RDS 5.9.0 Preview 1

RDS 5.9.0-rc.1 preview

This prerelease is a trial of the current main line after 5.8.0. It does not
include the open PR #189
for optimizing reusable proof plans. Use the release page's source and host
packages to get the exact published revision.

Preview scope

• Continue bounded project work from the same ledger after a pass yields, while
  preserving the original goal, contract, reservations, costs and evidence.
• Request model repairs with bounded, hash-checked task-input and original-output
  excerpts. The shared excerpt allowance is limited to 64 KiB; this is not the
  full task, dataset or execution environment.
• Offer an optional evidence-bound working set assembled from owned project
  records, goals, budgets and missing evidence, reducing irrelevant context
  without treating retrieved material as verified fact.
• Retain alternative problem structures, their dependencies and rival
  predictions, then use verified feedback to activate, reject or roll back a
  candidate and replan within the declared budget.
• Route supported exact-math requests through bounded tools and count a result
  only after the registered independent checker accepts its certificate.
• Preserve oversized or incomplete outputs as unresolved evidence rather than
  treating a successful process exit as a correct research result.

These are engineering capabilities with declared scopes. The structural
exploration work has not established that a model independently discovers a
better research strategy. No end-to-end scientific capability gain is claimed;
the matched-model evaluation remains in Issue #166.

Trial status

This is a release candidate, not a stable compatibility promise. The release
page identifies the exact tagged commit and its CI run. CI establishes only the
software checks that ran; research benefit, unsupported domains and unmeasured
resource costs remain unknown.

Published revision and downloads

• Tag: v5.9.0-rc.1.
• Commit: b52f7582b4e6e10f0c4de33b9a484fa215f163e0.
• Exact-commit main CI: all 7 applicable jobs passed.
• Downloads: source ZIP and Claude, Codex, and OMP harness bundles; acceptance record and SHA256 checksums are included below.

RDS 5.8.0

RDS 5.8.0

Stable compatible feature release from main 4b26a182e1d36a07176916feb2ebc547306e0243.

Delivered behavior

• Program-owned goal/dependency state, original-goal review, bounded alternative/next-move selection, and owned Advisor execution preserve declared inputs, budgets, receipts and recovery boundaries.
• Conditional theory-tool lookup and eight runnable operator scaffolds include supported bounded finite-model→EGraph→native-Lean composition. Each stage retains its actual assurance and checked transport; unsupported or missing evidence stays UNKNOWN.
• Artifact and receipt identity/integrity checks reject aliases, tampered bodies and invalid observations before they can count as goal support.
• CLI/Skill entry and checkpoint/recovery improvements preserve real attempt and cost history. Owned Advisor action/protocol admission now rejects unusable frozen contracts at initialization, with actionable binding diagnostics.
• Recent repairs cover valid checkpoint review, identical concurrent dependency publication, contract reads after math binding without project initialization, and automatic usage fallback for accepted abbreviated/attached root options.
• CHANGELOG.md was intentionally removed; release notes carry delivered behavior and verification, and exports/development snapshots no longer require that file.

Installation and compatibility

Assets include source plus separate Claude, Codex and OMP plugin archives. Each plugin archive retains its native host manifest and bundled Skill/resources; see the included installation guides for the host's supported loader. SHA256SUMS.txt binds published assets. CLI5.8.0, Skillv5.8.0 and enginerds-cli-v5.8 are aligned.

Six real v5.7.0-writer synthetic ledger reading paths passed with all original bytes unchanged. This MINOR release preserves those supported legacy read paths; it does not guarantee every private historical schema. Source/environment binding changes still obey the execution contract guard.

Verification

Exact-main CI37122874906 passed all seven jobs. Four platform/Python suites each ran1289 tests, with23 Ubuntu/16 Windows skips; native parity and Lean/statistical coverage passed separately. Direct main acceptance supplied the positive native progression skipped by ordinary CI: four declared pairs checked by Lean, plus seventeen boundary tests, no failures/skips. Actual relocated archives passed all twelve bounded CLI calls, including public empirical and mixed examples; old-ledger compatibility and the six-test package suite also passed. See RELEASE-ACCEPTANCE.md and ci-case-map.json for scopes and skip boundaries.

