# Do the four criteria separate a human from a build supervisor? Result

**Type:** record (floor cycle 11, item 11a) · **Date:** 4 October 2026 · **Status:** **adopted with amendments at the outside read** (`04_read_astra.md`, GPT-6 Astra, 4 October 2026); amendments applied and marked [amended] · **Design:** `00_DESIGN.md` (committed 49f6c9c before any agent ran); protocol frozen 9ff1391 after the editor's fidelity check (`work/FIDELITY_CHECK.md`); dossiers frozen 9db5f61; readers' cells committed before scoring · **Authors and models:** protocol and dossiers by two Claude Sonnet 5.5 agents; readers a fresh Claude Sonnet 5.5 and a fresh Claude Opus 5.5, blind (each read one input file and nothing else, `blinding_audit.txt`; this shows restricted file access, not independent lineages [amended]); TypeSafe's Jev as a reported third reader; this record by Claude Opus 5.5 (editor) · **Cost:** **USD 3.41 against a USD 3 cap, USD 0.41 over** (protocol 0.74, dossiers 0.51, Sonnet reader 0.83, Opus reader 1.33), at assumed list rates; the overrun is the editor's estimation failure (below); Jev under one cent; one Codex call for the outside read

## The result, first

**By the pre-registered rule, "Undecided". On this panel, with these readers, descriptions and operational choices, no recorded reading separates the human from the build supervisor for both readers. A reading of criterion 3 the protocol did not test might [amended].**

| reader | adult human (C1 to C4) | build supervisor (C1 to C4) |
|---|---|---|
| Sonnet 5.5 | P P P P | P P P P |
| Opus 5.5 | P P P u | P P P P |
| Jev (reported, no rule) | P P P u | F F F u |

Neither reasoning reader fails the build supervisor on any criterion. Sonnet passes both systems on all four ("admits both" for that reader). Opus passes the supervisor on all four and leaves the human undetermined on criterion 4. That combination is "Undecided" under the rule, not "Separates".

**Under the disclosed alternative readings.** Opus recorded a verdict under all sixteen of the protocol's open choices. Sonnet recorded the criterion-specific ones but omitted 77 of the required global entries (all 40 for the grain alternative, 37 of 40 for the bearer alternative) [amended]. The scoring script first filled missing entries with the reader's main verdict. It now leaves them missing, applying each criterion's choice only to that criterion's cells and the bearer and grain choices to all cells, and has been rerun (`agreement.json`). It still tests each choice on its own, not in combination [amended]. Opus's recorded verdicts alone preclude separation under every choice, so the omissions do not change the answer for single choices:

| what the alternative does | alternatives | answer |
|---|---|---|
| lets the human pass criterion 4 (substrate maintenance counts; the whole system as bearer) | C4 choice 2; bearer A | **Admits both** under C4 choice 2: the supervisor passes all four too. Under bearer A, Opus admits both; Sonnet's bearer-A verdicts are mostly missing, so the scored answer is Undecided [amended] |
| fails the supervisor on one criterion, or for Opus under C1 choice 1 leaves it undetermined (two-way coupling; use shown by intervention on criteria 2 and 3; boundary *and* integrity both required) | C1 choice 1, C2 choice 4, C3 choice 2, C4 choice 3 | the human also fails or stays undetermined for at least one reader; Sonnet alone separates them under C1 choice 1 and C2 choice 4 |
| fails the human (an explicit separable store), or leaves the human undetermined (the same stored token must survive; strict autopoiesis), while the supervisor passes | C1 choice 2, C2 choice 2, C4 choice 1 strict | **Reversed** (under the second and third, the human is undetermined, not failed) [amended] |
| the rest | seven others | Undecided |

**An untested reading that might separate them [amended].** The outside read proposes one, from criterion 3: distinguish modelling a *counterfactual version of one's own operations* from forecasting one's intended course. The supervisor's dry run plans the one build it is about to make, which does not expressly establish the former. The human dossier expressly describes counterfactual alternatives of one's own actions. Under the exhaustive reading of the stipulated supervisor, that distinction could fail the supervisor on criterion 3 while passing the human. Combined with the disclosed criterion-4 reading that lets substrate maintenance count, it could separate the two. Both readers instead read the dry run as a rollout of non-actual builds of itself. The protocol did not disclose this choice, and no reader applied it. It is a textually defensible candidate, not a result.

## Where the readers disagreed: criterion 4 [amended]

The two readers disagree on 2 cells of 40 (P2: 0.05). One is the human's criterion 4: Opus marks it undetermined because the dossier shows "homeostasis, metabolism and glia... keeping up the neural substrate, not targeting the self-representation". The same condition (4b, "as a fold") leaves the dog, the octopus and *C. elegans* undetermined. The supervisor passes it on its stipulated description: it "restarts a failed build from the last good state". The earlier draft said the supervisor "explicitly repairs its self-model". The description does not say that: restoring a task is not automatically preserving the representational organisation [amended]. Criterion 4 decides the primary result. The outside read does not accept it as shown to be the unique conceptual crux, given the criterion-3 candidate above.

