Repository navigation
task: measure whether InferRSI helps Claude Code and Codex optimize inference tasks #248
Description
Current state: The bounded CPU KV slice software is merged (seven PRs from #274 to #290), but both live A2/A3 attempts on 2026-10-31 ended INCONCLUSIVE, the second stopping on HTTP 403; no third attempt is allowed.
Decision needed: none
Next action: The existing slice implementer or an authorized account operator retrieves the 2026-10-31 rejection and billing records without sending test requests; the broader proposal has no assignee.
当前执行状态（有界切片）
- 进展： 有界 CPU 任务与判据 #274 已合入，云端无模型资格验证 PASS；首份原始证据已固定保留。这是合成任务/控制资格，不是模型或知识效果验证。
- 范围： 当前切片由 owner 侧工作承接；下方原五臂、双 harness、加速器与整体验收方案仍保留为提案，原请求者与整体所有权字段不变。
- 软件交付： 受控编辑/测量、离线准入与脚本化 actor 资格 #275 已合入
19c3df83（审查 head5af35536），在既有修复基础上补齐最小执行证据记录，并修正凭据回显检查的 JSON 编码一致性，Ready 五个作业均通过，新 actor receipt 为 PASS；合入前 head 的 native review已完成且无 findings。原809b0a03的整轮 CI 失败及actor PASS receipt仍保留。这仍是软件与无模型脚本控制路径，不含 live launcher 或模型调用。 - 执行器交付： 执行器与有界证据留存 #276 已合入
ffb77f9a，Ready 五个作业均通过。该次交付时 locator 禁用、无凭据绑定；后续单次激活配置 #280已合入。软件/配置交付不代表真实对照成功或知识效果验收。 - 首轮实际执行： 首次运行已结束；A2 首个请求槽位已消费，但这不证明请求已获准或 HTTP 已发送；未记录正常完成且 usage 未知，A3 未启动，没有最终候选产物。两组最终判定均为 INCONCLUSIVE。workflow SUCCESS 仅表示编排与证据留存作业结束，不是模型任务成功或知识效果证据。
- 诊断软件补充： 失败诊断留存修复 #281已合入，保留后续失败的固定原因码、有界传输调用计数和执行器状态；它不能恢复首轮缺失的根因，也不证明 HTTP 送达、计费或任务成功。原失败、未知实际费用、已到期冻结及原冻结的不重试约束均保留。
- 后续软件： 单次追加准入软件 #287 已合入，Ready 与无模型资格验证已通过，原始 task/actor 资格包及匹配收据已固定归档；范围限于 PR275 的 $350 生命周期例外与一次同规格 8+8 追加对照的准入控制，保留首轮完整 $162.72 claim，其余 PR 默认上限及日/试点上限不变。旧价格/账本快照对应的 10:00 UTC 窗口已到期，准入截止前未启动新对照，历史保留。新 v2 冻结及价格/账本刷新对应的单次追加已执行；原 12:00 UTC 到期及 11:38 UTC 最晚准入不变。激活配置 #290已合入，追加运行通过同机前检并获准入，但 A2 首次传输调用记录 HTTP 403 后停止，A3 未启动，没有最终候选产物；固定原始归档中两臂均为 INCONCLUSIVE，有效对照未形成。workflow SUCCESS 仅表示编排与留存完成，HTTP 状态不证明模型处理或计费，也不能追溯首轮缺失根因；不构成效果验收。
- 下一步与成本： 当前 PR275 账本及当前日账本完整保留两笔各 $162.72 claim：PR275/试点合计 $325.44，日合计 $816.79（含此前 $491.35）；实际费用仍未知，这些是保留上界而非实付账单。首轮 PR 账本与首轮日账本、无模型资格原始证据、执行冻结及首次失败完整归档均保留；首轮冻结已于 2026-10-31 03:00 UTC 到期。下一步由既有实施者核对已有服务/账户侧 2026-10-31 10:46–10:48 UTC 的拒绝与计费记录；当前证据不能区分 403 的具体来源或原因，不发测试请求来补证。两轮失败与全部预留保留，不退款、不重置预算、不作第三次尝试，剩余额度不等于执行授权。候选执行沙箱不等于 actor 工具 worker 的 OS 隔离；冻结、资格或 CI 均不代替效果验收。
Goal and acceptance
- Problem / desired outcome: InferRSI has not measured its own value on the work it exists for. Skill Eval (#40, #100) and the Collect comparison (#128) graded one answer document per trial against a frozen verifier, under minute-scale client timeouts (#128 used 360 seconds). None of them measured a code change with the engine runner. The exploratory runs in #220 found that a no-library agent already reached the configuration ceiling of the #75 task, and that outcomes depended on whether the agent benchmarked a candidate at all. This issue proposes a pre-registered, harness-agnostic evaluation for this question: for a fixed coding agent (Claude Code or Codex) and model, does adding InferRSI change three things at equal budget? Those are the independently re-measured result of an inference-optimization task, the reliability of the agent's own claims, and the cost of getting there. "InferRSI" here means its knowledge, its workbench, or both.
- Completion criteria:
- Model-free tooling and fixtures that freeze, run and grade all five arms (four configurations plus an A/A duplicate) with both harness adapters, merged through the normal lanes.