Known limitations and follow-up

• #108 remains OPEN. A hosted Windows concurrent fallback-start test previously reported SQLITE_BUSY/start record not confirmed despite twelve successful command exits. The test stopped before physical-row assertions; missing-row count and trigger remain UNKNOWN. Later green CI does not establish a fix. The owner explicitly accepted this as a release limitation; incomplete usage logging must not be treated as a complete count.
• #120 retains host execution coverage outside supported RDS entries. Package manifests/hooks do not establish universal enforcement of arbitrary agent commands.
• Optional NumPy/PyTorch, native Lean and pinned Formal dependencies have explicit availability limits. Their absent/skipped cases are not passes. Native adapters retain bounded budgets and can return UNKNOWN on timeout; a finite checked example is not a general symbolic proof pipeline.
• Engineering tests and examples do not establish causal scientific gain, research autonomy, universal measurement independence or RSI policy improvement. Example outcomes remain scoped development evidence.

Original failures, receipts and UNKNOWN evidence are preserved. Issue #53's retained supported composition scope is accepted; withdrawn additional operator proposals are not revived or called delivered.

RDS v5.7.0 — Scoped exploration, resource planning and native Lean gates

RDS 5.7.0 adds compatible research-workflow capabilities to the changes published in 5.6.0-rc.2: scoped exploration control, bounded theory work, native formal gates and resource-aware batch advice.

Release revision: bcff923c89411db95e7a3a6bb328e51e50a3a86a on main.

All five checks for this exact release revision passed: Windows and Linux on Python 3.11/3.13, plus native Lean build, axiom audit and backend/statistical replay.

• Scoped exploration and evidence: Advisor consumes durable rejection checkpoints, distinguishes diagnostic coverage, and records evidence-led advancement. Failed or unknown attempts retain their evidence and costs.
• Theory before execution: bounded formal checks account for their own budget before empirical admission. Unresolved application premises remain blocked.
• Native Lean verification: pinned Lean/Mathlib 4.33.1 builds contraction, Gershgorin, scale invariance, Ville, Hoeffding and data-processing proofs. CI requires native backend/statistical replay and audits library axioms.
• Honest backend provenance: --tactics lean4 requires native Lean; missing dependencies return UNKNOWN. --tactics rational performs exact Python rational verification with separate backend and assurance labels.
• Resource planning: supported capacities, priorities, dependencies and protected budgets inform proposed parallel batches. The planner advises; it does not allocate physical GPUs or launch jobs.
• Version policy: compatible features advance MINOR, fixes advance PATCH, and incompatible public behavior advances MAJOR. Published tags are retained. See the versioning policy.

Validation of the feature tree: all ten candidate checks passed for 4eb15bf0c38909de52105b8955bdbd30e8a8bb74; all five merged-main checks passed for 3c4fbca8d6ed62e756a5b54326a6335b61989017. Native CI executed 42 backend/statistical checks, with no skips. The actual RDS self-development loop executed 521 passing checks, five documented environment skips, zero failures (526 total), and imported the original results through Advisor, rule replay, isolated adoption and rollback. A new RDS self-development iteration also executed 14 concurrency/compatibility checks for the 5.7.0 update, with no failures or skips, preserving readable rc.2 ledgers and the stale-engine execution guard.

Native statistical replay verifies conditional theorems; it does not prove that an arbitrary experiment satisfies their premises. Automatic empirical premise closure and statistical early stopping remain future work. These engineering checks do not establish research-policy gains. The original ExplorationBench evaluation has not run; author access is pending.

Use the GitHub source archives or clone this repository and check out v5.7.0. Existing ledgers remain readable; create a new execution contract when upgrading changes the bound engine hash. The earlier rc tags remain available.

Full comparison: v5.6.0-rc.2...v5.7.0.

v5.6.0-rc.2 — Advisor budget and background execution fixes

Fix three practical issues found during local scientific benchmark preparation and real GPU execution:

• Advisor resource identity: CPU and GPU costs are no longer ordered or checked against another resource's budget merely because units match. Mixed or unknown resource identities remain incomparable.
• Upgrade compatibility: v5.6.0-rc.1 ledgers remain readable in rc.2; stale execution contracts still cannot authorize new work.
• Background completion: settled Task Scheduler attempts and receipts now record their terminal RDS lifecycle, preserving task IDs and startup/recovery guards.

Validation: 301 passed, 4 optional-dependency skips, 0 failures (305 total) in the actual RDS self-development loop, including original process receipt, record import, Advisor inspection and isolated finite RSI replay/adoption/rollback. Two skipped checks need configured native Lean; two need the optional PyTorch dependency. An actual Windows Task Scheduler CPU acceptance job also passed, with run/scheduler COMPLETED and receipt SUCCEEDED.

