[cosmicbboy]

Adds the "non-fancy deep learning" baseline the tuned policy must justify itself against.

Approach chosen: quantile gradient-boosted trees

Per the resource-prediction literature — HPC memory-requirement predictors, Slurm ensemble predictors, and Google Autopilot's percentile framing — tree ensembles over engineered features, predicting a high quantile of peak usage so the safety margin is part of the loss:

• memory: HistGradientBoostingRegressor(loss="quantile", q=0.9) on log2(peak MiB)
• cpu: quantile 0.75 on cores
• gpu: classifier over {none, T4, L4, L40S} + a VRAM quantile regressor that picks the cheapest fitting card
• features: sampled params, input-profile numbers, lexical code flags, author prior, run history — what a scheduler sees without a language model

Options considered and rejected for the first cut

• Bayesian ridge (μ + k·σ headroom): interpretable, closed-form uncertainty — but linear, and the corpus's param→footprint curves aren't (predictive-uncertainty review)
• Autopilot-style percentile-of-history: strongest with run history, no cold-start answer; the tune service's ledger-history path already covers that regime
• k-NN / collaborative filtering over code similarity: needs embeddings to compete — that's the fancy path this baseline exists to contrast

Wiring

• eval_tuner renders a three-way table (policy vs rule baseline vs ML baseline) incl. GPU-task success, and reports ml_baseline_* keys
• the lineage dashboard's reward-shape comparison shows the ML $/task-hr beside the policy's
• driver image gains scikit-learn (baseline trains in-process inside eval, seconds on CPU)

Tests

5 new (162 total passing): feature purity/determinism, grid-snapped proposals, GPU recall >90% with <5% spurious on heldout, and the existence proof — the GBT baseline matches the family-median rule baseline's fit rate with no extra waste in the simulator.

🤖 Generated with Claude Code

https://claude.ai/code/session_016huuSPKsdzj5YwzV846E8e

[cosmicbboy]

Live three-way result from uvl4sb6j7ttlwvqktwz5 (r8-c-cost checkpoint, 256 heldout, sim scoring):

estimator	fit	$/task-hr (requested)	$/successful task-hr
LLM policy	55%	$0.1188	$0.216
rule baseline	43%	$0.1717	$0.399
ML baseline (quantile GBT)	99%	$0.2163	$0.219

The GBT nearly never fails and ties the LLM on effective cost — full read in the round-10 research-log entry on this branch. The comparison table is now standing in every eval report + the lineage dashboard.

[cosmicbboy]

GBT is now a first-class citizen, verified live on rt-tune:

• fit_ml_baseline (usgk2plg) fitted on the 500k corpus and published ml-baseline-model v1 (joblib + manifest with sim self-scoring vs the rule baseline in its report)
• tune service loads it on /reload (ml_baseline_loaded: true in /health) and serves it three ways: {"estimator":"ml"} per request → source=ml_baseline ✅ verified; RT_TUNE_ESTIMATOR=ml fleet-wide; and as the middle rung of the degradation ladder (LLM → ml_fallback → prior)
• estimator is part of the proposal-cache key — A/B arms never share answers (verified: same context, different digests per arm)

One honest serve-time note: at serve time the GBT sees no params_json/family features (cold-start regime), so its answers there are more conservative than its 99%-fit eval numbers, which include generator params. The eval caveat in the round-10 log covers this.