Clarify inclusion learning and browsing context

[isomorphisms]

Summary

• correct the SVM terminology and make the baseline explicit: learned affine planes, separate high-recall proposal thresholds, per-example slack variables, support-vector diagnostics, positive-unlabeled evidence, overlapping category decisions, and optional bagged or hierarchical refinements
• make intentional human organization typed supervision while keeping model proposals, candidate retrieval, and accepted durable membership distinct
• establish the first two frontends as visual pre-paint and a ChatGPT-like text-only-by-default task workbench
• specify focus-region prefetch, bounded GitLab/software and arXiv investigations without speculative tab creation, cheap early source-backed summaries, and a scoped provider-neutral browser-to-assistant context bundle
• add a disposable explicit-hyperplane probe with reproducible embedding/input provenance, held-out isolation, authoritative overrides, explicit (w) and (b), support examples, slack violations, and unique provisional-unlabeled resampling
• recover only explicit prior user assertions into a provenance ledger, pin them to exact sanitized source rows, and expose a three-fit/two-held-out reading probe without converting old assistant partitions into ground truth

Experiment scope

The probe shortlists mxbai-embed-xsmall-v1, e5-small-v2, static-retrieval-mrl-en-v1, and snowflake-arctic-embed-xs. The first real benchmark now covers only mxbai-embed-xsmall-v1; the other shortlisted encoders remain unbenchmarked.

The recovery audit did not find a complete authoritative membership artifact for either the 491-row exercise or the 227-row reading/non-reading split. Both complete partitions were assistant-generated, and the saved 227-row output files were not recoverable. They remain excluded. The checked-in ledger contains only exact user assertions that could be tied to exact URLs. Five reading positives occur in the pinned 219-row sanitized fixture, enough for a positive-unlabeled reading probe with three fit and two row-disjoint held-out positives. This unlocks a limited experiment; it does not claim reproduction of either old partition.

First real embedding result — 2026-09-13

GitHub Actions run 34777706264 at head f02f0f94bcfab94c243843f6dff98b5da9233c2c completed successfully and preserved the input text, real vectors, embedding provenance, full probe JSON, and diagnostic summary as artifact ib-mxbai-reading-probe.

The run used mixedbread-ai/mxbai-embed-xsmall-v1 revision b0561d9a97e6b298da39f0ef3e7d3cf153b1b29a, onnx/model_quantized.onnx SHA-256 952f996d8cf46c311ee8654a750fa942b71c8b94aabe69d043dbb2bcaff5528e, ONNX Runtime 1.19.2, mean pooling, 384-dimensional float normalized outputs, and the existing URL-only url: <url> grammar with a 256-token budget. No fit/evaluation labels or probe hyperparameters were changed after seeing the held-out rows.

The fixed probe retained 16 unique planes, each fitted from the same three authoritative positives plus six provisional-unlabeled rows. Every one of the nine training rows was a support vector in every plane. Pairwise cosine between normalized plane normals ranged from 0.578891 to 0.772758, with median 0.670730, so changing only the provisional-unlabeled sample materially changes the separator.

The learned high-recall proposal threshold was 0.298998456 and produced only three model proposals: exactly the three fit positives. Held-out positive recall was 0/2. The Rezchikov visualization ranked 37/219 with aggregate score -0.016488046 and was above the fitted zero surface in 6/16 planes. The Decent One ranked 58/219 with aggregate score -0.041750192 and was above the zero surface in 2/16 planes. Neither passed the untouched proposal threshold.

The fit positives were Problems I Like, Haaretz offers, and the Benn Peifert X post. Problems I Like and Haaretz had only tiny maximum reported slack (0.000657 and 0.000335). The Benn Peifert URL was the difficult positive: it had nonzero slack in 14/16 planes and maximum reported slack 0.343927. Because the input artifact contains URLs rather than page semantics, an opaque X status URL is particularly weak evidence from which to infer a reading category.

This is a useful negative/diagnostic result for the present URL-only three-positive probe, not evidence that the held-out assertions are wrong and not a measurement against either old assistant-generated partition. Unlabeled rows remain unlabeled: slack on a provisional-unlabeled row means tension with that temporary negative role, not a false-positive or false-negative judgment. The fact that every training row is a support vector makes support-vector membership itself non-selective here; bag-to-bag rotation, slack magnitude, held-out ranking, and zero-surface disagreement are more informative diagnostics. The run does not establish that mxbai is generally unsuitable; the supervision and URL-only input are too sparse for that claim.

Verification

• python3 -m py_compile experiments/category-hyperplanes/probe.py experiments/category-hyperplanes/build_url_inputs.py experiments/category-hyperplanes/validate_recovered_labels.py
• python3 experiments/category-hyperplanes/probe.py --self-test
• python3 experiments/category-hyperplanes/validate_recovered_labels.py
• pinned 219-row canonical-input build and count validation
• end-to-end CLI/report validation with recovered row-disjoint labels and dummy test-only vectors
• real mxbai-embed-xsmall-v1 ONNX embedding and positive-unlabeled probe in GitHub Actions run 34777706264
• preserved embedding/probe artifact with checksummed model and tokenizer provenance
• independent reviews of the SVM semantics, model provenance, and prefetch/frontend consistency