PR #10 passed all eight candidate push/PR checks across Windows/Linux and Python 3.11/3.13. README files are unchanged. The experimental two-fidelity Advisor planner remains in draft PR #9 and is not included in this release.

The four checks on merged main also passed. This tag targets 1f1554736bd4e2c926340c6a5de202324c5fbb40, whose Git tree matches the tested candidate.

Clone the repository and check out v5.6.0-rc.2 to use this prerelease. Native GPU task execution has now been exercised locally, but passing software tests does not establish improved scientific research policy.

First real task check: the original MLGym CIFAR-10 L1 task, with its unchanged official grader on native Windows, produced 87.42% for the plain control and 90.59% for the crop+flip arm (+3.17 percentage points). Both arms used seed42, 80 epochs, the same initial weights/data order/14,080 updates, and validation-only checkpoint selection before final grading. The task's published threshold is 49.71%; it is not the matched local control.

Actual RDS tools executed the training, imported original records, inspected Advisor suggestions, froze the models, executed official grading and imported the derived paired gain. GPU-task process-wall accounting was about 314 seconds, including a 15-second reserve for earlier failed/runtime probes. Total conservatively accounted execution, including failed scoring and CPU evaluation, was about 347 seconds; download and human work are outside those figures. Original local logs and receipts retain a cuDNN environment failure and a byte-decoding scoring-wrapper failure. This is one original task result, not the full MLGym agent-harness score or a measured independent Advisor/RSI policy gain.

Research Direction Selector v5.6.0-rc.1

Research Direction Selector v5.6.0-rc.1 connects the first M01–M06 tool workflow and repairs issues found while using RDS to develop RDS.

• Import original experiment records into Advisor with source locations, conflicts and unknowns; compose finite experiment candidates and retain resource units.
• Execute locked project commands with logs, receipts and recovery. Reuse controls only when their arm, command and protocol match; treatment receipts and unknown identities cannot become cached controls.
• Restore usable reference run IDs and validate current inputs. Failed and timed-out reference executions return a failing exit code.
• Read the actual project ledger in dashboard exports. Distinguish measured usage, unknown usage and charged estimates; render Advisor actions, evidence requests, costs and outcome-dependent decisions directly, with original JSON available in folded details.
• Reject conflicting Advisor options and report corrupt explicit literature libraries. Preserve UTF-8 paths on Windows and normalize Windows short-path aliases in tests.
• Evaluate rule candidates before adoption, retain rollback records, and support the full suite in the RDS self-development executor.
• Repair four independently reproduced errors: duplicate action IDs masking READY rules, receipt identity conflicts accepted as observed costs, decisions attributed to the wrong branch, and recovery interrupting a newly starting live process. Validate malformed Advisor candidates before HTML export.
• Provide separate English, Simplified Chinese and Japanese READMEs in the agreed framework, separate English/Chinese contribution guides, a static Penrose SVG banner and optional scientific preferences.

Validation: 298 cases — 294 passed, four optional-dependency skips, zero failures. The actual RDS project executor ran the full suite, imported its results, inspected Advisor output and exercised isolated rule adoption plus rollback. All eight dashboard regression tests passed; Chromium checked 96 snapshot views, four actual Advisor views and six export-validation views with no page errors. Six historical replay checks passed.

The four independently reported defects were rechecked by Astra in fresh CLI workspaces and no longer reproduced. The Windows/Linux and Python 3.11/3.13 CI matrix passed. This release is published from main after merging PR #4, PR #6, PR #7, PR #5 and PR #8.

The project-runner acceptance is CPU-based. Native Lean and PyTorch checks need their optional local configuration or dependency. End-to-end scientific-policy gains remain unmeasured. No GPU training, paid model API or multi-seed campaign was run.

See CHANGELOG.md and the development loop for scope and commands.

