Model router

examples/model-router.algal.json routes a first-session task prompt to frontier or standard: the binary capability tier a coding-agent model picker decides at session start. It is small, fast, and replayable: two expr cells compute deterministic shape features and the routing score; one decide cell asks a typed-decision provider (Jev) for six scores.

bun cli.ts run examples/model-router.algal.json \
  --args '{"src":{"task":"migrate the session store and update every caller"}}' \
  --jev

Output: {"route": "frontier"|"standard", "score": <logit × 1000>, "kind": <task kind>, "gate": "kind"|"score"} plus a one-line summary.

Shape

src (task text)
 ├─→ feats  (expr)  words, imperative, resume, verbs_c
 └─→ probe  (decide) difficulty, scope, ambiguity, stakes, kind, frontier
        ↓
     combine (expr) linear head + kind gates → route, score, kind
        ↓
     fmt    (expr)  "route → frontier (score 648, kind resume)"

The deterministic features are deliberately weak alone; the typed scores are deliberately generic. The fitted head combines them:

z = -1.7683
  + 0.5271 · min(words,400)/400
  + 1.4028 · imperative          (first word is an action verb)
  + 0.4491 · resume              (session take-over / continuation marker)
  + 0.4438 · min(distinct verbs,8)/8
  + 0.0045 · difficulty/5
  + 0.9353 · scope/5
  - 0.1459 · ambiguity/5
  - 0.5129 · stakes/5
  + 0.6107 · frontier-noul
route = frontier  if z ≥ -0.619  (p ≥ 0.35), else standard

question and probe kinds gate to standard unconditionally: the observed operator never routed a pure question to a frontier model, and a capability probe does not need one.

Provenance and limits

The head has two generations, both fit on one operator's local Codex, Claude Code, and Devin CLI stores and labeled by the model that served the first turn: claude-fable-* and gpt-6-astra* count as frontier; swe-2-*, gpt-5-6-sol*, gpt-5.6-*, claude-opus-*, claude-sonnet-*, and the remaining standard tier count as standard. Harness-generated prompts (eval batteries, organism cells, guardian probes) are excluded, and probe-kind sessions are dropped from fitting.

Generation 0 (September 22, 2026) was an L2 logistic fit on 306 first prompts from July 30 to September 21: 8-fold cross-validated AUC ≈ 0.74, in-sample accuracy ≈ 0.77 at the shipped threshold, predicted frontier share 37% against 36% observed. Intercept -1.8027; weights 0.4530, 1.5943, 0.6881, 0.4405, -0.0755, 0.8716, -0.1508, -0.4787, 0.5851 in the feature order above.

Generation 1 (September 24, 2026, shipped above) is a prior-anchored update of generation 0, using the same rule xcb's route reflex learns with (xcb_core::reflex::fit: full-batch logistic descent with a Gaussian penalty of strength 24 centred on the parent weights, 600 iterations, learning rate 0.5). It was fit on the 84 labeled, non-probe first prompts from September 2026, the only window in which the frontier choice existed on that machine (every May to August pick is standard). On those 84 sessions generation 0 scores AUC 0.63 and accuracy 0.58 at the shipped threshold; generation 1 scores 8-fold cross-validated AUC 0.64, in-sample AUC 0.65, accuracy 0.58, and predicts a 58% frontier share against 50% observed. The gain is small and within noise; the update is shipped because the shipped head should track the newest labeled window, not because it is a better router.

Two readings of the new window matter more than the weights. First, the decision moved from prompt to tool: 34 of 36 September Codex sessions ran gpt-6-astra, 12 of 20 Claude Code sessions ran fable, and 0 of 35 Devin sessions left the standard tier. Second, within that window the pick reads as difficulty-calibrated for the first time: longer prompts (words AUC 0.71, median 69 frontier vs 27 standard), higher Jev difficulty (0.70) and scope (0.69) all point to frontier, and the continuation marker that carried generation 0 is neutral (AUC 0.50). A head fit from zero on September alone reaches cross-validated AUC 0.65 with difficulty at +0.72 and resume at -0.32; it is published for comparison on hraness.com/prompting and not shipped, because 84 examples do not justify replacing the parent. A head fit across the whole May to September corpus reaches AUC 0.82 and is not shipped either: every frontier pick is a September prompt, so that fit learns the calendar through prompt-style drift.

Every feature is task-intrinsic: no operator identity, repository names, timestamps, or history. Local stores rotate, so the September 22 corpus is not reconstructible from today's stores; the aggregates behind both generations are published at hraness.com/prompting/data.json.

Cost and speed

One routing decision costs one bounded decide call (~1k tokens in, ~0.2k out with jev-latest) plus two fueled expr evaluations — no agent cell, no generated text. All six questions ride in a single ask.

Reuse

Any typed-decision executor can serve the decide cell — --jev, --executor-cmd, or a route/preset executor. Retrain the head by re-fitting the nine weights on a new labeled sample and editing the combine program; the cell boundaries do not change.