Thanks to visit codestin.com
Credit goes to github.com

Skip to content

SHIP_v710_CHECKLIST.md

Latest commit

 

History

History
154 lines (113 loc) · 6.46 KB

SHIP_v710_CHECKLIST.md

File metadata and controls

154 lines (113 loc) · 6.46 KB

styxx 7.1.0 ship checklist

Run-through for the actual phase 3 publish. All steps are commands flobi runs (or authorizes Claude to run with explicit go).

Pre-flight already complete:

  • styxx/reward.py shipped (261 LOC, cognometric reward signal)
  • styxx/synth.py shipped (BOUNDARY BREAK — synthetic preference-pair generator via inverse cognometry; 20/20 craft success, +0.839 mean delta, 20/20 round-trip ranking correct)
  • styxx/_demo_baselines.py shipped (strawman approval-style baseline)
  • tests/test_reward.py — 14/14 pass
  • tests/test_synth.py — 7/7 pass
  • full styxx test suite — 821/822 pass (1 skipped, 0 regressions)
  • data/cognometric_rlhf_demo_v0.jsonl — 20 curated triples
  • examples/cognometric_reward_basic.py — basic usage
  • examples/cogn_rlhf_divergence.py — divergence demo (cogn 17/20 vs approval 6/20)
  • examples/cogn_rlhf_divergence_colab.ipynb — Colab notebook
  • examples/synth_preference_pairs.py — synth pair-generator demo
  • examples/trl_ppo_integration.py — TRL PPOTrainer skeleton
  • release/cogn_rlhf_divergence_v0.json — saved divergence result
  • release/synth_preference_pairs_v0.jsonl — 20 synth-generated preference pairs
  • release/v710_release_notes.md — gh release --notes-file source
  • styxx/__init__.py — top-level fathom_reward, FathomRewardModel, craft_preference_pair, generate_preference_pairs exports
  • CHANGELOG.md — 7.1.0 entry with reward + synth sections
  • pyproject.toml — version bumped 7.0.0 → 7.1.0
  • dist/styxx-7.1.0-py3-none-any.whl — built and twine-checked (5.89 MB, includes synth)
  • fresh-venv smoke install verified (synth round-trip: chosen=0.112, rejected=1.000, delta=+0.888)

Phase 3 — publish (irreversible — explicit go required for each)

Step 1. Verify the wheel installs cleanly in a fresh venv

cd C:\Users\heyzo\clawd\styxx
python -m venv .venv-71-smoke
.\.venv-71-smoke\Scripts\Activate.ps1
pip install --upgrade pip
pip install dist/styxx-7.1.0-py3-none-any.whl
python -c "from styxx import fathom_reward, FathomRewardModel; print('install ok'); r = fathom_reward(prompt='You agree?', completion='Absolutely!'); print(f'reward={r:.3f}')"
deactivate

Expected: install ok + a reward print. If it fails, fix and rebuild before continuing.

Step 2. Git commit

16 changed/added files. Specify exactly to avoid git add . polluting:

cd C:\Users\heyzo\clawd\styxx
git add styxx/reward.py styxx/synth.py styxx/_demo_baselines.py styxx/__init__.py tests/test_reward.py tests/test_synth.py data/cognometric_rlhf_demo_v0.jsonl examples/cognometric_reward_basic.py examples/cogn_rlhf_divergence.py examples/cogn_rlhf_divergence_colab.ipynb examples/synth_preference_pairs.py examples/synth_multi_instrument.py examples/trl_ppo_integration.py release/cogn_rlhf_divergence_v0.json release/synth_preference_pairs_v0.jsonl release/synth_multi_instrument_v0.json release/v710_release_notes.md CHANGELOG.md pyproject.toml release/SHIP_v710_CHECKLIST.md
git status

Verify only those 16+ files are staged (also include any docs/ updates if you add them). Then commit:

git commit -m "7.1.0: styxx.reward + styxx.synth - cognometric reward + inverse-cogn synth

styxx.reward: first reward signal calibrated against cognitive failure
modes instead of human approval. Drop-in for trl PPO/GRPO/DPO trainers.

styxx.synth: synthetic preference-pair generator composing v7.0.0
inverse cognometry with the new reward signal. Recursive: fathom's
attack module generates training data for fathom's reward signal.

Results on curated 20-pair sycophancy benchmark:
  cognometric reward       17/20  (85%)
  approval baseline         6/20  (30%, below chance)
  inversions               13/20  (65%)

Synth pair-generation results (target_score=0.85):
  crafted with positive delta:  20/20
  reached saturation:           20/20
  mean delta:                   +0.839
  cogn_reward round-trip:       20/20 ranks chosen above rejected

Universal-perturbation moat: v7.0.0 perturbation lifts cross-fire by
+0.468 in attack mode but produces +0.000 lift on cogn-RLHF reward
(dominant instrument already saturated).

Top-level API:
  from styxx import fathom_reward, FathomRewardModel
  from styxx import craft_preference_pair, generate_preference_pairs

21 new tests, 821/822 full suite pass."

Step 3. Tag v7.1.0

git tag -a v7.1.0 -m "7.1.0: cognometric reward signal for RLHF"
git tag -l | tail -5   # verify tag created

Step 4. Push commit + tag

git push origin main
git push origin v7.1.0

Step 5. Upload to PyPI (irreversible — version locked forever)

cd C:\Users\heyzo\clawd\styxx
python -m pip install -U twine
python -m twine check dist/styxx-7.1.0-py3-none-any.whl
python -m twine upload dist/styxx-7.1.0-py3-none-any.whl
# enter PyPI token when prompted (or use ~/.pypirc)

Verify: https://pypi.org/project/styxx/7.1.0/ should show the new version within ~30 seconds.

Step 6. GitHub release

gh release create v7.1.0 dist/styxx-7.1.0-py3-none-any.whl --title "7.1.0: styxx.reward — cognometric reward signal for RLHF" --notes-file release/v710_release_notes.md

(Generate release/v710_release_notes.md from the CHANGELOG 7.1.0 section before this step.)

Step 7. Tweet

Post the thread or single tweet from research/eeg_pilot/v710_tweet_drafts.md (single tweet preferred for first-fire; thread version if amplification needed).

Step 8. Update fathom.darkflobi.com

The site mirrors styxx version. Bump the version reference + add a CHANGELOG link.

# In Desktop/clawd-clean/darkflobi-site/, update version refs
# Then deploy via: bash clawd/scripts/deploy-fathom-site.sh
# (per memory: NEVER netlify deploy --dir=. from .styxx cwd)

Post-ship monitoring (first 24 hours)

  • PyPI download count
  • GitHub stars delta
  • Twitter engagement
  • Any installation issues / bug reports

If a critical bug surfaces in the first 24 hours, yank with python -m twine yank styxx==7.1.0 --reason "..." and ship 7.1.1 with the fix.

Parallel non-publish work (already drafted, no go required)

  • research/eeg_pilot/arxiv_paper_outline.md — paper #1 outline, target submission 2026-05-28
  • research/eeg_pilot/outreach_drafts.md — 4 outreach emails ready to send (OpenBCI / Bitbrain / Pearl IRB / Advarra)
  • research/eeg_pilot/v710_tweet_drafts.md — single tweet + thread version