Research Direction Selector v5.5.0-rc.2

Research Direction Selector v5.5.0-rc.2

Research Direction Selector 自动化辅助人类科研：把研究问题转成可检验实验，保存真实证据，再决定下一步。面向真人及真人交给 AI 使用，固定五组件：Skill、执行与验收内核、研究状态与记忆、Advisor、RSI。

本版从 main 发布，已合入数学执行内核和 Advisor/科研工作台。展示名称使用全称，仓库现已按所有者确认改为 kongtou20070406/research-direction-selector。三语 README 首页新增实测成绩、原始结果和优势示例，指南同步 main 已实现的接口。

2026-09-30 后续 main 文档更新（83c6986a71d9323ffdf5f2e2b0d8a91951547df4）：采用 Kramer 等人的已有科研自主框架，当前对应局部 L2，近期目标是全环节 L2，长期研究方向为 L4／L5；优先真实项目的体验与实用，横向比较暂缓。具体改法、交付物与通过条件见开发计划。本 release 的源码 tag 仍固定为下方发布提交，后续文档在 main，没有改写旧 tag。

已公开验证	实际结果	能说明什么
回归检查	208 项：204 通过，4 可选跳过	已覆盖功能保持正常；跳过项为本轮未配置原生 Lean、未安装 PyTorch的检查
历史检查	6/6 通过	已知失败与决策边界的有限回放
合成攻击场景	4/4 通过	四个已声明的攻击条件被处理
公开任务改编组件挑战	8/8 案例、28/28 合同检查通过	这组手工适配条件下，建议随合同、来源、依赖和预算变化
同一组件挑战处理时间	83.0743 ms	已保存的单次本地元数据 fixture 处理；不代表真实科研任务速度

可以从逐项结果直接复核的行为：单个 loss 记录保持缺证据；目标合同从 20 项变为 10 项时候选随之改变；删除事实来源后生成证据查询；未知成本不被当作可承担；缺少前置完成记录时不跳过依赖；预算为零时阻止有成本候选；不同 capsule 身份阻止错误归属。测试报告、逐项原始结果

公开任务改编分数不是 ScienceAgentBench/CORE-Bench 的原始端到端分数；案例是 development fixtures，未测量科研建议质量、对其他系统的领先幅度或实际 GPU 节省。本轮没有 GPU、付费 API 或多 seed 实验。Lean 数学义务只覆盖已声明的窄范围，数学证明不代替训练或机制证据。

同时恢复三处判断图措辞：只有已观察到可能改变决策的 seed 不稳定性，才考虑有预算的追加。版本升级后仍可读取 rc.1 账本；旧执行契约绑定旧代码，新执行需新契约，相关兼容回归已扩展。

旧 tag v5.5.0-rc.1 保留。下载本版本的 GitHub Source code，或按 README 使用 v5.5.0-rc.2 安装。

发布提交：b0689fd618cba67414ea2de2ad380242d8bacf23，位于 main 的历史中。发布时 GitHub Actions 的 Windows/Ubuntu × Python 3.11/3.13 四组回归与历史检查均通过。

RDS v5.5.0-rc.1 — 科研工作流与证据驱动 Advisor

RDS v5.5.0-rc.1

RDS 自动化辅助人类科研：把研究问题变成可检验实验，保存真实证据，再决定下一步。面向真人和真人交给 AI 使用，固定五组件：Skill、执行与验收内核、研究状态与记忆、Advisor、RSI。

本版整合：

• 三语 README、真人入口、可复制 AI 指令、组件职责与未来计划。
• Advisor 的结构化诊断与有界判断图搜索：带来源事实、三值条件、依赖组合、竞争解释、结果分支和成本偏序。五条规则有可执行绑定；缺失证据与成本仍是未知，不执行训练。
• 9 条有适用范围与原文定位的文献记录；导入摘录保持未审查，保留 WAL、去重和显式旧库迁移。
• 只读离线科研工作台，以及保留依赖和结构化规则的无依赖图读取/保存。
• PR #2 的数学义务、独立证书、有限原生 Lean4 闭合有理数检查与事务修复。

测试：208 项回归（204 通过、4 可选跳过：原生 Lean 未配置、缺少 PyTorch）；6 项历史检查；4 个合成攻击场景；8 个公开任务改编组件挑战、28 条检查全部通过。解压发布包后再次运行 8 个挑战，全部通过。

公开任务组件检查不是 ScienceAgentBench/CORE-Bench 的端到端分数，未测量科研建议质量或真实 GPU 节省。没有运行 GPU、付费 API 或多 seed 实验。示例 HTML 是明确标注的合成介绍，不含真实实验结果。

这是从整合提交 5d7001e 发布的预发布版；开放的 PR 尚未合并到 main。安装时选择本 tag，而不是 main。旧账本可以读取，但新执行需要与新代码绑定的契约。

发布后跟踪 GitHub CI、tag、安装包与入口到北京时间 2026-09-30 20:00；异常时修复并发布新的候选版本，不覆盖已有 tag。