# The sharper filter on three more systems: result

**Type:** record (floor cycle 13) · **Date:** 4 October 2026 · **Status:** **adopted with amendments at the outside read** (`02_read_astra.md`, GPT-6 Astra, 4 October 2026); amendments applied and marked [amended] · **Design:** `00_DESIGN.md` (committed 1a47535 before the readers ran) · **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`; Sonnet's structured output was submitted twice, the second accepted) · **Editor:** Claude Opus 5.5 · **Cost:** USD 1.38 of a USD 1.50 cap (Sonnet 0.60, Opus 0.78), at assumed list rates; one Codex call for the outside read

## The result, first

Applied jointly, the two sharper readings (criterion 3 requiring a model of a counterfactual version of one's own operations, not a mere forecast; criterion 4 letting substrate maintenance count) give the same verdict from both readers on all 12 cells:

| system | C1 | C2 | C3 | C4 | status |
|---|---|---|---|---|---|
| language-model planning agent | pass | pass | pass | **fail** | **excluded** |
| self-modelling robot (Kwiatkowski and Lipson 2019) | pass | pass | pass | undetermined | open |
| domestic dog | pass | undetermined | undetermined | pass | open |

- **The planning agent is excluded, on criterion 4, not criterion 3.** Both readers pass its criterion 3 under the sharper reading: its planner compares several alternative plans of its own actions, which they read as modelling counterfactual versions of its operations. It fails criterion 4 because "hosting, storage integrity and compute come from infrastructure outside the loop", and letting substrate maintenance count relaxes condition 4b (what is maintained), not 4a (that the system's own operations maintain it).
- **The robot is open on criterion 4.** The frozen dossier leaves it unclear whether the arm's own operations contribute to its maintenance. The protocol requires an internal contribution to maintenance and expressly allows standing maintenance; it does not require that the robot trigger repair autonomously [amended]. **Source check by the outside read [amended]:** the published paper (Kwiatkowski and Lipson 2019, "Task-agnostic self-modeling machines") attributes damage detection and self-model retraining, with about 10 percent additional data, to the robot. Columbia's account and the authors' lab page say the robot can restart data collection after detecting divergence. That supports robot-led adaptation, but does not document whether the experimental retraining was launched without a human command. A source-informed reassessment of criterion 4 is warranted; the panel result stands as conditioned on the frozen dossier.
- **The dog is open on criteria 2 and 3.** Its body representation guides action (criterion 1), and its homeostasis maintains the substrate its fold runs on (criterion 4 under the reading). The dossier gives no evidence about its own past or future being used, or about alternatives to its own actions. These are gaps in the dossier's evidence, not uncertainty about whether dogs are conscious [amended].

This matches the editor's pre-registered expectations on all three systems.

## Together with cycles 11 and 12 [amended]

Under the two sharper readings, composing verdicts across cycles, the adult human passes all four criteria and the build supervisor fails one (criterion 3; it passes criterion 4). In this cycle, both readers exclude the specified, unbuilt planning-agent design on criterion 4. The robot and the dog are open on the frozen dossiers. The criterion-4 alternative was in cycle 11's frozen protocol before its results. The criterion-3 reading, and the choice to combine the two, came after them.

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

- These two readers exclude this specified planning-agent design on criterion 4, condition 4a: its upkeep is supplied from outside its loop. They do not exclude it on criterion 3, which its comparison of alternative plans of its own actions meets under the sharper reading.
- That is a result about five descriptions, drawn from cycle 11's dossiers, the cycle in which the readings were proposed, and read by two Claude models. It does not validate a general boundary between AI agents and organisms, and it does not show that self-maintenance is where such a boundary lies.
- No transfer; it goes to the author.

## What would count against this

- A system everyone would exclude that passes all four under the two readings, or one everyone would include that fails.
- The robot's open cell resolved by checking the paper: if the arm triggered its own retraining, it would pass criterion 4 under the reading and be admitted on all four, a case the site would then have to address.