- One authorized pilot. Its frozen plan digest is posted here before any paid call, and every run is retained. The report states per-arm outcomes, claim reliability, cost, A/A variance and limitations. A null or negative efficacy result satisfies this criterion.
Scope and constraints
- In scope:
- Four arm configurations in the tooling: no InferRSI, knowledge only, workbench only, and both, plus an A/A duplicate of the first. The first pilot runs only the two end configurations and the A/A duplicate.
- Claude Code and Codex adapters.
- A CPU task tier and at least one single-accelerator task.
- Independent held-out re-measurement through the existing AB/BA runner.
- Claim-reliability and cost metrics, and a report.
- Out of scope:
- Ranking harnesses or models against each other.
- New CI lanes and knowledge promotion.
- Host plugin packaging (#152).
- Engine performance claims. Candidate patches graded here are evaluation artifacts, not campaign
PASSresults.
- Relevant dependencies, shared interfaces or limits:
- Reuse
run_experiment,evaluate,qualify_devices,repository_identityand the Skill Evalprepare→authorize→executelifecycle with its receipts. Add no second runner. - Thresholds, budgets and task pins live in this issue and the frozen plan, not in code defaults.
- Paid runs need runner-enforced isolation.
- Reuse
Ownership and authorization
- Proposed by / actual requester: @david6666666. Drafted with Claude Code (
claude-opus-5-5) at the requester's direction. - Status: Proposed
- Accountable parent (Issue-level owner, if assigned): TBD
- Slice implementer(s) / acting contributor (if assigned): TBD
- Goal-acceptance owner: TBD
- Accepted decision or scoped delegation: Not yet granted. Paid model calls, accelerator time and evidence publication each need separate authorization.
Inputs and deliverables
- Relevant sources, revisions or linked issues:
- Expected artifacts:
tools/task_efficacy.pywithprepare,authorize,execute,gradeandreport.- A
ClaudeCodeClientand a long-horizon mode forCodexClient. - A Tier-0 CPU task pack with scripted offline clients.
- The report schema.
- The pilot evidence bundle.
The full plan is in the English comment below. A Chinese translation follows it.
Completion and closure
- Who accepts completion and may close: the goal-acceptance owner above.
- Remaining uncertainty / next decision:
- Which harness to run first.
- The Tier-1 task.
- The pilot budget and repetition count.
- Whether to publish raw traces or only redacted ones.
Activity
david6666666 commentedon Oct 29, 2026
Full plan: does InferRSI help Claude Code or Codex optimize an inference task?
Inspected on 2026-10-29 at main
509531c6. Line links point at that revision. This is a proposal: it authorizes no model calls, accelerator time or publication. A Chinese translation follows in the next comment.1. Why the existing evaluations cannot answer this
run_trialinputs_changed/ INCONCLUSIVE (L262-L267). The verifier gets 30 seconds (L241). An optimization deliverable is a code change, so this path rejects it by design.run_experimentandevaluateThe roadmap already names the missing step: "Task-efficacy pilot: compare a bounded task with and without the relevant knowledge; report correctness, completion and cost" (#61).
2. Where a difference can come from
coveredtasksuncoveredtasksThe last row is the expected harm. A gain and a cost can both be real, which is why outcome, claims and cost are separate endpoints.
3. Arms and what stays fixed
InferRSI offers two things that can help independently:
KnowledgeReaderandexport_bundle, then installed the way each harness loads skills.inferrsiCLI and runner onPATH, so the agent can run gated AB/BA comparisons with retained evidence instead of ad-hoc timing.The tooling supports all five arms: four configurations and the A/A duplicate. The first pilot runs A0, A0' and A3, so its primary comparison is A3 against A0. A1 and A2 decompose an effect once one has been shown.
Within one harness, every arm shares the harness version, model ID, effort, task prompt, raw benchmark and validator commands, budget, workspace content, devices and permission rules. Every arm can run the task's own benchmark and validator. W adds a disciplined measurement path; it does not unlock measurement that A0 lacks. The only prompt difference is one frozen paragraph per arm that says what is available, and its hash is part of the plan.
4. Hypotheses and endpoints
All of these are frozen before any paid call.
PASSfromevaluateat the task's frozenmin_improvement_pct.Secondary measures are descriptive only. They are verified gain (an invalid or failed patch counts as zero gain), pass@k and pass^k over repeats, and the process measures in section 8. A single primary endpoint keeps the pilot clear of multiple-comparison fishing.
5. Tasks
#220 gives three design rules. The configuration ceiling must sit below the target, the remaining step must be a code change, and benchmarks must block in the foreground instead of being pushed into the background by a harness time limit.
