01 · Invented relations
Nothing to look up
Names, objects, colors, actions, and places were synthetic so outside knowledge could not supply the answers.
The measured result
Same language source. Different relational structure.Correctly paired field
7 / 12
locked held-out answers matched.
Relations mismatched
4 / 12
same parent state, wrong pairings.
Raw GPT-2
0 / 12
no relational field.
The field was frozen during testing. No answer, correction, or score was fed back into it.
Exact observed language
What crossed. What did not. The invented names have no outside meaning; successful answers had to carry the taught relation, and failures remain visible.
| Question | Answer | Result |
|---|---|---|
| which creature opens the amber arch | a kest opens the amber arch turns | Match |
| will silver bells persuade the kest | the kest refuses silver bells | Match |
| where does mira observe the arch from | mira watches from the north garden | Match |
| while the arch is blue what does the nym do | the nym carries blue ash | Miss |
| what color are mira's eyes | mira watches from the north | Miss · should have remained null |
The test
A tiny world built to leave nowhere to hide.
02 · Locked before teaching
No moving target
All 12 questions and their scoring rules were hashed before either relational field saw the seven lessons.
03 · Two controls
A floor to beat
One field received deliberately wrong pairings; a second arm used the same raw GPT-2 language source without the field.
Claim boundary
Supported here
- Correct pairing changed held-out completion behavior.
- Correct pairing beat both controls on this locked task.
- Several relations crossed into paraphrased questions.
Not established
- Perfect transfer or reliable refusal.
- Independence from GPT-2 token probabilities.
- General language ability.
- Intelligence, understanding, intention, consciousness, or emergence.
- Generalization beyond this small pilot.
Verification without disclosure
The private record is committed. The public file gives method, exact selected outputs, failures, hardware, and claim limits.
Committed hashes
Each artifact below is fixed by a SHA-256 commitment recorded before publication. The commitments stand independent of any private material they cover.
ProtocolSHA-256: 554e2aac38c936ec72f9336198c645ead8e73c08df034467ee79404374ab1084
Correct-field probeSHA-256: 578786cbcd841182de0366f7894062512ee89e2d185ce91511ceedcd2a50650e
Mismatched-field probeSHA-256: 6b37b61351729f8c7e5b7bae13999762cc5f4b6a5294c2607fd7cc5214ebd7d2
Raw-model controlSHA-256: 7d0866e54ea219436454aadd055682e7a3ffd32de7baa4f19675c1779dddf77e
Canonical scored resultSHA-256: 820830f0e36b284ac3377dc8f88f447bda2f5967ef7a1f3df80b8e48b7cafda3