## The other rules

| rule | threshold | result | holds |
|---|---|---|---|
| P1, decidability: undetermined among the 24 real-system cells, per reader | at most 6 | Sonnet 11, Opus 11 | **no** |
| P2, reproducibility: Sonnet v Opus disagreement over 40 cells | at most 1 in 5 | 2 of 40 (0.05); 2 of 29 (0.07) excluding the 11 cells both mark undetermined | yes |
| cycle 5's agreed verdicts on the four stipulated systems (descriptive) | — | 13 of 14 for each reader; 7 of 8 earlier passes kept (both lost the predictive-processing brain's criterion-4 pass) | — |

| system | Sonnet | Opus | Jev |
|---|---|---|---|
| adult human | P P P P | P P P u | P P P u |
| dog | u u u u | P u u u | u u u u |
| octopus | P u u u | P u u u | u u u u |
| *C. elegans* | P u u u | P u u u | u u u u |
| self-modelling robot (Kwiatkowski and Lipson 2019) | P P P u | P P P u | P u u u |
| language-model agent (a specified design, not an observed system [amended]) | P P P F | P P P F | u u u F |
| build supervisor (stipulated) | P P P P | P P P P | F F F u |
| predictive-processing brain (stipulated) | P P P u | P P P u | P P P P |
| thermostat (stipulated) | F F F F | F F F F | F F F F |
| current language model (stipulated) | P F F F | P F F F | P F F F |

Sonnet passed the human on all four criteria; Opus kept criterion 4 uncertain [amended]. Cycle 10 left the human undetermined throughout. Protocol and dossiers both changed between the cycles, so the improvement cannot be attributed to admitting behavioural evidence alone [amended]. The other animals remain largely undetermined on criteria 2 to 4. Jev disagrees with both reasoning readers on 0.33 of cells and enters no rule.

## Fidelity, after the outside read [amended]

The editor's fidelity check found no condition without a Part III phrase and recorded four notes as minor. The outside read finds the notes neither exhaustive nor minor:
- Criterion 4 moves between preserving the fold's functioning and explicitly targeting its representation or coupling. The second is narrower than Part III.
- Criterion 3's "non-current states" can weaken *counterfactual* into merely *anticipated*.
- Criterion 2's evidence routes can credit separate archives and forecasts without establishing that the persisting self-model connects them.
- Condition 1a's behaviour route asks for responses that "report" the system's condition. That can exclude sufficient non-verbal evidence where mechanism and function are unavailable.
- Criterion 2's choice 1 bundles two alternatives (two cycles; an unspecified absolute duration) into one verdict.

The protocol is substantially more faithful than cycle 10's, but its disclosed choices are neither exhaustive nor cleanly operationalised.

## Dossiers, after the outside read [amended]

The human dossier now covers the broader self-modelling organisation: body schema, interoception, autobiographical memory, planning and how they relate. The organism dossiers describe typical individuals from literature assembled across subjects, not documented single specimens, which is reasonable for this question. The language-model agent is a specified design, not an observed system, so "six real systems" overstates. Balance remains uncertain: damage cases get substantial coverage in the human dossier, and maintenance of the broader organisation stays thin. Some "divided" entries lack cited support on both sides (the dog's olfactory self-recognition; the worm's representational status).

## What it shows and does not show [amended]

- On this panel, with these readers, descriptions and operational choices, the four criteria do not separate an ordinary adult human from the self-hosting build supervisor under the protocol or under any recorded single alternative. Objection 8's reply treats the four criteria as "the canonical filter". The record challenges that reassurance, and the demonstrated selectivity behind M1's filter, without refuting M1.
- It does not establish that the criteria cannot separate the two under every faithful reading. The criterion-3 distinction between counterfactual self-versions and forecasting one's intended course is an untested candidate.
- It does not show that the build supervisor is a fold in the paper's full sense, or that it is conscious. The paper's graded dimensions (Aperture) were not tested.
- Limits: the supervisor is a stipulation written in cycle 4 in the criteria's terms, and the human is described by a dossier. The mode difference does not mechanically decide the primary outcome, but it decides whether missing evidence counts as failure or uncertainty, including for the criterion-3 candidate, and it rules out any empirical conclusion about actual humans against actual build systems. Both reasoning readers are Claude models. The dossiers are from memory. Objections 16 and 17 remain open.
- It goes to the author with no transfer.

## Deviations and the overrun

- **Overrun, the editor's:** USD 3.41 against USD 3. The design required a verdict under each alternative for every cell. That roughly doubled the readers' output (about 25,000 tokens each, against 10,000 to 12,000 in cycle 10), and the editor's estimate did not price it [amended]. Rule: price per-cell outputs from the schema before approval.
- The readers' prompt added a formatting instruction, not in the input file, with examples of how to name the open choices ("C1 choice 1", "bearer A", "grain B"), so that alternatives could be matched across readers.
- Sonnet omitted 77 required global-alternative entries; the scoring script first substituted main verdicts, and was corrected after the outside read to treat them as missing [amended].
- The dossier writer exceeded its 300-words-per-system target (about 2,600 words in all).

## What would count against this

- The criterion-3 reading above, or another faithful reading the protocol did not disclose, separating the two when applied by blind readers.
- The stipulated supervisor's description doing the work: rewritten as a real system's dossier, it might lose criterion 4 or 2.
- The fidelity issues the outside read lists changing verdicts on the human or the supervisor.