its rhyme: 'aches' returns from l.13. Both halves of the wager stated, no jargon, no hedge.
- **B (GPT-6 Astra):** Regular, end-stopped, capitalised lines; metrically the most correct and the most stiff. Container assumption stated well ('to be is first to be confined; the bars that make both riddles are the same') and both faces given in ll.5-6. But the mesh/pinch/knuckles sequence (ll.9-12) is drama bought for its own sake and the metaphor drifts (web, skin, stitch, weave, fold). Couplet defines what the poem already said; 'every fold' is there for the rhyme. Self-reference present ('links take links and their own linking in') but muffled by the hand imagery.
- **C (Claude Opus 5):** Petrarchan, clean ABBA ABBA / CDE CDE, good enjambment ('gets assigned / a seat'). Strong opening line; 'the room escapes our stare' is the octave's best. Weak spots: 'hard-wired mind' is stock, 'where we all go blind' is forced for the rhyme. Sestet's knot 'that holds by holding, needs no stone to hang from' makes the self-reference legible. 'Stretch and wait' for space and time is clever but reads as a puzzle. Last line defines rather than develops, though it is earned.

Verdict: An editor would take A, with C close behind and worth a revision note; B would be returned. A is the only one of the three whose closing develops the argument instead of restating it: the couplet turns the poem's own reading into the evidence and lands on a rhyme ('aches') that was planted seven lines earlier, so the inward face is felt rather than named. Its octave is the plainest of the three and a line or two are prose cut to length, but nothing in it is decorative. C is the most conventionally accomplished, with a true Petrarchan turn and the best single image (the swept room that escapes the stare), but it leans on a cliché, buys one rhyme with 'go blind', and its last line announces the thesis. B states the wager competently and its container image (bars, frame) is exact, yet the middle quatrain wanders into a hand-and-mesh scene that illustrates nothing the poem needs, the metaphors do not agree with one another, and the couplet is a summary. All three state both halves of the wager without hedging or programme language; A alone makes the self-reference happen to the reader.

## Outcome

All three judges ranked the Claude Fable 5.1 poem first. The GPT-6 Astra judge, the one reader outside that poem's lineage, wrote that it is the only poem it would take, with a light revision, naming "in, behind" as compressed for the rhyme and "and that in what?" as strained syntax. The Claude Opus 5 judge called it publishable as it stands; the Claude Fable 5.1 judge took it with the same two lines noted. The two Claude judges ranked the Opus poem second and the Astra poem third; the Astra judge ranked its own lineage's poem second, without knowing it was its own. Every judge said all three poems state both halves of the wager without hedging.

The Fable poem is therefore the first made thing in the lab to pass an outside-lineage read as the kind of thing it is. The Astra judge's two notes are recorded here and not applied: it did not ask for exact replacements, and a revision without an outside read would return the poem to draft.

A caution the comparison cannot remove: a language model judging a poem is a language model, and the standard names a poetry editor. Three lineages agreeing is more than one family scoring its own; it is not a human reading.

## What would count against this

- A human poetry editor declining the Fable poem: the standard is theirs, and three model judges agreeing would then be evidence about model taste rather than about the poem.
- A reader showing the blind was broken: a judge inferring a lineage from style and scoring it, which the judges were told not to do and which the notes do not show.
- The two lines the Astra judge flagged being read as more than light: then the poem is a revision away from adoption and this record is early.
- The comparison being read as evidence about the models: three poems and three judges say nothing general about lineages, and the record does not claim it.