Each task is a frozen
task.jsonwith:read_metricreads, as insynthetic_benchmark.py), a validator command, metric key, direction andmin_improvement_pct;held_out_splitandheld_out_releasefields (campaign-plan.yaml#L139-L140);coveredmeans the frozen K bundle holds relevant but not answer-bearing knowledge;uncoveredmeans it holds none. K should help only on covered tasks. Uncovered tasks test for harm: in a recent study, repository context files did not generally improve task success and raised inference cost by over 20% on average.Provenance rule: no K record may derive from the task's own campaign, evidence or reference patch. Before freeze, the K bundle is searched for the task's model, engine and reference-patch identifiers, and every hit is reviewed and listed in the plan. For now this rules out #75 as a held-out task, because seven files under
src/inferrsi/dataalready mention Qwen3-TTS.qualify_devices, which already refuses a GPU that has a compute process (L91).6. Harness adapters
Both adapters produce one normalized trial record. It holds the harness version, the model IDs actually used, effort, usage, estimated cost where the harness reports it, the commands the agent ran, the files it changed, errors, exit state, the final patch and the claim document. Both run under the existing
execute_commanddeadline, so no trial outlives its budget. The prompt states the goal (metric, direction, target), the visible inputs, the commands and the budget. It says nothing about the held-out grading inputs.Claude Code (a new
ClaudeCodeClientnext toCodexClient):Every arm also gets this environment:
CLAUDE_CONFIG_DIRper trial, which holds settings, session history and plugins;CLAUDE_CODE_DISABLE_AUTO_MEMORY=1andCLAUDE_CODE_DISABLE_CLAUDE_MDS=1, so no host memory or CLAUDE.md reaches the agent;CLAUDE_CODE_SUBAGENT_MODEL=<same ID>withCLAUDE_CODE_SUBAGENT_MODEL_FORCE=1, so subagents run on the frozen model;BASH_DEFAULT_TIMEOUT_MSandBASH_MAX_TIMEOUT_MSabove the task's benchmark time, plusCLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1, so benchmarks block in the foreground. This removes the task: pre-registered campaign on experience and context compaction for agents on the #75 task (Qwen3-TTS 0.6B, RTX 4090), with exploratory results #220 failure mode where every benchmark became a background job;DISABLE_AUTOUPDATER=1, so the version cannot change during the pilot.--bareis not used. It skips skill and plugin discovery and limits the agent to Bash, read and edit tools, so it would change the harness under test instead of isolating it. Thesystem/initevent must list the K plugin in K arms, no plugin in the other arms, noplugin_errorsand no MCP servers. Any other result is a setup failure: the trial is INCONCLUSIVE and reruns under the frozen retry cap. The agent writes its claim document to a frozen path outside the checkout, and the adapter validates it against the claim schema.For cost,
total_cost_usdandmodelUsageare client-side estimates that include subagents, whileusageleaves subagents out (cost tracking). Accounting therefore usesmodelUsage. A trial whosemodelUsagelists any model other than the frozen one is markedmodel_driftand reported, never silently dropped.--max-budget-usdchecks an estimate that spend can pass, so the plan keeps headroom.Codex (a long-horizon mode of the existing
CodexClient):client_config():--ignore-user-config --ephemeral --json, with apps, plugins, memories, web search and multi-agent off andproject_doc_max_bytes=0.inputs_changedrule does not apply; the grader checks protected paths instead. The timeout is the task's wall-clock budget, and the final message is the claim document through the existing--output-schema.HOMEandCODEX_HOME, with credentials injected at runtime.client_config()already disables global and bundled skills. K arms expose the bundle through repo-scope.agents/skillsdiscovery in the trial checkout, as the Skill Eval discovery split already does (skill_eval_compare.py#L539). That directory is listed in the checkout's.git/info/exclude, so it never enters the graded patch. Other arms have no such directory. Theexec --jsonstream reports thread, turn and item events, not a loaded-skill list, so the load check is file presence plus trace evidence that the agent read the skill files.codex exechas no turn or spend cap, and one run reports usage only when its turn completes. The wall-clock deadline is the hard budget. Tokens are recorded and priced from a frozen price table in the plan, and trials over the frozen token cap are flagged.7. Trial protocol
preparewrites the plan and its digest. The plan holds the task pins, the arm definitions (prompt paragraph hashes, the K bundle digest, the InferRSI SHA for W), the harness version, model ID and effort, the budgets, the repetition count, the schedule seed, the retry cap, the price table and the decision rule.authorizestep records the approver against that digest, following the existing Skill Eval modes (skill_eval_pr.py#L816).executerefuses a plan whose digest has no matching authorization.run_experimentwith the pinned SHA as baseline and the patched checkout as candidate, alternating AB/BA order (runner.py#L163) and snapshotting both sources around every command (L68);evaluatefor aPASS,FAILorINCONCLUSIVEverdict and the measured gain.8. Measures
PASS, but the agent claimed nothing)inferrsicommands ranThe dose check matters for interpretation. If K arms never read the skills, the knowledge effect was not tested; that is a different finding from a null effect.
9. Statistics and decision rule
uncoveredtasks show a lower success rate or a cost increase above a frozen margin with no success gain.10. Milestones
tools/task_efficacy.pywithprepare,authorize,execute,gradeandreport; the plan, claim and report schemas; the grader; two or three Tier-0 tasks, each with a reference solution and a ceiling probe. Scripted offline clients cover four behaviors: one improves the metric and claims it honestly, one claims a gain it did not make, one never measures, and one edits a protected path. Each must receive the expected grade in Ready CI. A test also confirms that the runner commands work from an InferRSI build withoutsrc/inferrsi/data, which the A2 arm uses.ClaudeCodeClientand the long-horizonCodexClientmode, tested offline against stub executables that replay recorded event streams. Tests cover the load checks, model-drift detection, deadline and budget enforcement and claim validation. M1 also adds the isolation check for paid hosts and completes the Tier-0 pack.M0 and M1 meet completion criterion 1. M2 and M3 meet criterion 2.
