v0.8.481

• meta endpoint DAG changes; targets can now be optionally specified
• fix logic around storing sessions
• Regression UQ: cross-fold capture reports out-of-fold UQ, not in-sample
• raise instead of silently dropping oversized parameters
• fix logic for serialized molecules
• blog update
• Parse injected hyperparameters in the fingerprint template
• blog changes
• Add --foundation to the CYP2D6 model
• State what the anchor gate actually measures
• Pool public CYP potency into the scored column
• bumping version of workbench for all images

v0.8.480

• View cleanup gc (#615)
• Replace view existence full-scan with a Glue point lookup
• UQ blog rewrite around v1, drop knn_target_count, shuffle calibration folds
• replace deprecated ml.t2.medium with ml.c7i.large for endpoints
• bumping version of workbench for all images

v0.8.479

• adding an instance ladder quota check script
• adding an instance ladder quota check script
• changes to the instance ladders
• cyp_compare: read the capture models actually have, and split by activity band
• CYP2D6: --low-weight and --deep-weight for the band the model cannot rank
• CYP: --censored takes all three public sources, not just ChEMBL
• bumping version of workbench for all images

v0.8.478

• sample_weights: honour them in chemprop and pytorch, not just XGBoost
• sample_weights: honour them in chemprop and pytorch, not just XGBoost
• adding an instance ladder for single GPU jobs
• Task weights: fix the float32 noise at the source, shorten the rest for display
• CYP union: name and gradient share stop going stale as tasks are added
• bumping version of workbench for all images

v0.8.477

• adding a way to track head counts
• updating our use of public data
• CYP: say variant, not arm, for an experiment variant
• CYP: record MultiTaskAlignment's aux verdicts before the arms are built
• CYP: a censored ChEMBL arm for the union model
• Public CYP data: correct the sub-4.0 claim to its own isoform
• Censored CYP data: publish assay bounds under the {target}_lt convention
• Regression UQ: out-of-fold error-model calibration, v1 default, drop Gaussian CRPS
• bumping version of workbench for all images

v0.8.476

• Multi-target remap: only drop standard columns in a per-target frame
• Regression UQ: calibrate every target, not just the primary
• conf_weights_with_fallback: missing confidence falls back instead of poisoning the row
• CYP: Tox21 anchor check and a planning-doc claims verifier
• EnsembleSimulator: analyze a chosen target of a multi-target model
• MultiTaskAlignment: volume alone no longer earns a Use
• FingerprintProximity: selectable binary fingerprints, sturdier inputs
• Add a multi-task guide for Bosco
• MultiTaskAlignment: score on rank and magnitude, fix coverage lookup
• adding an update and create list at the bottom
• endpoint modified() anchors on EndpointConfig, not LastModifiedTime
• adding tox21 to the cyp datasets
• adding tox21 to the public datasets

v0.8.475

• change start up dialog when there's no Bosco key
• docs: README revamp — local-mode-first Getting Started
• docs: Bosco blog — resident ML agent framing + 11-turn worked session
• cyp contest updates

v0.8.474

• adding local mode prompt to Bosco
• aws setup intro message

v0.8.473

• Llm refactoring (#614)
• Local mode support (#613)
• bumping version of workbench for all images

v0.8.472

• Api key changes (#612)
• updating the HPO documentation
• updating the chemprop hyperparameters documentation
• Revert CYP2D6 placement, drop the offline ST-RAE optimizer
• CYP2D6's placement is now derived by minimising ST-RAE rather than R²-optimal moments
• some cleanup/reorg for our cyp scripts
• edits for a more concise blog
• cyp challenge blog edits (a bit more concise)
• cyp challenge blog edits (a bit more concise)
• cyp challenge blog edits (a bit more concise)
• cyp challenge blog edits (a bit more concise)
• CYP challenge: ensemble submission and blog rewrite
• moving old ml_pipelines to storage
• bumping version of workbench for all images