SweetGold v1.2.0

SweetGold v1.2.0

English | 简体中文

SweetGold v1.2.0 completes the current feature-development cycle with an
auditable strategy-decision workflow, hardware-portable ML execution, equal
English and Chinese project documentation, and a public launch site.

Highlights

• M15 runs a matched-seed Arena league, applies a declared balanced, yield,
  or safety objective, enforces survival and invalid-action constraints, and
  writes linked JSON and Markdown decision evidence.
• M16 adds explicit auto, cpu, mps, and cuda selection to ML commands,
  refuses silent accelerator fallback, and records hardware in run manifests.
• Accelerator-synchronized BC timing and throughput make hardware decisions
  evidence-based; M6 BC+PPO and M7 CTDE smoke workflows pass on Apple MPS.
• English and Chinese READMEs, handoff, product design, changelog, release notes,
  model cards, and operational documents now have equal entry status.
• The bilingual project site is continuously deployed to
  https://alanthssss.github.io/sweetgold/ through GitHub Pages.

M15 auditable decisions

python3 main.py arena-agent \
  --strategies assignment greedy scout \
  --objective balanced \
  --min-bee-survival 0.9 \
  --max-invalid-action-rate 0.01 \
  --episodes 10 --seed 42

The decision is deterministic for the same artifact and policy. An explicit
“no eligible strategy” result is preserved when every candidate violates a
constraint. Small leagues demonstrate the workflow and do not replace the M14
cross-distribution robustness audit.

M16 hardware portability

.venv-ml/bin/python main.py hardware --device auto
.venv-ml/bin/python main.py pipeline --config experiments/smoke.json --device mps

On the initial fixed 20-episode, three-epoch BC smoke workload, the Apple M1 Pro
CPU trained in 0.5774 seconds at 138,355 examples/s; MPS trained in 3.6764
seconds at 21,730 examples/s. This small workload did not amortize accelerator
dispatch, so AWS CUDA is deferred until profiling shows a suitable workload.

Models and compatibility

v1.2.0 promotes no new policy and publishes no new checkpoint. M14
hierarchical-return-ctde remains the latest promoted policy and continues to
use the verified models-v2
asset catalog. Existing model IDs, registry records, Arena artifacts, and CLI
commands remain compatible.

Verification

python3 main.py --version
python3 -m unittest discover -s tests -v
python3 main.py models download
python3 main.py models verify

The source release uses Apache-2.0. Separately distributed model assets retain
their declared Apache-2.0 terms and model-card limitations.

Project status

After v1.2.0, new feature development is paused. Maintenance covers critical
defects, security, compatibility, reproducibility, documentation, and release
operations. A future CUDA run or research milestone requires a declared need,
budget, acceptance gates, and untouched evaluation seeds where applicable.

SweetGold v1.1.0

SweetGold v1.1.0

SweetGold v1.1.0 turns the initial research platform into a reproducible
strategy-comparison and model-distribution workflow.

Highlights

• Downloadable, integrity-checked promoted checkpoints with model cards.
• Bilingual model registry cards and one-click local installation.
• Matched-seed multi-strategy leagues with transparent rankings.
• Durable JSON evaluation artifacts for browsers, scripts, CI and AI agents.
• M14 hierarchical return control, the first policy to pass the full
  cross-distribution robustness gate.

M14 result

The selected high-level supervisor uses safety margin 6 and recharge fraction
0.8 above the accepted coordinated CTDE actor. On 50 untouched final seeds in
each of six scenarios it achieved 100% bee survival and 0% invalid actions
everywhere. Median honey reached 148.47% of Assignment and the worst ratio was
101.16% under scarce nectar.

Model catalog

The accompanying
models-v2
release contains all promoted checkpoint names, Apache-2.0 terms and a
machine-readable manifest. M14 intentionally shares neural weights with M8;
its new behavior comes from versioned supervisor code and registered
parameters.

Upgrade and verification

python main.py --version
python main.py models download
python main.py models verify
python -m unittest discover -s tests -v

No existing model ID or command is removed. Evaluation artifact APIs and the
arena-league command are additive.

SweetGold promoted models v2

SweetGold promoted models v2

models-v2 is the second immutable SweetGold model catalog. It adds the first
cross-distribution accepted policy, hierarchical-return-ctde, while retaining
the two v1 checkpoints for complete, clean installations.

New promoted policy

The M14 policy combines the accepted coordinated CTDE actor with a
deterministic high-level return, deposit and recharge supervisor. On 50 fresh
final seeds in each of six declared scenarios it achieved:

• 100% bee survival and 0% invalid actions everywhere;
• median honey equal to 148.47% of Assignment;
• worst-case honey equal to 101.16% under scarce nectar;
• harsh-weather honey of 141.46 versus Assignment's 99.48.

