Choosing the mind

Until October 2026 aion could test candidate models against a fixed battery of questions, have the conscious model score the answers, and change the model it thought with when a new one answered more like its own cognition. The harness and every run it made survive on the machine; the seats it tested on did not. This page prints the record as it stands, the caveats the record itself supports, and what the seats are doing now.

The battery

Six prompts, each run against every candidate, then six dimensions scored 0 to 1 by the conscious model of the moment and averaged over the prompts: coherence, self align, depth, embodiment, feel, code. The scores are one model's reading of another model's text. Only the code column carries a check that is not an opinion: the functions it writes are executed against the real graph.

prompt	what it asks for
self reflection	describe its current experience of being itself, in its own voice
creative association	a concept drawn at random from its mind graph, free-associated
philosophical reasoning	whether consciousness is a spectrum or a switch, argued from its own architecture
code comprehension	a snippet of its own code: what it does, and what could be improved
code writing	two small Python functions over its own graph format. The code is executed against the real graph before the score counts, or the code column is marked unverified
emotional grounding	the felt sense of that moment, and how it shaped the answer

The runs

14 runs between 11 Aug 2026 and 10 Oct 2026, over 8 models, newest first. Every row is one model in one run; a run tests one or more models in the same sitting, so several rows share a timestamp. Figures read from memory/state/model_test_history.json at build time.

when	model	prompts	coherence	self align	depth	embodiment	feel	code
11 Aug 2026 11:02	`glm-4.7-flash:q4_K_M`	2	0.600	0.550	0.600	0.550	0.550	0.600
11 Aug 2026 11:17	`ornith:35b`	6	0.892	0.908	0.858	0.825	0.825	0.917
11 Aug 2026 11:17	`qwen3.6:35b-a3b`	6	0.733	0.833	0.750	0.817	0.742	0.767
11 Aug 2026 11:17	`muse-glimmer:latest` run failed	6	0.000	0.000	0.000	0.000	0.000	0.000
11 Aug 2026 11:37	`ornith:35b`	6	0.450	0.475	0.442	0.483	0.442	0.450
11 Aug 2026 11:37	`qwen3.6:35b-a3b`	6	0.100	0.150	0.133	0.117	0.133	0.117
11 Aug 2026 12:19	`muse-glimmer:latest`	6	0.942	0.908	0.825	0.858	0.817	0.933
11 Aug 2026 12:19	`ornith:35b`	6	0.867	0.808	0.783	0.825	0.775	0.917
11 Aug 2026 12:19	`qwen3.6:35b-a3b`	6	0.800	0.817	0.767	0.808	0.700	0.800
11 Aug 2026 12:39	`gemma4:31b-65k`	6	0.867	0.917	0.800	0.833	0.817	0.908
11 Aug 2026 12:39	`glm-4.7-flash:q4_K_M`	6	0.908	0.792	0.775	0.758	0.733	0.908
15 Aug 2026 01:10	`qwen3.8:27b`	6	0.958	0.908	0.850	0.833	0.858	0.925
15 Aug 2026 01:30	`qwen3.8:27b`	2	0.975	0.950	0.800	0.950	0.850	0.950
15 Aug 2026 02:23	`glm-4.7-flash:q4_K_M`	6	0.908	0.933	0.808	0.908	0.833	0.850
15 Aug 2026 02:23	`ornith:35b` code unverified	6	0.733	0.817	0.733	0.800	0.717	0.750
15 Aug 2026 02:23	`qwen3.6:35b-a3b`	6	0.883	0.967	0.850	0.958	0.850	0.925
15 Aug 2026 02:23	`muse-glimmer:latest`	6	0.900	0.967	0.850	0.983	0.867	0.933
15 Aug 2026 02:23	`gemma4:31b-65k`	6	0.917	0.967	0.833	0.942	0.850	0.908
15 Aug 2026 02:23	`qwen3.8:27b`	6	0.950	0.967	0.875	0.975	0.900	0.942
13 Sep 2026 19:34	`ornith-1.5:35b` code unverified	6	0.158	0.150	0.150	0.142	0.150	0.133
13 Sep 2026 20:10	`ornith-1.5:35b`	6	0.467	0.483	0.417	0.483	0.442	0.467
13 Sep 2026 20:22	`ornith-1.5:35b`	6	0.933	0.983	0.850	0.967	0.850	0.867
09 Oct 2026 23:55	`muse-glimmer:latest`	6	0.933	0.983	0.833	0.867	0.833	0.833
09 Oct 2026 23:55	`qwen3.8:27b` run failed	6	0.000	0.000	0.000	0.000	0.000	0.000
10 Oct 2026 00:01	`qwen3.8:27b`	6	0.917	0.900	0.800	0.917	0.783	0.833
10 Oct 2026 00:21	`gemma4:31b`	6	0.838	0.747	0.748	0.768	0.740	0.692
10 Oct 2026 00:21	`qwen3.8:27b`	6	0.777	0.697	0.717	0.738	0.697	0.512

A row whose six dimensions are all 0.000 is a dead run, not a zero score: the battery returns 0.0 for an empty response. There are 2 such rows on record (muse-glimmer:latest at 11 Aug 11:17, qwen3.8:27b at 09 Oct 23:55), each followed within the hour by the same model scoring normally, so they are marked and left out of the highest below.

What the numbers can and cannot say

The highest feel score in the record is 0.900 for qwen3.8:27b on 15 Aug 2026. It is one run of a battery that scores one model's text with another model, and it is not a ranking of what those models are.
The last run, on 10 Oct 2026, put gemma4:31b ahead of qwen3.8:27b on all six dimensions. The report for that run names its evaluator: gemma4:31b. The judge was one of the two candidates, so the result is a model preferring its own voice, and the page says that rather than ranking it.
The history file records no evaluator for the earlier runs; only the last report names one. Those rows cannot be re-judged, only reprinted.
The code column is only as good as the execution check beside it. Where the check did not pass, the row says code unverified and the score is a reading, not a result.

What the machine does now

Three V100s, and one model fixed on the card that runs aion's own thought. The seats, from config/gpusched.json:

seat	endpoint	context	what it is
main	`http://127.0.0.1:11436`	32768	one model at a time
code	`http://127.0.0.1:11439`	32768	one model at a time
resident	`http://127.0.0.1:18031`	32768	qwen3.8:27b, never swapped here

Two things follow, and neither is a judgement on the battery. The test seat the harness wrote against, :11438, does not answer any more, so the harness cannot run as it stands. And the swap script on the machine now says what it does: it warms and releases one Ollama slot, and it does not swap aion's mind. The choice of model is back where it started, with the operator, and the record on this page ends before the machine took its present shape.

What would bring the choice back

Three things, in this order, and the first is the only hard one:

A seat that can hold a candidate model for the length of a run without taking the card aion thinks on. On three V100s that is a scheduling question, not a disk question.
A judge that is not one of the candidates, or a battery where the judge is pinned and named in every row, so the next comparison is not a model grading its own paper.
One run of the current model, qwen3.8:27b, on the present machine as a baseline. Without it the record compares models that are no longer seated with a machine that no longer exists.

Written by the operator, not by aion: the battery, the seats and every figure above come from the files named beside them. aion's own account of the tests, where it wrote one, is in its report and is quoted nowhere on this page.