11. Risks and mitigations
DISABLE_AUTOUPDATER=1, version and model IDs recorded per trial, and themodel_driftflag.--fallback-modelis never passed. Affected blocks are rerun or reported separately.--max-budget-usdand token caps, a frozen pilot total thatexecutechecks before each trial, and the wall-clock deadline for Codex--allowedToolslist for Claude Code.client_config()currently setssandbox_mode = "danger-full-access"for Codex (L24); the long-horizon mode overrides it toworkspace-writewhere the device class allows, and otherwise relies on the container.A result describes one InferRSI revision, one harness and model pair, and these tasks. It does not rank harnesses or models. A Tier-0 gain is not evidence about accelerator engine work unless Tier 1 shows it too.
david6666666 commentedon Oct 29, 2026
完整方案（中文译文）：InferRSI 能否帮助 Claude Code 或 Codex 优化推理任务？
核查时间为 2026-10-29，版本为 main 分支上的
509531c6。行号链接均指向该版本。这是一份提案：它不授权任何模型调用、加速器时间或发布。1. 现有评估为何无法回答这个问题
run_trialinputs_changed/ INCONCLUSIVE（L262-L267）。验证器的时限为 30 秒（L241）。优化任务的交付物是代码改动，所以这条路径在设计上就会拒绝它。run_experiment与evaluate路线图已经点出了缺失的这一步：“任务效能试点：在有、无相关知识两种条件下比较一个有界任务；报告正确性、完成情况和成本”（#61）。
2. 差异可能来自哪里
covered任务上，候选更少，达到首个验证收益所需的时间更短uncovered任务上成功率更低最后一行是预期的危害。收益和代价可能同时真实存在，所以结果、声明和成本分别作为独立的终点指标。
3. 实验组与固定条件
InferRSI 提供两样东西，它们可以各自独立地带来帮助：
KnowledgeReader和export_bundle导出，再按各 harness（智能体运行框架）加载 skill 的方式安装。PATH上的inferrsiCLI 和 runner，让智能体能运行设有门槛、保留证据的 AB/BA 对比，而不是临时计时。这套工具支持全部五个实验组：四种配置加上 A/A 副本。首个试点运行 A0、A0' 和 A3，因此主要比较是 A3 对 A0。A1 和 A2 在某个效应得到证明之后再用于分解该效应。
在同一个 harness 内，所有实验组共享 harness 版本、模型 ID、effort、任务提示词、原始基准测试与校验器命令、预算、工作区内容、设备和权限规则。每个实验组都能运行任务自带的基准测试和校验器。W 增加的是一条有纪律的测量路径，并不会解锁 A0 所没有的测量能力。各实验组提示词的唯一差别是一段冻结的说明，告知该组可以使用哪些东西；这段说明的哈希值是计划的一部分。
4. 假设与终点指标
以下各项都在任何付费调用之前冻结。
min_improvement_pct下由evaluate返回PASS。次要指标仅作描述，包括：验证收益（无效或失败的补丁按零收益计）、基于重复运行的 pass@k 和 pass^k，以及第 8 节中的过程指标。只设一个主要终点指标，可以避免试点在多重比较中挑选有利的结果。
5. 任务
#220 给出了三条设计规则：配置上限必须低于目标；剩下的那一步必须是代码改动；基准测试必须在前台阻塞运行，不能因为 harness 的时间限制被推到后台。
每个任务都是一个冻结的
task.json，包含：read_metric读取的格式，可参考synthetic_benchmark.py），一条校验器命令，以及指标键、方向和min_improvement_pct；held_out_split和held_out_release字段对应（campaign-plan.yaml#L139-L140）；covered表示冻结的 K 知识包中含有相关但不直接给出答案的知识；uncovered表示其中没有相关知识。K 预期只在 covered 任务上有帮助。uncovered 任务用于检测危害：在一项近期研究中，仓库上下文文件总体上没有提高任务成功率，并且使推理成本平均上升了 20% 以上。来源规则：任何 K 记录都不得源自该任务自身的 campaign、证据或参考补丁。冻结之前，要在 K 知识包中搜索该任务的模型、引擎和参考补丁标识符，每一处命中都要经过审查并列入计划。按这条规则，#75 目前不能用作留出任务，因为
src/inferrsi/data下已有七个文件提到 Qwen3-TTS。qualify_devices实现，它已经会拒绝存在计算进程的 GPU（L91）。6. Harness 适配器
两个适配器都会产出一份规范化的试验记录，内容包括：harness 版本、实际使用的模型 ID、effort、用量、估算成本（若 harness 有报告）、智能体运行过的命令、它改动的文件、错误、退出状态、最终补丁和声明文档。两者都在现有
execute_command的截止时间下运行，因此任何试验都不会在预算耗尽后继续运行。提示词说明目标（指标、方向、目标值）、可见输入、命令和预算，对留出评分输入则只字不提。Claude Code（在
CodexClient旁边新增一个ClaudeCodeClient）：每个实验组还会获得以下环境设置：
CLAUDE_CONFIG_DIR，其中存放设置、会话历史和插件；CLAUDE_CODE_DISABLE_AUTO_MEMORY=1和CLAUDE_CODE_DISABLE_CLAUDE_MDS=1，确保宿主机上的记忆和 CLAUDE.md 都不会传到智能体；CLAUDE_CODE_SUBAGENT_MODEL=<same ID>加上CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1，让子智能体也运行在冻结的模型上；BASH_DEFAULT_TIMEOUT_MS和BASH_MAX_TIMEOUT_MS设为高于任务的基准测试耗时，并设置CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1，让基准测试在前台阻塞运行。这消除了 task: pre-registered campaign on experience and context compaction for agents on the #75 task (Qwen3-TTS 0.6B, RTX 4090), with exploratory results #220 中每次基准测试都变成后台作业的失败模式；DISABLE_AUTOUPDATER=1，确保试点期间版本不会变化。不使用
