Phase 3 classification harness - #3

Merged

johnoooh merged 31 commits into

mainfrom

phase-3-classification-harness

May 7, 2026

Commits on May 7, 2026

feat(phase-3): wire classification harness to live on-device model
johnoooh
committed
179501f
fix(phase-3): serialize classify calls — WebLLM engine is not parallel-safe
johnoooh
committed
26daf7f
feat(phase-3): expand fixture from 6 to 20 trials
johnoooh
committed
f15a555
feat(phase-3): add 4 realistic-length trials (~3kB eligibility each)
johnoooh
committed
34f8665
feat(phase-3): "copy results as markdown" button
johnoooh
committed
fa40555
tweak(phase-3): default prompt biases toward POSSIBLE over LIKELY
johnoooh
committed
379ff3b
tweak(phase-3): rewrite prompt with ordered rules + no-assume guard
johnoooh
committed
c66fc36
tweak(phase-3): few-shot prompt to fix literal-VERDICT bug + collapse rules
johnoooh
committed
8244252
tweak(phase-3): collapse to binary LIKELY/UNLIKELY classification
johnoooh
committed
d41cd78
feat(phase-3): register Llama 3.2 3B Instruct as a model option
johnoooh
committed
15d05b4
fix(phase-3): register Qwen2.5-1.5B + diversify few-shot examples
johnoooh
committed
c0051a3
feat(phase-3): multilingual preset switcher + custom-model stub
johnoooh
committed
b0f088f
feat(phase-3): translate-then-classify pipeline as harness toggle
johnoooh
committed
2ce00be
fix(phase-3): include translate-first state in markdown report
johnoooh
committed
6a9948f
fix(phase-3): wire translate-first state through markdown report (followup to a628b7a)
johnoooh
committed
c05ac74
feat(phase-3): production-realistic agreement toggle
johnoooh
committed
6f04fb4
feat(phase-3): wire stage-1 classifyAll into ResultsList + fit dot in TriageRow
johnoooh
committed
69f6fae
fix(simplifier): replace bare trailing labels with directive instructions
johnoooh
committed
44b9e7b
feat(phase-3): register Gemma 2 2B q4f16_1 as a faster alternative
johnoooh
committed
a5aca29
revert(phase-3): remove gemma_fast — q4f16_1 broke simplifier output
johnoooh
committed
2875992
revert(simplifier): restore PLAIN-LANGUAGE OUTPUT: label
johnoooh
committed
3845332
fix(phase-3): defer simplification until classification completes + raise frequency_penalty
johnoooh
committed
1f09331
feat(phase-3): pipeline progress caption in detail pane
johnoooh
committed
cd30a67
feat(phase-3): drop fit narrative section + add oncologist disclaimer
johnoooh
committed
9597334
feat(phase-3): two-tier disclaimer + carry redesign hot-fixes
johnoooh
committed
45d8839
chore(phase-3): remove disabled sort chips from toolbar
johnoooh
committed
73b809b
chore: post-merge follow-ups (matchMedia, regression test, doc move)
johnoooh
committed
1b3e759
fix(phase-3): re-run classification when patient description changes
johnoooh
committed
d473033
fix(phase-3): address PR #2 review must-fixes
johnoooh
committed
71926e6
chore(phase-3): post-review cleanup + disable classify-in-results
johnoooh
committed
e3cefa6
chore(phase-3): clear lint diagnostics introduced by this PR
johnoooh
committed
f99743b

Navigation Menu

Phase 3 classification harness - #3

Phase 3 classification harness#3
johnoooh merged 31 commits into
mainfrom
phase-3-classification-harness

Commits on May 7, 2026

feat(phase-3): wire classification harness to live on-device model

fix(phase-3): serialize classify calls — WebLLM engine is not parallel-safe

feat(phase-3): expand fixture from 6 to 20 trials

feat(phase-3): add 4 realistic-length trials (~3kB eligibility each)

feat(phase-3): "copy results as markdown" button

tweak(phase-3): default prompt biases toward POSSIBLE over LIKELY

tweak(phase-3): rewrite prompt with ordered rules + no-assume guard

tweak(phase-3): few-shot prompt to fix literal-VERDICT bug + collapse rules

tweak(phase-3): collapse to binary LIKELY/UNLIKELY classification

feat(phase-3): register Llama 3.2 3B Instruct as a model option

fix(phase-3): register Qwen2.5-1.5B + diversify few-shot examples

feat(phase-3): multilingual preset switcher + custom-model stub

feat(phase-3): translate-then-classify pipeline as harness toggle

fix(phase-3): include translate-first state in markdown report

fix(phase-3): wire translate-first state through markdown report (followup to a628b7a)

feat(phase-3): production-realistic agreement toggle

feat(phase-3): wire stage-1 classifyAll into ResultsList + fit dot in TriageRow

fix(simplifier): replace bare trailing labels with directive instructions

feat(phase-3): register Gemma 2 2B q4f16_1 as a faster alternative

revert(phase-3): remove gemma_fast — q4f16_1 broke simplifier output

revert(simplifier): restore PLAIN-LANGUAGE OUTPUT: label

fix(phase-3): defer simplification until classification completes + raise frequency_penalty

feat(phase-3): pipeline progress caption in detail pane

feat(phase-3): drop fit narrative section + add oncologist disclaimer

feat(phase-3): two-tier disclaimer + carry redesign hot-fixes

chore(phase-3): remove disabled sort chips from toolbar

chore: post-merge follow-ups (matchMedia, regression test, doc move)

fix(phase-3): re-run classification when patient description changes

fix(phase-3): address PR #2 review must-fixes

chore(phase-3): post-review cleanup + disable classify-in-results

chore(phase-3): clear lint diagnostics introduced by this PR