Equal Evidence, Different Outcomes
A process-isolated evaluation of governed external reasoning around small language models.
HeadSoft Research / Active program
Amoeba Core v2
A governed intelligence substrate that treats language models as proposal-producing components while deterministic memory, evidence authority, selectors, proof obligations, and cost policy control what becomes actionable knowledge.
Status experimental research system
Paper empirical v0.1
Primary finding
Governance changed outcomes under matched evidence
In the frozen 48-case target, the raw Gemma 3 1B proposal scored 18/48, flat prompting with acquired evidence scored 14/48, and the governed runtime scored 48/48. Flat and governed conditions used the same evidence and matched normalized work.
6 processesFresh provider process for every phase-condition pair
48/48Raw and governed began from identical model proposals
p = 1.164e-10Exact paired sign test, governed versus flat
Evidence map
What is measured, and what remains open
Amoeba separates mechanism demonstrations, recognized-dataset capability runs, generated ablations, and transfer studies. Results from one tier are not silently promoted into another.
| Evidence | Measured result | Interpretation |
|---|---|---|
| Equal-evidence target | 18/48 raw, 14/48 flat, 48/48 governed | Primary controlled mechanism result |
| HumanEval local run | 129/164 raw, 164/164 Amoeba primary route | Closed-book, nonofficial, and benchmark-developed; not a leaderboard claim |
| 72-case synthetic transfer | Top-4 retrieval and typed Amoeba both 72/72 | No accuracy advantage over the strongest retrieval comparator on this set |
| 16-case untouched holdout | 14/16 selected policy vs 12/16 static, p = 0.5 | Small, non-significant transfer result |
Publications and artifacts
Research you can inspect
The web paper, PDF, Markdown source, evidence ledger, verification receipt, and source manifest are published together. The architecture program remains separate from empirical claims that have already been earned.
Machine-verifiable ledger
Eleven evidence entries and 219 integrity checks connect headline numbers to source artifacts and hashes.
Governed intelligence substrate
Typed memory, compositional process IR, provenance, scientific-method gates, cost policy, and evolutionary search are active architecture subjects. Their evidence levels differ.
Current claim boundary
The evidence supports a narrow causal claim: governed deterministic evidence processing can change a fixed model's outcomes beyond flat evidence injection in a controlled family. It does not yet establish general superiority over retrieval, official benchmark standing, broad autonomous science, frontier-model equivalence, or general-purpose transfer.