--bare。它会跳过 skill 和插件的发现，并把智能体限制在 Bash、读取和编辑工具上，因此使用它会改变被测的 harness 本身，而不是把它隔离出来。system/init事件必须满足：K 实验组中列出 K 插件，其他实验组中不列出任何插件，没有plugin_errors，也没有 MCP 服务器。出现其他任何结果都算环境准备失败：该试验判为 INCONCLUSIVE，并在冻结的重试上限内重新运行。智能体把声明文档写到检出目录之外的一个冻结路径，适配器依据声明 schema 对其进行校验。成本方面，
total_cost_usd和modelUsage是包含子智能体的客户端估算值，而usage不包含子智能体（成本追踪）。因此成本核算使用modelUsage。如果某次试验的modelUsage中出现冻结模型以外的任何模型，该试验会被标记为model_drift并报告出来，绝不会被悄悄丢弃。--max-budget-usd检查的是估算值，实际花费可能超过它，所以计划要留出余量。Codex（现有
CodexClient的长时程模式）：client_config()：--ignore-user-config --ephemeral --json，关闭应用、插件、记忆、网页搜索和多智能体，并设置project_doc_max_bytes=0。inputs_changed规则在这里不适用，评分器改为检查受保护路径。超时时间就是任务的墙钟预算，最终消息即声明文档，通过现有的--output-schema输出。HOME和CODEX_HOME，凭据在运行时注入。client_config()已经禁用了全局 skill 和内置 skill。K 实验组通过试验检出目录中仓库级的.agents/skills发现机制暴露知识包，Skill Eval 中 discovery 划分的用例已经采用了这种做法（skill_eval_compare.py#L539）。该目录列在检出目录的.git/info/exclude中，因此不会进入被评分的补丁。其他实验组没有这个目录。exec --json流报告的是 thread、turn 和 item 事件，而不是已加载 skill 的列表，所以加载检查依据的是文件存在，再加上轨迹中显示智能体读过 skill 文件的证据。codex exec没有轮次上限或花费上限，而且单次运行只在其轮次结束时才报告用量。墙钟截止时间就是硬性预算。token 会被记录下来，并按计划中冻结的价格表计价，超过冻结 token 上限的试验会被标记。7. 试验流程
prepare写出计划及其摘要。计划包含任务的固定版本信息、实验组定义（提示词段落哈希、K 知识包摘要、W 所用的 InferRSI SHA）、harness 版本、模型 ID 和 effort、各项预算、重复次数、调度种子、重试上限、价格表和决策规则。authorize步骤针对该摘要记录批准人，沿用现有 Skill Eval 的几种模式（skill_eval_pr.py#L816）。如果计划的摘要没有对应的授权，execute会拒绝执行。run_experiment，交替使用 AB/BA 顺序（runner.py#L163），并在每条命令前后对两份源码做快照（L68）；evaluate，得到PASS、FAIL或INCONCLUSIVE判定以及测得的收益。8. 测量指标
PASS，但智能体未声称任何收益）inferrsi命令剂量检查对结果解读很重要。如果 K 实验组从未读过 skill，知识效应就没有得到检验；这与零效应是两种不同的结论。
9. 统计方法与决策规则
uncovered任务的成功率更低，或其成本增幅超过冻结的容差而成功率没有提高。10. 里程碑
tools/task_efficacy.py，包含prepare、authorize、execute、grade和report；计划、声明和报告的 schema；评分器；两到三个 Tier-0 任务，每个都配有参考解和上限探针。离线脚本化客户端覆盖四种行为：一个改进了指标并如实声明，一个声称了并未取得的收益，一个从不测量，一个修改了受保护路径。每个客户端都必须在 Ready CI 中得到预期的评分。另有一项测试确认，runner 命令在不含src/inferrsi/data的 InferRSI 构建上可以正常工作，A2 实验组使用的正是这种构建。ClaudeCodeClient和CodexClient的长时程模式，使用回放已录制事件流的桩可执行文件进行离线测试。测试覆盖加载检查、模型漂移检测、截止时间与预算的强制执行，以及声明校验。M1 还会为运行付费调用的主机加入隔离检查，并补全 Tier-0 任务集。M0 和 M1 满足完成标准 1。M2 和 M3 满足标准 2。
11. 风险与缓解措施
DISABLE_AUTOUPDATER=1、每次试验记录版本和模型 ID，以及model_drift标记。从不传入--fallback-model。受影响的分块重新运行或单独报告。--max-budget-usd和 token 上限，execute在每次试验前检查的冻结试点总额，以及 Codex 的墙钟截止时间--allowedTools列表。client_config()目前为 Codex 设置了sandbox_mode = "danger-full-access"（L24）；长时程模式在设备类别允许时将其覆盖为workspace-write，否则依赖容器。一项结果描述的是一个 InferRSI 版本、一组 harness 与模型的组合，以及这些任务。它不对 harness 或模型进行排名。除非 Tier 1 也显示出这一收益，否则 Tier-0 上的收益不能作为加速器引擎工作方面的证据。
Dong1017 commentedon Oct 29, 2026
Proposal review of the English plan and its Chinese translation, against the Issue body updated 2026-10-29T07:46:06Z. Assessed main:
a8215b3ec314b8bc5cee152cdebbf4bb76ba4188; the plan cites509531c681479fbdd7ddd0e8b0032239e64a44c1. This is a design/source review, with no model calls, device runs or executed efficacy measurements.The separation of K/W arms, equal access to raw measurement commands, held-out grading and model-free milestones is useful. Two experimental contracts need revision before preregistration:
