Laya-GUT-Finetune

Laya-GUT-Finetune — monitor-triage · GUT (Guided Unconscious Thinking) by qalarc

GUT stands for Guided Unconscious Thinking — the fast, instinctive judgment layer (System 1) that acts before slow reasoning is needed. This checkpoint is a fine-tuned Laya 421M (Convai Innovations, Apache-2.0) that reproduces qalarc's cloud decision verdicts locally through the same wire protocol (System One typed decisions: choice / score / noul with calibrated probability distributions).

This model decides monitor alerts: severity, wake-the-owner urgency, and likely cause. It is one of ten GUT specialists. Live instance, the full pipeline, and the other domains: gut.qalarc.com · github.com/qalarc/gut-finetuned · qalarc.com/projects/gut

⚠️ Research checkpoint — did NOT pass the calibration gate at time of upload. Teacher-argmax agreement: 0.9565 (bar: ≥0.85 ✓) · ECE post-temperature: 0.3964 (bar: ≤0.10 ✗). Not registered for production serving. Published for transparency and reproducibility; the gate is the product. See the eval ledger below.

How it was trained (methodology)

1. Jev-shadow collection — every verdict the qalarc fleet's cloud teacher (TypeSafe Jev) makes in production is logged as a training row with its soft probability distribution (soft labels preserve the teacher's hedging).
2. RLCD loop — GRPO-style group sampling (G=4, σ 0.4→0.1, zero-mean projected) scored by a proper scoring rule (log + spherical + RPS), policy gradient + full soft cross-entropy guidance, AdamW with encoder/head learning-rate groups, cosine schedule.
3. Per-bucket calibration — LBFGS temperature fitting per (question-type, option-count bucket) on the held-out split.
4. The gate — held-out teacher-agreement ≥ 0.85, ECE ≤ 0.10, no crash buckets. No pass → no production registration. Ever.

Domain: monitor-triage

Decides, from a monitor-alert state:

• severity — choice: low / medium / high / critical
• wake — truthy: wake the owner at 3am?
• cause — choice: origin_down / network / tls / upstream

Data: 1,200+ Jev-shadowed monitor states from qalarc's live operations (alert transitions, site incidents, contact-form traffic), soft probability labels, row-level 90/10 split (no state leakage into held-out).

Usage

from laya import Agent
a = Agent("Qalarc/Laya-GUT-Finetune")
print(a.predict("site x HTTP 503 4 consecutive checks, 12min, cert valid",
                {"severity": {"type": "choice", "instructions": "Severity",
                              "criteria": {"low": "cosmetic", "medium": "degradation",
                                           "high": "outage", "critical": "emergency"}}}))

Or serve it behind the exact Jev wire protocol:

POST http://127.0.0.1:8798/v1/systemone
X-Laya-Domain: monitor-triage

(serve.py in github.com/qalarc/gut-finetuned)

System requirements

• Serve: 8 GB RAM, any CPU (~2.5 s/decision) — or any 4 GB GPU (17-23 ms/decision)
• This fine-tune: trained on free Kaggle 2×T4s; small domains fine-tune on a 16 GB-RAM workstation in about an hour

License

Weights and datasets: CC BY-NC 4.0 (free for non-commercial use; commercial licensing via qalarc.com). The Laya base model remains Apache-2.0 (Convai Innovations) — see ATTRIBUTION notes in the pipeline repo.

Citation

@software{qalarc_gut_2026,
  title  = {GUT — Guided Unconscious Thinking: distilling a cloud decision engine into an open 421M model},
  author = {qalarc},
  year   = {2026},
  url    = {https://gut.qalarc.com}
}

Eval ledger (updated per run)