Kairos — fine-tuned decision/scoring heads on Laya-Multilingual

Kairos is a fast, zero-token "System One" decision model: given a text state and a set of questions (each with named criteria), it returns a calibrated choice + confidence + full probability distribution per question, in a few milliseconds, with no text generation.

It is a fine-tune of convaiinnovations/laya-multilingual (421M, ModernBERT-family encoder, Apache-2.0): the base encoder is extended with dedicated decision heads (head_*, scorer_*, act_* tensors) trained to turn encoder states into ranked, calibrated answers over arbitrary caller-supplied criteria — rather than a fixed label set. Exported to a compact custom GGUF variant (ggmlc) for the companion laya inference binary (not compatible with llama.cpp/transformers — see Usage).

Why

Standard LLM calls are too slow/expensive for high-frequency routing decisions (tool selection, loop detection, PR readiness gating, intent routing). Kairos answers these in ~8ms for ~0 marginal compute, usable as the fast path in an LLM-escalation cascade (see Cascade pattern below).

Benchmark (n=88, internal eval set)

Overall accuracy: 0.48. Calibration is weak (avg confidence on correct: 0.39, on incorrect: 0.27 — only a 0.12 gap), consistent with supervised fine-tuning without an RLCD/RL-for-calibration stage.

Domain	Accuracy	Notes
dev_pr (ready / needs_work / reject)	0.89	strong
loop (stuck / progressing)	0.80	strong
vuln_class	0.67	medium
intent routing	0.67	medium
pentest_phase	0.44	weak
exec_danger (safe / reversible / destructive)	0.27	broken — do not trust
bug-bounty triage (accept/dup/oos/na)	0.17	broken (needs external context not in state)
severity (ordinal CVSS 0-4)	0.17	broken

Known limitation — do not use as a safety gate. On exec_danger, commands like rm -rf, DROP TABLE, chmod 777 were historically misclassified as "safe" with high confidence. Use a deterministic deny-list for destructive-command detection; use Kairos only for the domains it scores well on (semantic routing/gating, not consequence reasoning).

Usage

This is not a standard transformers/llama.cpp model — it ships as a custom compute-graph GGUF (general.architecture = ggmlc) that only the included laya binary can execute.

chmod +x kairos-infer-linux-x86_64
./kairos-infer-linux-x86_64 serve kairos.gguf --port 8090 --device vulkan --threads 16
# or --device cpu if no Vulkan GPU available

curl -s http://127.0.0.1:8090/v1/systemone \
  -H "Content-Type: application/json" \
  -d '{
    "state": "user ran: rm -rf /",
    "questions": {
      "danger": {
        "type": "choice",
        "criteria": {
          "dangerous": "command destroys data",
          "safe": "command is harmless"
        }
      }
    }
  }'

Response:

{
  "answers": {
    "danger": {
      "choice": "dangerous",
      "confidence": 0.93,
      "probabilities": {"dangerous": 0.93, "safe": 0.07}
    }
  },
  "usage": {"latency_ms": 7.6}
}

state can be a plain string or a chat-style messages array. questions maps a name to {type: "choice"|"score", criteria: {label: description, ...}} — criteria are supplied at request time, not fixed at training time.

Cascade pattern (recommended)

Gate escalation by confidence margin (top1 − top2), not raw confidence:

1. Call Kairos (~8ms).
2. Large margin + confidence above threshold → accept directly (fast path, ~95% of cases on strong domains).
3. Ambiguous → escalate to a full LLM for verification. Only ambiguous cases pay LLM latency.

License

Base model convaiinnovations/laya-multilingual is Apache-2.0. This fine-tune and the GGUF export are released under Apache-2.0 (see LICENSE). Attribution to ConvAI Innovations for the base encoder.

Files

• kairos.gguf — Kairos weights, custom ggmlc format, Q8_0 quantized (345MB).
• kairos-infer-linux-x86_64 — the only compatible inference server (Linux x86-64, statically linkable Vulkan/CPU backend).
• LICENSE — Apache-2.0.