eval: replace saturated demo-fixture gate with a falsifiable synthetic battery (FIX-13)#32
Merged
ChelseaKR merged 2 commits intoJul 12, 2026
Conversation
…c battery (FIX-13) The merge-blocking eval gate ran on exactly one hand-built demo fixture (ingest/demo.py) and scored content/hybrid at a perfect 1.0 on every metric — trivially separable, so it could never catch a real regression short of catastrophe. recommender/synth.py adds a seeded, deterministic synthetic-library generator (random.Random(seed) only, no numpy): a noisy canon/distractor tag split, popularity anti-correlated with canon-fit, and a reader taste profile split between theme-only discoveries and an author-loyalty pathway. recommender/battery.py runs content/hybrid/popularity across 10 default seeds and gates on median MAP@5 uplift >= MARGIN (0.5, calibrated with real headroom on both sides) plus a no-losing-seed guarantee, and tracks two ablations per seed (drop curated lists; shuffle 20% of candidate tags) that never improve content MAP. ingest/cli.py's `stacks eval` now gates on the battery by default, writing docs/audits/eval-battery.json alongside the existing (now informational) demo-fixture report; --no-synthetic keeps the legacy path. tests/test_synth_eval.py proves the gate is actually falsifiable: zeroing recommender/model.py::AUTHOR_BONUS measurably narrows the median uplift and flips `passed` to False. docs/audits/source-ethics.md explains the distribution instead of citing the saturated 1.0, and docs/ideation/02-large-scale-fixes.md marks FIX-13 done. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
# Conflicts: # docs/audits/coverage.xml # docs/ideation/02-large-scale-fixes.md
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The merge-blocking eval gate ran on exactly one hand-built demo fixture
(ingest/demo.py) and scored content/hybrid at a perfect 1.0 on every
metric — trivially separable, so it could never catch a real regression
short of catastrophe.
recommender/synth.py adds a seeded, deterministic synthetic-library
generator (random.Random(seed) only, no numpy): a noisy canon/distractor
tag split, popularity anti-correlated with canon-fit, and a reader taste
profile split between theme-only discoveries and an author-loyalty
pathway. recommender/battery.py runs content/hybrid/popularity across 10
default seeds and gates on median MAP@5 uplift >= MARGIN (0.5, calibrated
with real headroom on both sides) plus a no-losing-seed guarantee, and
tracks two ablations per seed (drop curated lists; shuffle 20% of
candidate tags) that never improve content MAP.
ingest/cli.py's
stacks evalnow gates on the battery by default,writing docs/audits/eval-battery.json alongside the existing (now
informational) demo-fixture report; --no-synthetic keeps the legacy path.
tests/test_synth_eval.py proves the gate is actually falsifiable: zeroing
recommender/model.py::AUTHOR_BONUS measurably narrows the median uplift
and flips
passedto False. docs/audits/source-ethics.md explains thedistribution instead of citing the saturated 1.0, and
docs/ideation/02-large-scale-fixes.md marks FIX-13 done.
Co-Authored-By: Claude Fable 5 noreply@anthropic.com
Backfill PR for a completed roadmap item (
roadmap/fix-13-falsifiable-recommender-eval-synt, 1 commit(s) overmain). Part of the portfolio roadmap batch.