agentic-rl-wordle v1.0.0

Final main SHA: a04bc93c9b18b4a607d584c2cfe001a899d668ce

agentic-rl-wordle v1.0.0

Status: stable source release; the annotated v1.0.0 tag and GitHub Release are published.

This is a source-only portfolio release. The GitHub Release must have zero additional assets:
no model weights, word lists, raw episodes, training outputs, wheel, sdist, or other binary artifact
is attached.

Portfolio result

Multi-turn GRPO changed the fixed evaluation from base 0/463 wins and no legal/protocol actions to
tuned 13/463 wins (2.81%), 2749/2753 protocol-adherent turns (99.85%), and 2748/2753 legal actions
(99.82%). The base Wilson 95% CI is 0.00%–0.82%; the tuned interval is 1.65%–4.74%. The two-sided
exact paired McNemar value is 0.000244140625; the conservative two-look Bonferroni-adjusted value
is 0.00048828125.

Protocol learning succeeded; strategy learning remained limited; the 2.81% win rate is not a practical Wordle solver. This release does not claim practical capability, full strategy
learning, production readiness, general RL superiority, or complete exclusion of reward
exploitation.

Evidence

• Immutable research/evidence source commit:
  1a077a45e309594e5bb43743a8b84d89155595d4
• Evidence source URL:
  https://github.com/kuotunyu/agentic-rl-wordle/commit/1a077a45e309594e5bb43743a8b84d89155595d4
• Stable release URL:
  https://github.com/kuotunyu/agentic-rl-wordle/releases/tag/v1.0.0
• Aggregate report: results/full_463_report.json
• Deterministic recomputation: results/full_463_analysis.json
• Claim traceability: docs/claim-matrix.md
• HF pre/post inventories and README-only transaction receipt: docs/huggingface-audit.md

The research/evidence source commit is not the final release commit. The annotated tag and GitHub
Release metadata bind the final source identity to this published closure.

Model repository coordinates

• LoRA adapter: steven0226/qwen2.5-1.5b-wordle-grpo
  • approved pre-update revision: ef1e98ce214921049b86dce7c104c88875130023
  • verified post-update revision: e95fc44d5914d800483a847e8768b86f33719f12
  • authoritative README SHA-256:
    ab9c473a8eb9efaf2ddc32873d405bb5eb6b5e305a0dddcdf774f8b7a77a0e6b
  • immutable adapter weight LFS SHA-256:
    92e6379ed7ddf363e7f500b143afa7a2dc725d3e86bd87bc9eb933831c7d68b7
• Merged full model: steven0226/qwen2.5-1.5b-wordle-grpo-merged
  • approved pre-update revision: a59a4fb4c26e5d0612ce3a3574193ec58d46fc64
  • verified post-update revision: 94c0524cf963f7b22f1dc253eda5b4ef5a075956
  • authoritative README SHA-256:
    b9f67c27188839000385ec85900a1d6825157aec1516af883349b7e52efb8e47
  • immutable merged weight LFS SHA-256:
    b6c55086e798e1f62e6d970f07ee97ab39c1e0af3ee4b6ecdb2a349e485087af

Both complete eight-file inventories were re-read after the transaction. Only each README.md
blob and size changed; every non-README blob, ordinary size, LFS SHA-256, and LFS size remained
identical. No weight, tokenizer, configuration, or training artifact was changed.

Limitations

• The committed 463-game evaluation evidence is aggregate-only; full per-episode records are unavailable.
• The historical GPU environment is not bit-for-bit reconstructable.
• The exact upstream Qwen commit and an end-to-end run→code→prompt→bundle→model identity chain were
  not preserved.
• The adapter-to-merged relationship is documentary lineage, not complete cryptographic proof;
  no merge command or manifest proves the current merged bytes from the current adapter bytes.
• The cfreshman word lists are fetch-only and have no explicit license. Apache-2.0 does not license
  those third-party lists.
• The 463-word result is the adapter evaluation. No independent evaluation was run against the
  merged bytes, so it is not an independent merged-model replication.

Publication boundary

The stable closure includes a green exact-head PR, verified README-only HF receipts, a later Git
evidence-closure commit, green final-main and tag CI, an annotated v1.0.0 tag, and a non-draft,
non-prerelease source-only GitHub Release with zero additional assets.