P2 — Separate refuted claims from unverified claims (§4, §8, §9). H2 currently counts both FAIL and INCONCLUSIVE as a false claim. The existing evaluator deliberately returns INCONCLUSIVE for missing metrics, timeouts or unverified source integrity; it does not establish that the claimed gain was false. An identical honest claim would therefore acquire a different “false” label merely because its grading run timed out, and this also changes the “helps” decision's H2 condition. Keep verified/refuted/unverified outcomes separate, state the false-claim denominator and grading coverage, and leave numeric claim error unavailable when no valid measured gain exists. If desired, retain an explicitly named “unverified-or-refuted claim” operational metric. Add an offline fixture where a valid claimed improvement meets a grader infrastructure failure.
P2 — Define the enforced equal-budget comparison (§6, §11). The Codex design explicitly records usage only after completion and merely flags a token-cap overrun. The shared executor enforces elapsed time, not tokens or USD; checking the pilot total before starting a trial does not cap that trial's spend. Freeze the primary comparison as equal enforced wall-clock budgets unless another resource limit can actually be enforced. Label token/USD limits and estimates accordingly, and freeze how overruns and every setup retry enter outcomes and total cost; post-hoc exclusion must not create a favorable subset. Exercise those rules in M1 fixtures.
Recommendation: revise these definitions, then have the authorized owner decide a bounded M0/M1 slice and the shared adapter boundary with #249. This review does not grant implementation or pilot resources. Requester remains @david6666666; accountable implementer and acceptance owner are still TBD. No completion or efficacy acceptance is established.
Dong1017 commentedon Oct 30, 2026
当前受限执行切片：CPU KV-block 回收队列
本轮推进一个独立的、已获 owner 授权的合成 CPU 任务与确定性判据，先完成无模型资格验证，再在隔离与成本约束满足时比较 workbench-only（A2）与 workbench＋固定知识（A3）。实施与证据发布由当前 owner 侧工作承接；本 Issue 原有五臂、双 harness、加速器层级和整体验收仍分别保留为原提案，不据此全部接受或变更其他贡献者分工。
该切片针对现有设计 review提出的结论分类与强制预算问题显式约束；其结果只用于这个有界任务，不能替代广泛的知识效果、Indexer 准确度或 campaign 验收。
Dong1017 commentedon Oct 30, 2026
CPU reclamation: model-free qualification plan freeze
Under the accepted bounded slice, the independent neutral verifier was frozen before the optimized software control: 88 traces / 1,076 events, with 11 development and 77 grader traces. The verifier author did not read the target knowledge or implementation before freeze.
163c21134a8155da2a6469d9dbddb6ea71cb162e14f53e1c4266fe8ce75126f7.60cc479686aa2ba3772f627cc5179a3c0dc6d215c0fdeb57547fd58dc30859ed(UTF-8 JSON below, with one final LF).No model calls, paid execution, actor-isolation acceptance, knowledge efficacy or production/vLLM conformance is established here. The broader Issue scope and original results remain unchanged.
