Research
Two streams, one notebook. Model development — the Kaer Reasoners line, and the architecture and training work underneath it. And measurement — what we and everyone else can be shown to do once the weights are frozen. Numbered entries are studies; most link to a full write-up on the blog, the rest are partner reports available on request.
A 7B reasoner trained to decide, per question, how long to think. A learned halting head plus budget-matched supervision closes 71% of the gap to a model eleven times its size on our hard-reasoning suite, at 8% of the inference cost — and, more usefully, tells you when it is out of its depth.
Ablating a 240B-token pre-training mix one slice at a time. 61% of the tokens are within noise of contributing nothing to any downstream metric we track, 4% actively hurt calibration, and the slice that mattered most cost less than the compute we spent measuring it.
Paraphrase sensitivity across 312 models: a 19-point median accuracy spread across twelve phrasing families, against an 11-point gap between adjacent-tier models. Question-order dominates; typos barely register.
Longitudinal trace of one deployed assistant: nine behavioural step-changes (six unannounced), flat task accuracy, monotonic style drift, and one month where date arithmetic quietly broke and recovered.
A minimal decision-grade evaluation recipe: 400 stratified scenarios, pinned local judges, batch pricing. Kendall τ = 0.83 rank agreement with the full Atlas protocol at roughly 1% of the cost.
Hand-coding of 9,400 refusals yields a six-shape taxonomy. Silent scope-narrowing — answering an easier nearby question with no signal — is the most common shape in the flagship class and invisible to standard rubrics.
Scoring models on realistic traffic: consented, hand-paraphrased scenarios with domain-worker judges and a helped/neutral/harmed scale. Leaderboard rank explains 41% of helped-rate variance; harmed-rates spread 4–19%.
Pre-deployment and live-traffic evaluation for three London borough services: harmed-rate bounds, deadline-question safeguards, and the human-fallback usage metric now adopted by all three partners. Partner data; summary available on request.
Foundation document for the Model Atlas: domain taxonomy, blind pairing design, judge-panel construction and pinning policy, paraphrase-spread reporting, and the five published axes. Updated with each release.
Programme one
Models that spend their compute where the difficulty actually is. Adaptive halting, budget-matched supervision, and ruthless data ablation (R-09, R-08) — then the same fixed probes, fixed seeds and published variance we hold everyone else to (R-07, R-03).
Produces the Kaer Reasoners line. Feeds the Atlas reliability and cost axes.
Programme two
An answer you can't trace is a rumour with good grammar. Refusal taxonomy (R-04), answered-the-actual-question checks, and mechanism tracing — exhaustive on weights we hold, best-effort and clearly labelled on everyone else's.
Every Kaer release ships with its interpretability card. Feeds the Atlas legibility axis.
Programme three
Behaviour under pressure, over months. The eight-month drift trace (R-06) is the template: frozen probes, change-point detection, manner metrics that move before competence does. We run it against our own checkpoints on the same cadence.
Gates every Kaer release. Feeds the Atlas human-fit and drift axes.
Every lab says "aligned with human values" and hopes nobody asks which ones. We keep a list. It's short, unglamorous, and non-negotiable — the working faculties that make a human, human. A system passes when it strengthens them in the person using it. It fails when it quietly rents them out.
This is not a slogan we bolt on after training. It is a release gate. A Kaer checkpoint that scores well on reasoning and badly on rented judgement does not ship, however good the headline number looks — and we have held two back on exactly that basis. Our rented-judgement metrics (in R-06) are the first attempt to put numbers on it.
See it measured