Two earlier splits reshaped Everwhen's AI task set. #162 narrowed the Judge to pure writing craft (it no longer sees the goal); #163 pulled the anachronism "reach" rating out of the Timeline Engine into its own blind reach-rater. Once the tasks change shape, their per-task model settings in ai-routing.ts deserve a re-tune — that is this issue (#164).

This PR is the mechanical half of that re-tune. The routing table's settings are normally decided by a live bake-off (eval:bakeoff, judge-tune.eval.ts), which makes real model calls and needs both an Anthropic and an OpenAI key. Those keys weren't available here, so this pass lands only what is safe without a live run, and captures the one balance-sensitive change — the Judge — as a written plan for a keyed session. Net diff: one number changes (Timeline's token cap), the rest is provenance and a plan.

#163 removed the anachronism field from the Timeline Engine's output — it's an input now, rated upstream by the reach-rater. What's left is four small fields: a roughly nine-word headline, a one-to-two sentence detail, an integer swing, and a valence enum — under ~150 output tokens. But #163 left the maxTokens at 600, sized back when the output was larger and padded.

With thinking off (this route sets no adaptive thinking), Anthropic's max_tokens is a true output ceiling — no hidden reasoning is billed inside it — so 300 fits the real shape with better than 2x headroom. temperature and effort are untouched: the last bake-off held Timeline at 4/4 band ordering and a healthy neutral-EV corridor on those exact values, and nothing this pass moved them.

The issue asks for the cheapest model that rates reach reliably. #163 already pinned that — Haiku at temperature 0 is the canonical cheap-classifier tier — and the output is two small fields (a one-phrase reach gloss plus the tier enum), about 40 tokens. So the right re-tune outcome here is "hold," not "trim": the cap saves no money (you pay for real output, not the ceiling) and a lower cap would only risk clipping the reach phrase that anchors the tier.

The change is a comment that records this as a decision, so a future reader doesn't mistake the untouched route for an overlooked one.

The meatiest part of #164 is the Judge. Playtests showed "masterful" is almost unreachable (about 1 in 98 expert messages), and the bake-off flagged Haiku grading the middle rungs a notch harsh. The owner-approved direction is to bump the Judge to Sonnet and nudge the rubric so the top grades are reachable on pure writing craft.

That change is held back on purpose. The Judge sits on the dice-tilt and gains-multiplier path the run economy rides on, and the repo's rule is that the bake-off decides routes — shipping it unvalidated would break that contract. So the route is unchanged and a breadcrumb points to the plan.

The full plan lives in a results note: the rationale, the exact change to make (including that a Sonnet route must pin effort explicitly, which the routing spec enforces), and the commands that gate it — judge-tune.eval.ts for absolute calibration, eval:bakeoff for the re-pin evidence, and npm run eval to keep the standing "sharp ≥ vague" behavior green.