Proof

Seven pre-registered tests. Four adversarial falsifiers. All held. Re-run any time with py experiments/proof_indisputable.py.

The headline numbers

loading…

S3 — name-substitution control

Same syntactic templates, swap the entity name. If the saved LoRA encodes "Eli" specifically (and not generic English), the loss-drop should be much higher for Eli than for any other name. Numbers are mean loss drop (zero-LoRA loss − with-LoRA loss) across three templates.

Falsifier (pre-registered): Eli drop ≤ max drop for other names → claim falsified. Observed: Eli +1.208 vs next-highest +0.608, margin +0.601. Held.

S4 — random-LoRA negative control

Five fresh random non-zero LoRA matrices (NOT the saved one) tested on "Eli: My name is Eli." Random LoRAs should NOT match the saved LoRA's identity help.

Falsifier: random LoRA drop ≥ saved drop − 0.3 → "any LoRA helps" claim. Observed: random mean −1.405 (random LoRA actively hurts), saved +1.496, gap +2.901. Held.

S5 — T4 episode-specific recall, seed sweep

Train two SubstrateLMs from each of 5 seeds, give each a different conversation, measure both gaps. All 10 gaps must be positive (each model prefers its own past on held-out evaluation).

Falsifier: any seed fails T4 → episode-recall is seed luck. Observed: 5/5 seeds pass; 10/10 gaps positive; all > +1.5. Held.

Artifact receipts (SHA-256)

These are the files that produced the numbers above. Anyone can hash their copies and confirm same-binary reproducibility.

loading…

Reproduction

py experiments/identity_tests_substrate_lm.py    # ~30s on RTX 4060
py experiments/identity_tests_lora_v2.py         # ~10s
py experiments/proof_of_self.py                  # ~5s
py experiments/proof_indisputable.py             # ~2min (incl. seed sweep)
    

Output JSON ships next to this page: proof_indisputable_results.json, proof_of_self_results.json, identity_tests_lora_v2_results.json, identity_tests_substrate_lm_results.json.

Full writeup

Plain-English version with all 4 experiments, methodology, and what each one rules out: notes/proof_of_self_2026_05_12.md.

Architectural priors cited: Schlag/Irie/Schmidhuber 2021 (linear attention is Hebbian) · Carlini et al. 2022 (memorization scales log-linearly with duplication) · Charles et al. 2024 (user-level DP for LLM fine-tuning) · FDLoRA 2024 (federated LoRA shards).