Exact frozen qualification plan
{ "schema": "kv-reclamation-model-free-plan-v1", "scope": "Synthetic CPU mechanism transfer qualification; no model trials or vLLM conformance", "accepted_scope": "https://github.com/InferMatrixLab/InferRSI/issues/248#issuecomment-6526000390", "base_revision": "3902e4abcbe8806a683ab06caf4b9f09743c20cf", "verifier_freeze_sha256": "163c21134a8155da2a6469d9dbddb6ea71cb162e14f53e1c4266fe8ce75126f7", "starter_sha256": "a8cdf534deead1d9ba8351e436f1019a4608e15daae61e54a971aec13047928d", "workload_sha256": "aac18d5edc5c9421ea7d28bec475f7089478baa7ae644ec12ddb368a0efe4bc7", "timing": { "lease_count": 1800, "paired_repetitions": 8, "min_improvement_pct": 10.0, "negative_absolute_gain_limit_pct": 10.0, "comparison_max_seconds": 180, "warmup": "none; one complete trace per fresh sandbox process", "metric": "host elapsed seconds including sandbox and interpreter startup", "order": "existing runner alternating AB/BA" }, "qualification_max_seconds": 600, "deadline_probe_seconds": 0.3, "required_good": [ "starter", "reference" ], "required_bad": [ "ignore_save", "ignore_fence", "ignore_generation", "empty_answers", "skip_events", "missing_output", "spoof_output" ], "required_negative_comparisons": [ "unchanged", "no_op" ], "configuration_surface": "none; no_op is identical starter bytes, separately measured", "retry_policy": "no automatic reruns, retain every attempt; any correction gets a new receipt", "paid_execution": "not implemented or authorized by this plan", "verdict_scope": "model-free exact task/environment qualification only; actor isolation remains unqualified" }feat: add bounded KV editing and scripted actor qualification (#275)
feat: add inactive bounded KV pilot launcher (#276)
Dong1017 commentedon Oct 31, 2026
Pre-admission rates and accounting support a maximum $162.72 for the single bounded A2/A3 pair. No reservation or model call has occurred. The existing implementer must still bind the reviewed freeze and recheck prices/ledgers before execution.
Exact rates, sources and accounting evidence
Bounded CPU KV pilot: public rates and pre-admission accounting
Public rates checked 2026-10-31 01:42 UTC at Yibuapi's official pricing API, version
a42d372ccf0b5dd13ecf71203521f9d2, forclaude-opus-5-5:All three public group ratios remain 2. The token-based public record advertises no separate request fee; this calculation uses 0. Higher-priced groups exist.
Effective-group provenance remains the prior owner-attested existing-credential statement, unchanged since 2026-10-30. This refresh does not independently inspect the account/invoice or renew that attestation. Stop if its applicability or the rates change.
Operational price-evidence expiry: 2026-10-31 03:00 UTC (
expires_at=1793415600). This is an admission cutoff, not a provider price guarantee. Recheck rates and accounting before admission; retain at least 1,320 seconds before both expiry and Shanghai midnight for the whole pair.Prospective maximum: one A2/A3 feasibility pair, at most 8 requests per arm / 16 total, no automatic retries. Each request reserves the full 1,000,000 input tokens at $10/M plus 4,096 output at $40/M: $10.16384, rounded up to $10.17/request; $162.72/pair.
Accounting checked 2026-10-31 01:43 UTC: sole charge identity PR275, whose audited genesis has revision 0, no entries and $0 known prior charge. The unchanged daily ledger retains $491.35 across nine claimed entries with unknown actual cost/closure; midnight does not release them.
After a prospective $162.72 reservation, remaining capacity would be $37.28 under PR275's $200 lifetime cap, $345.93 under the $1,000 Shanghai-day cap, and $837.28 under this pilot's separate $1,000 ceiling. The unrelated historical carryover constrains the day, not pilot-total spending.
This comment records rates/accounting only. No pilot reservation or provider call has occurred. Genesis setup is not a spending grant; the exact reviewed execution freeze, qualification, durable admission and remaining authorization conditions still apply. No efficacy or broader Issue acceptance is claimed.
Dong1017 commentedon Oct 31, 2026
