Representation & reusable skills
Can a small model compose verified operations on unfamiliar problems? We compare raw language, representations, primitives, reusable compounds, and a non-neural search baseline.
Read E01 protocol ↗Model systems laboratory · division of constrained computation
The Smol Model Company
of New Haven · small models, real problems.
We are a small local-model research program studying how much reasoning, composition, and adaptation can be obtained under practical compute limits — and how to carry earned improvements into an offline-capable assistant.
NHV-MARK-REV-A
compression glyph · signal unit
Research program / phase zero
A question worth making smaller
Under bounded training and inference resources, which allocations of computation, representation, memory, and reusable skills improve transfer to unfamiliar tasks?
Current state
The written research design, evaluation protocol, task-generator specification, and minimum-apparatus plan are public. No experiment runs or benchmark findings have been published. We count failures, resource limits, and contradictions as part of the record.
Recurrent computation
Can reusing parameters improve reasoning beyond ordinary depth or additional answer sampling? The initial work separates mechanism evidence from practical deployment claims.
Read E02 protocol ↗Temporary adaptation
When a task introduces a new rule, when does a bounded update help more than examples in context, retrieval, or explicit rule search — and what does reset cost?
Read E03 protocol ↗Evidence ledger / no decorative numbers
The public record
A score is not a theory. A pleasant demo is not a transfer result. The laboratory publishes the protocol, provenance, costs, errors, and limits alongside any claimed gain.
“Keep development feedback separate from final evaluation. Preserve unsuccessful runs and contradictions. A small clean negative result is a legitimate deliverable.”
Research conduct / current policyCurrent register
- Published results
- None yet. The results ledger is intentionally empty until a run earns an entry.
- Compute envelope
- Measured local desktop capacity guides the program; full fine-tuning is not presumed feasible.
- Review model
- Independent Forgejo identities, protected main, required review, and CI-backed documentation checks.
- Public source
- Forgejo is canonical; GitHub mirrors the public record.
Instrument card / practical local constraints
Measure the whole machine
We count model weights, caches, retrieval, learned libraries, verifiers, preprocessing, CPU memory, wall time, and external services. “Small” is a system property.
LOCAL RESEARCH ENVELOPE[▪] NHV
- Primary host
- RTX 3080 / WSL2
- Planning VRAM
- 7.33 GiB free
- Research form
- Local-first
- Rental status
- Not authorized
- Result policy
- Manifested
- Run ledger
- Empty by design
Operating principles
- 01Separate an intrinsic model capability claim from a complete local-system capability claim.
- 02Freeze protocol, scorer, and final split before confirmatory evaluation — not after we like an answer.
- 03Record OOMs, timeouts, invalid outputs, and missing artifacts in the denominator instead of quietly retrying them away.
- 04Move to a useful assistant only after a mechanism survives controlled evidence and an independent reproduction.
Open research / careful claims
Follow the work where it actually happens.
Issues, branches, reviews, and decision records live in the canonical Forgejo forge. The GitHub repository is the public mirror for readers and discoverability. If you find a mistake in a claim or a protocol, that is useful research feedback.