Reward function experiments: audited three-seed mechanism study

Completed nine-worker mechanism study

This release adds the frozen, independently audited A/B/C study across paired seeds 103, 211 and 307. Each fresh Qwen3-4B policy completed 64 groups of four trajectories and evaluated 96 previously untouched final tasks twice. The 34-page paper retains its derivations, earlier development study and all eight figures, and adds the executed training recipe, qualification, recovery and limitations.

Primary full-penalty B minus homogeneous-success-mask C: -29.564236 tokens (-1.804218% relative to C), descriptive 95% crossed seed/task interval [-76.662413, +15.480469]. Accuracy: +3.993056 percentage points, interval [-0.520833, +8.506944]. The registered promising token-cost response criterion was not met. Only three seeds and restricted synthetic tasks limit inference; a smaller effect is not ruled out. No adaptive-controller advantage is established.

Public scientific reconstruction

Includes all 2,304 training trajectories, 1,728 trained-policy final records, the one 192-record initial reference, frozen scientific source/configuration/tasks, raw numerical gates and independent audit receipts. CPU reconstruction checks complete original member hashes and reproduces recorded scoring, rewards/advantages/masks, update support, pairing and the registered bootstrap. The downloadable archive is the complete reviewed public package; verify its file manifest before analysis.

Actual model/optimizer/RNG binaries, encrypted operational archives, credentials, provider account/support records and the original private Git history are excluded. Recorded tensor hashes are audit receipts, not independently rehashed absent tensor bytes. No model or cloud connection is required for public reconstruction. Training-source availability does not authorize another paid run.

The paper, website visualizer and Work entry reference this same immutable version. Earlier releases and historical evidence remain preserved. Original code: Apache 2.0; original text/tasks/measurements: CC BY 4.0, subject to third-party notices.

Reward function experiments: research v1.0.1

The experiment repository is now arpjw/reward-function.

This release updates the repository identity, canonical links, citation metadata, experiment-focused README and paper PDF/LaTeX. It preserves the prior paper-v1.0.0 release. The mathematical derivations, raw scientific inputs and measured results are unchanged. Public CPU reconstruction and private tensor/operational archive boundaries remain the same.

Contains the refreshed 26-page PDF and complete reviewed research archive. This is a publication/link maintenance revision, not a new experiment result.

Reward contrast: public research release v1.0.0

Versioned public release of the completed exploratory study.

Contains the 26-page paper, expanded LaTeX derivations, frozen source, raw scientific records, integrity manifests and CPU analysis reconstruction. The analysis exactly reproduces the recorded diagnostics; four focused evidence tests pass.

Penalty: 54/64 at 1,567.890625 generated tokens. Reused control: 52/64 at 1,556.6875. Tokens increased 0.719677%; principal descriptive intervals include zero. One seed, reused inspected development tasks and differing deterministic runtime controls limit interpretation. No savings, reliable accuracy gain or generalization is demonstrated.

Fresh Git history excludes private account/billing/support records, credentials, model/optimizer/RNG tensors and complete encrypted operational archives. Scientific input hashes are unchanged. Public CPU analysis reconstruction is distinct from the historical full tensor/restart audit. No prospective sealed test data are included.