Research — raising coherence + structured verdicts on the tiny (25M) head

Date: 2026-08-07. Situation: fsi-anomaly, ~25M liquid-architecture model, CPU-only ARM tablet, needs (a) coherent chat and (b) reliable structured verdicts. Stage-A SFT produced garbled verdict strings (eval 0.0) — the head held the analyst register but not clean sentences.

1. "Higher-quality, more diverse gold" — VERIFIED (with a nuance)

LIMA / "The False Promise of Imitating Proprietary LLMs" (arXiv 2305.12227): 1-2k hand-curated, high-quality examples can beat 100x more imitation data. Supports the owner's "quality over quantity" thesis — but the key word is DIVERSE + CONSISTENT FORMAT, not just "more." Concentrate on variety across the 10 SOP/probe domains and a few hundred CONSISTENT templated rows, not bulk.
phi-1 "Textbooks Are All You Need" (2306.11644): tiny models leave clean when trained on clean, textbook-grade data — coherence follows data cleanliness.

2. "Train with thinking/reasoning traces" — VERIFIED, current practice

ReasonLite (2025) and Skip-Thinking (2025) (Semantic Scholar): distilling long chain-of-thought "traces" into a small LM makes it reason better and more coherently on multi-hop tasks. Our <|scratchpad|> is the right scaffold; the upgrade is LONG, structured multi-step traces (look-up -> decompose -> check -> verdict), not short one-liners.
STaR lineage / Self-Taught Reasoner-with-Tools (2025), and CoT (Wei et al.): a loop where the model proposes a trace, we judge the answer, feed back the good traces = self-improvement for the head. Cheap for us: our verdicts are deterministic/checkable, so the su git "reward" is free.

3. The biggest "software we can build" lever: GRAMMAR-CONSTRAINED DECODING

Why the eval scored 0.0: the head proved {supports, refutes, not enough info, unclear} and confidence, but the free string came out garbled ("confinmpanislerue") and the string-parser couldn't credit it.
Fix (No retraining): constrain token choice at decode time to an ALLOWED set for the verdict + confidence fields (llam2.cpp grammars / Outlines / guidance-style, done lightly in our own decoder). The head only has to PICK the right verb — a 25M model can do that. This turns a structural 0.0 into a real score, today.
- Works WITH a closed loop: constrained decode → parse → run via the suit (verify.py, dual-mind) → write goal to memory. Same closed loop, but the brain is never allowed to free-form the critical field.
Confirm software: llama.cpp grammars, Outlines, Guidance (all public); our own research/structured.py already does constrained report-decode — we harden it.

4. "Sandbox / closed loop / scratchpad" — already in skill; deepen reuse

Room/workspace + write-back + replay loop exist (tiny-model-suit). The new bit from research: the trace loops emulate self-improvement without a big teacher — generate trace, check verdict with the rule spine, keep only winning traces. This is STaR-style and cheap because the reward is checkable.

5. Dark hole / unverified this session (honesty)

arXiv API (export.arxiv.org) and Liquid's docs site timed out / are JS-gated, so I could NOT cite Liquid AI's exact written recommendations this session. Liquid's own story: gated recurrence + strong long-context/edge focus; treat that as context, not verified quotation. Re-run this note when connectivity allows.

Conclusion / next actions (ordered)

Constrain decoding of the verdict/confidence fields now (NO training). Fast, measurable: Stage-A eval score should move from 0.0 to a real number.
Expand gold with reasoning-trace rows (multi-step CoT: look-up→check→verdict; diverse across per probe domains) + a fluency-retention mix so the head stops fragment-echoing. Retrain, re-probe.
Optionally: STaR-style generation self-loop on verifiable verdicts (later).