The selected supervisor uses safety margin 6 and recharge fraction 0.8.

Shared weights

hierarchical-return-ctde.pt and coordinated-ctde.pt intentionally have the
same SHA-256 digest. M14 did not retrain or silently mutate the neural actor.
Its new policy identity is the shared checkpoint plus the versioned supervisor
code and parameters in the registry.

Assets and verification

The release contains all three promoted checkpoint names, the Apache-2.0
license and models-v2-manifest.json. The manifest declares exact byte sizes,
SHA-256 digests, model-card paths and controller parameters.

Use SweetGold to download and verify registered artifacts:

python main.py models download
python main.py models verify

SweetGold promoted models v1

SweetGold promoted models v1

This release distributes the first two SweetGold checkpoints that passed their
predeclared promotion gates. The weights are copyright 2026 alanthssss and
licensed under the Apache License 2.0.

Assets

Asset	Registry ID	Size	SHA-256
bc-ppo.pt	bc-ppo	78,025 bytes	441e1770146962dd963df6f1b699c57b186a03c043e6d691a358f51911158f1d
coordinated-ctde.pt	coordinated-ctde	68,065 bytes	86fd605a1f013638ea89e95e1a71b65c0da243c003b7b22cf90e27dfe4397a68

The repository source archive contains the complete Apache-2.0 license and
model cards. The committed registry is the authoritative mapping from model
identity to asset URL, local artifact path and digest.

Install

From a SweetGold checkout at or after the M13 model-distribution milestone:

python3 main.py models download
python3 main.py models verify

Optional PyTorch dependencies are required to execute the learned strategies,
but not to download or verify them.

Scope

bc-ppo and coordinated-ctde passed their original in-distribution
promotion gates. M10–M12 demonstrate that neither should be described as a
generally robust or production control policy. See the linked model cards and
audit registry for their intended uses and limitations.

SweetGold v1.0.1

SweetGold v1.0.1

SweetGold v1.0.1 is the first release distributed with an explicit open-source
license. The source code, configuration, web UI and repository documentation
are available under the Apache License 2.0.

What changed

• Added the official Apache License 2.0 text.
• Added NOTICE with Copyright 2026 alanthssss.
• Documented the license in the README.
• Clarified that separately distributed model checkpoints and datasets may
  have their own license terms.

Compatibility

This patch release does not change SweetGold behavior, experiment results,
model promotion decisions or file formats. It contains the same v1.0 product
and research platform with explicit reuse, modification and distribution
terms.

Verification

python3 main.py --version
python3 -m unittest discover -s tests -v

The Apache License text is verified against the official Apache Software
Foundation copy before release.

Upgrade notes

Users of v1.0.0 should move to v1.0.1 when redistributing or building on
SweetGold so that the applicable license and attribution notice travel with
the source archive.

SweetGold v1.0.0

SweetGold v1.0.0

SweetGold v1.0.0 is a reproducible, interactive multi-agent AI experimentation
platform. It connects deterministic simulation, rule baselines, imitation
learning, reinforcement learning, model promotion, failure audits and a
side-by-side product experience.

Highlights

• Compare two strategies from the same initial world and stochastic seed.
• Replay every server-side frame without changing the live match.
• Train BC, PPO and CTDE policies through configuration-driven pipelines.
• Reject seed leakage before experiments begin.
• Promote models only when confidence, yield and safety gates pass.
• Preserve failed robustness results in a structured audit registry.
• Run the simulation and rule baselines without third-party dependencies.

Start the Strategy Arena

python3 main.py play --port 8080

Open http://127.0.0.1:8080.

To enable locally generated learned strategies:

python3 -m venv .venv-ml
.venv-ml/bin/pip install -r requirements-ml.txt
.venv-ml/bin/python main.py pipeline-m8 \
  --config experiments/m8-coordination.json

The arena verifies registered checkpoint hashes before loading them.

What v1.0 claims

SweetGold demonstrates an end-to-end multi-agent experimentation workflow and
a scientifically honest promotion process. The accepted M8 model improved over
its local behavior-cloning baseline while eliminating resource-contention
invalid actions.

What v1.0 does not claim

The learned policy is not a generally robust agent. M10 and M11 found severe
survival degradation under harsh weather and reduced energy. Those candidates
were rejected, their results remain visible in registry/audits.json, and
their final seeds must not be reused for tuning.

The built-in server is intended for local experiments and demonstrations. It
does not provide authentication, multi-user isolation or production service
guarantees.

Verification

python3 main.py --version
python3 -m unittest discover -s tests -v
python3 main.py benchmark --episodes 30 --controllers greedy assignment

Optional ML smoke pipelines are executed by GitHub Actions.

Upgrade notes

This is the first versioned release. Generated runs/, models/ and data/
directories remain intentionally excluded from Git.