Commit History
27 days ago
-
Solomon v1.1: serve unscaled probabilities; headline and MMLU tables from one computation
Archer +1: Claude Opus 5 (1M context)- Readout temperatures set to 1.0 for every type: dev-fitted scalars worsened held-out test ECE in 8 of 10 type x modality cells (weighted 0.0212 fitted vs 0.0199 unscaled); fit recorded, not applied - All three serving bindings (bf16, fp32, int8) re-issued; live acceptance re-run: BF16 83/83, int8 subset 38/38 - Card: new Calibration section (per-cell unscaled vs fitted ECE, 0.03 target missed for yes/no and ordered); real test/dev and per-type counts and the MMLU table taken from the headline comparison (one definition for all models, external reference included); precision ECE unscaled; served score disclosure updated Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>6b9bcc4 -
Solomon v1.1: BF16 default, fp32 reference and int8 option, ranked evidence pointers, entity type removed
Archer +1: Claude Opus 5 (1M context)- Adapter d122466d and trained heads f766d752 (one merged yes/no head serves yes/no and multi-label) - Per-type readout temperatures (solomon-readout-temperature-v3), fitted on real dev from BF16 scores - One serving binding per precision (bf16 default, fp32, int8), identities captured live on the serving image - Experimental ranked evidence pointers (top 3 with scores; none when the answer is not stated) - Entity answer type removed from the API (old requests are refused with the replacement) - Model card rewritten with v1.1 measurements, accepted gate misses and AI-labelled evaluation disclosure - Live acceptance: BF16 83/83 on B200, int8 subset 38/38 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>d144fee -
Model card: add the BF16 MLX route alongside the float32 CUDA path
Archer +1: Claude Opus 5 (1M context)Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>8ba630c -
Clarify image support in the model card
Archer +1: Claude Opus 5 (1M context)Images are a supported input through the base vision tower. The adapter is text-trained and the shipped adapter has not been scored on images. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>ac4f9cc -
Document the evidence feature, and correct the headline claim
The card claimed in its opening sentence that answers come back "with the evidence spans each answer was read from", then documented nothing about it. That claim is false at the default evidence level and is now corrected: the answer is read from the whole document, and the spans are a separate word-overlap retrieval over the source text that the model never sees unless you ask for 'sufficiency' or above. Adds an Evidence spans section covering the four levels, their measured costs, the response shape, what is measured and on what population, and what is not established. Every figure is read from an artifact at generation time: - live acceptance (artifacts/scope9/acceptance/C4-49904e27814e): 20 of 57 checks are evidence checks; 0 / 10 / 16 fresh prefills for support / sufficiency / removal, and the per-unit cost rule is re-derived from the run's own questions and compared against those counts; - real-document demonstration (artifacts/scope9/demo/ scope9_demo_privacy-evidence_…7f592ed -
Pin the MLX hub loader to the published revision
The loader downloaded the Solomon snapshot at a revision from the repository's pre-publication history, which does not exist in the published repository and would have 404'd on first use. Repins SOLOMON_REVISION to the published commit and re-derives the converter code identity that guards it, so prepare.py's self-check still agrees. Historical conversion records under mlx/bf16/ keep the revision they were actually built from. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>b35fb3b -
Make MLX adapter-only and include the reproducible BF16 converter
The MLX setup reads the root adapter and heads, downloads the pinned Qwen base separately, and assembles a verified local BF16 model. Includes the exact historical converter, locked dependencies, source, integrity tests, setup instructions and dated CUDA-parity status. No new temperature fitting; complete held-out parity remains pending.5c0a4a8 -
Solomon v1.1
Archer HumeAdapter, trained readout heads, readout calibration and the serving layer for the Solomon decision model. The base model is referenced by pinned revision and is not redistributed; see NOTICE, MODIFICATIONS.md and MANIFEST.json.105f9ef -
initial commit
Verified signature340f8e3
