Skip to content

docs: refresh AgentSkill for Champion Challenge redesign#556

Open
catoneone wants to merge 1 commit into
AffineFoundation:mainfrom
catoneone:update/affine-skill-2026-06-redesign
Open

docs: refresh AgentSkill for Champion Challenge redesign#556
catoneone wants to merge 1 commit into
AffineFoundation:mainfrom
catoneone:update/affine-skill-2026-06-redesign

Conversation

@catoneone

Copy link
Copy Markdown
Collaborator

The skill/references/*.md still described the removed 4-stage ELO/Pareto pipeline, ~13 environments, and Chutes deployment. This refreshes the skill docs to match the current system (verified against e3a9fce).

Changes

  • skill/SKILL.md — add concise sections (in the existing project-notes style) for:
    • Model Requirements (Qwen3-32B / Qwen3.6-35B-A3B qwen3_5_moe, quantized/gated bans, one-commit-per-hotkey)
    • Scoring Model (Champion Challenge: partial-Pareto dethrone rule, winner-takes-all, comparator margins)
    • Active Environments (SWE-INFINITE, NAVWORLD, MEMORY, TERMINAL; DISTILL-V2 shadow)
    • Anti-Copy (CEAC) — decision-position logprob fingerprint; replaces the old cosine detector
  • references/scoring-deep-dive.md — rewritten for Champion Challenge (challenger queue, battle window, 95% completion, dethrone rule, weight emission).
  • references/rl-training-guide.md — 4 scoring envs, allowed architectures, one-commit rule; drop geometric-mean framing.
  • references/database-schema.md — 11 tables: drop task_pool/anti_copy_results, add inference_endpoints + anticopy_* (CEAC).
  • references/api-reference.md — serving base_url instead of Chutes slug; note scorer logic folded into the scheduler.

All facts verified by reading current source (comparator.py, challenger_queue.py, weight_writer.py, anticopy/, model_size_checker.py, system_config.json, docs/MINER.md, docs/VALIDATOR.md).

🤖 Generated with Claude Code

Update skill/SKILL.md and skill/references/ to match the current system
(verified against e3a9fce). The reference docs still described the removed
4-stage ELO/Pareto pipeline, 13 envs, and Chutes deployment.

- skill/SKILL.md: add concise sections for model requirements
  (Qwen3-32B / Qwen3.6-35B-A3B, one-commit-per-hotkey), the Champion
  Challenge scoring model, the 4 active scoring envs, and CEAC anti-copy.
- references/scoring-deep-dive.md: rewrite for Champion Challenge
  (challenger queue, battle window, partial-Pareto dethrone rule,
  winner-takes-all + optional past-N split).
- references/rl-training-guide.md: 4 scoring envs, allowed architectures,
  one-commit rule; drop geometric-mean framing.
- references/database-schema.md: 11 tables; drop task_pool/anti_copy_results,
  add inference_endpoints + anticopy_* (CEAC).
- references/api-reference.md: serving base_url (not Chutes slug); scorer
  logic folded into the scheduler.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant