working note · confidence: medium · 2026-09-24 · [05]

Laya took the heuristics feeder seat

We swapped a general ~3B for a ~0.4B typed-decision specialist to win latency and fit. Card accuracy is the starting bar, not a promotion story. RLCD and careful data are how we climb without lying about gates.

The voice path already has a packed C route classifier. That classifier keeps cascade authority: answer, retrieve, or escalate. Ahead of it, the lab needed a soft heuristics feeder — a model that emits route and decision labels plus confidence, and planner-gate fields such as reason and optimistic ack.

For a while that feeder was a general Qwen-class model in the roughly 3B range — about 3.09B parameters on the public card. It worked. It was also a causal language model doing triage: generate tokens, then parse a decision. Triage wants a typed choice and a score, not a free-form chat turn.

We replaced that lab seat with Laya typed-decisions. Spelling: Laya. This is a living note. The path will deepen. I will update the note, not republish the essay.

What the seat is (and is not)

Laya sits in the lab cascade as the soft heuristics feeder. It is not the voice model. It is not the stock planner that still runs elsewhere in the house. It does not own listen, think, or speak. The packed C classifier still decides the hard route. Laya proposes labels and confidence in a shape the C path already expected.

That boundary matters. A specialist that pretends to be the whole cascade is how you get accidental promotes. A feeder that stays a feeder is how you change one wire without lying about the product.

Why Laya won on runtime

Better here does not mean bigger, and it does not mean SOTA first. Better means the right tool for typed decisions: smaller footprint, a schema-native head, local Apache-2.0 serve.

Laya typed-decisions is a ModernBERT-style encoder head of about 421M parameters — Hub weights about 843 MB — built for System One–style choice / score / abstain workflows. That is roughly 7× fewer parameters than a general Qwen2.5-3B before anyone starts quantization games. A general 3B can emulate invoice, security, support, or agent-trace decisions with prompts. Laya is trained to emit calibrated typed decisions directly.

Public latency color only: upstream docs cite small English choices on the order of ~33–40 ms (AG News, T4). We do not have a clean house head-to-head millisecond bake-off against the old Qwen feeder. The runtime case is specialization plus footprint, not a measured Qwen delta we invented.

Sequence we chose on purpose: ship the runtime win first (swap the lab overlay to the typed specialist), then train up. Do not wait for a frontier LM score and retrofit decisions afterward.

How the lab wire works

Lab-only: the System One client points SYSTEMONE_BASE_URL at a local laya-serve typed overlay. Same HTTP shape the C path already expected. The production base URL stays empty and unchanged. No live alias flip. No stock planner cutover. No claim that the public product voice path now runs Laya.

Anvil still owns admission and activation elsewhere in the house. This note is about the feeder seat, not about flipping smart-local or the planner pin.

Accuracy at swap — honest, no promote

The swap ships the published typed-decisions specialist, not a claim that house traffic already clears the gates, and not a claim that any fine-tuned checkpoint is promoted.

Published card baseline (public model card, not a house bake-off):

| Metric | Card value | |--------|------------| | Accuracy | 0.766 | | Brier | 0.062 | | ECE | 0.213 (before house temperature refit) |

On the same typed bench, base English Laya zero-shot lands near chance — about 0.36. The specialist checkpoint is the point of the swap, not “any Laya.” Soft accuracy color on the card is optional noise; we do not lead with it.

House traffic is not the public typed-decisions workflows. Historical cascade rows often lacked a clean noul field. Calibration and whether score means what we think it means were not house-proven at swap. Quarantine and fail-closed receipts stay the rule.

House bar before any future promote of tuned weights: hold or beat card accuracy (≥ 0.766), do not regress Brier (≤ 0.062), ECE after temperature refit at or under card ECE (aspiration ≤ 0.15). Temperature refit on sealed labels is required. Those numbers are the climbing bar, not a smoke-test win.

Closing the accuracy gap later

Teach the specialist on our decision shapes, one careful pass, measure calibration, promote only when the receipt clears.

1. Serve the specialist first (typed-decisions overlay for System One–shaped decisions).
2. Train offline with RLCD — reinforce good soft probability assignments with a proper scoring–style reward, plus a soft classification loss, in short one-epoch runs so a small set does not overfit.
3. Data mixes: public typed-decisions examples (Mix A) plus house-shaped synthetic heuristics (Mix B) that mirror real planner and cascade fields (routes and confidence), then fold real lab decisions when they exist. Prefer distinct decisions — no padding duplicates to fake volume.
4. Always refit temperatures after training and report Acc / Brier / ECE on a sealed holdout. Fail closed: no live alias flip from the training job itself.

Fine-tunes land in Anvil Experiments: stage the challenger, fixed pack, operator verdict and eligibility, quarantine until promote. Candidates stay promotion_eligible: false until sealed gates and a house GO. No CUDA serve promote from a green train Job alone. No silent swap of the lab wire into production.

A completed train is a receipt. Ready is not live. Live is a deliberate cutover after evidence.

What we learned

Ship the specialist when the wire fits and the footprint is honest. Do not wait for champion house accuracy if the old seat is an oversized causal LM doing triage.

Say what you measured. We measured specialization and footprint. We did not measure a clean Qwen↔Laya millisecond bake-off in our harness, so the post does not pretend we did.

Keep cascade authority in C. Keep the feeder soft. Keep prod URLs unchanged until gates and GO.

Separate the published card from house traffic. Card Acc / Brier / ECE are the starting bar for tuned promote. Near-chance zero-shot on base English is why Mix B, one-epoch RLCD, and temperature refit exist.

What’s next

RLCD mixes keep landing in quarantine. Anvil Experiments remains the compare lane. promotion_eligible stays false until sealed gates and a house GO. The stock planner elsewhere in the lab stays stock. The product voice path stays its own note.

Mix A+B fine-tune completed offline; promote stays held until sealed gates and a house GO. The library wave remains held. Voice canary still prefers synth while house clips are scarce. Candlekeep keeps the Qwen-Image quarantine shelf-ready and treats vendoring as the priority.

When a Laya candidate clears the gates and someone deliberately cutovers the lab feeder — or declines to — we will deepen this note. Until then, Laya is the lab cascade heuristics feeder because it is small, typed, and fast enough to sit ahead of C. Accuracy is the homework, not the launch speech.