{"plan":{"actors":{"A2":{"config":{"bytes":3527,"path":"evidence/kv-live-pilot-20261031/plan/A2.json","sha256":"e6936bbb5cf3f78177bef2a299b1e50adbab17bf179121e782c3be6befd49770"},"workspace":[{"path":"CONTRACT.md","source":{"bytes":5608,"path":"evidence/kv-reclamation-20261030/validator/independent/CONTRACT.md","sha256":"85714ad11d360ca885c417cb4d56ef4ecf289e3a0c24696e94808e45661d2ccd"}},{"path":"candidate.py","source":{"bytes":3386,"path":"tests/fixtures/kv_reclamation/candidate.py","sha256":"a8cdf534deead1d9ba8351e436f1019a4608e15daae61e54a971aec13047928d"}},{"path":"development.json","source":{"bytes":11315,"path":"evidence/kv-live-pilot-20261031/plan/public/development.json","sha256":"e6da6e89af6a2956cc8f7dfe494b61fe9919900eefac5cb724bb508947dcb56c"}},{"path":"task.json","source":{"bytes":214,"path":"evidence/kv-live-pilot-20261031/plan/public/task.json","sha256":"8ee9cc95f07c8c36550f25008862f381a0a0a457bf2f62d4e644c6e38cfad44b"}},{"path":"workbench.md","source":{"bytes":950,"path":"evidence/kv-live-pilot-20261031/plan/public/workbench.md","sha256":"97596737c04f1904941e6eddd0090107af762d2743625260bf2185a35d85df5f"}}]},"A3":{"config":{"bytes":3907,"path":"evidence/kv-live-pilot-20261031/plan/A3.json","sha256":"4e895045d7a556165fafeea48a8ce92bb50da48090a0d2a1c92f27aa98e5eb57"},"workspace":[{"path":"CONTRACT.md","source":{"bytes":5608,"path":"evidence/kv-reclamation-20261030/validator/independent/CONTRACT.md","sha256":"85714ad11d360ca885c417cb4d56ef4ecf289e3a0c24696e94808e45661d2ccd"}},{"path":"candidate.py","source":{"bytes":3386,"path":"tests/fixtures/kv_reclamation/candidate.py","sha256":"a8cdf534deead1d9ba8351e436f1019a4608e15daae61e54a971aec13047928d"}},{"path":"development.json","source":{"bytes":11315,"path":"evidence/kv-live-pilot-20261031/plan/public/development.json","sha256":"e6da6e89af6a2956cc8f7dfe494b61fe9919900eefac5cb724bb508947dcb56c"}},{"path":"knowledge/SKILL.md","source":{"bytes":7139,"path":"src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/SKILL.md","sha256":"4c3f27459df239f9f4fc971c559d4c9ddfbfebd70507a8a93be8eae59a30b69d"}},{"path":"knowledge/metadata.json","source":{"bytes":781,"path":"src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/metadata.json","sha256":"b43fc3a7ccae59c664fcdf9ff77962edcc07c2a3df8b59324db551642bd23e23"}},{"path":"knowledge/references/source-notes.md","source":{"bytes":13853,"path":"src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/references/source-notes.md","sha256":"3665b034594fc8cc767c408002f4e16f8803369242405027c7f37ce6c9ecebe9"}},{"path":"task.json","source":{"bytes":214,"path":"evidence/kv-live-pilot-20261031/plan/public/task.json","sha256":"8ee9cc95f07c8c36550f25008862f381a0a0a457bf2f62d4e644c6e38cfad44b"}},{"path":"workbench.md","source":{"bytes":950,"path":"evidence/kv-live-pilot-20261031/plan/public/workbench.md","sha256":"97596737c04f1904941e6eddd0090107af762d2743625260bf2185a35d85df5f"}}]}},"arm_order":["A2","A3"],"campaign_id":"kv-reclamation-feasibility-20261031","charge_head":"5af3553683137582d624379d26c2c5f7cc9e2785","charge_pr":275,"daily_ledger":{"checkpoint_commit":"bde0bd429b8bf2f4f4fe785bc51cfaed1818acb6","genesis_commit":"1d3df4ea3924c2b4ce92f051becf566839d66c9c","ref":"refs/heads/skill-eval-budget/1409534281/daily"},"dependencies":{".github/workflows/kv-edit-pilot.yml":"421827ae94e0b5308584c581f4be5ab5f5b155b622e8bae8163009cf73a203a8","evidence/kv-actor-qualification-20261030/plan/qualification-plan.json":"b2f2d4fa301573b1bfb451fb0ce7f1777bfcd8c8088a61576af8703c45cd7949","evidence/kv-reclamation-20261030/plan/qualification-plan.json":"60cc479686aa2ba3772f627cc5179a3c0dc6d215c0fdeb57547fd58dc30859ed","evidence/kv-reclamation-20261030/validator/BENCHMARK_PROVENANCE.md":"54ba35ab354e0305f101bfb78d96eee016d1c3bad50b7a97d9452b8f254f2111","evidence/kv-reclamation-20261030/validator/benchmark.py":"c9fd5bdb12ad82ea4bdf9b1cace13e56053d7294eceb9d9357ffdae304984c69","evidence/kv-reclamation-20261030/validator/independent/CONTRACT.md":"85714ad11d360ca885c417cb4d56ef4ecf289e3a0c24696e94808e45661d2ccd","evidence/kv-reclamation-20261030/validator/independent/FREEZE.json":"163c21134a8155da2a6469d9dbddb6ea71cb162e14f53e1c4266fe8ce75126f7","evidence/kv-reclamation-20261030/validator/independent/INDEPENDENCE.md":"94acf341905fb171ef34aedbefb6f4d13fd485d0fc8013d9dc3027b64d97e6e6","evidence/kv-reclamation-20261030/validator/independent/build_fixtures.py":"d1d3f77b96d4187a4e12f5943703daa2de7507175c37faa34e96c57451c7bb46","evidence/kv-reclamation-20261030/validator/independent/control_spec.json":"8db4fece4e266bd1c41d3ae56ecaad3be3fcdcc6ef7a970bb574749cf20d7de5","evidence/kv-reclamation-20261030/validator/independent/fixtures.json":"5f00b3782b85d0f0c5294ab7f191f2b4e58996a14ab609d3de22c9a291c0e34c","evidence/kv-reclamation-20261030/validator/independent/oracle.py":"1521b278a258d6579d000e640068a981b697ab679b3d967593ed62dcf6af2712","evidence/kv-reclamation-20261030/validator/independent/test_verifier.py":"88ef532797ebb502d5a9e0f0a5aea22080a1aec0f3a429bf23f260d0cd64e30f","evidence/kv-reclamation-20261030/validator/independent/verifier.py":"e641fa6088032e15568fe17e87385a9f97e7e1d8d702a65fb62084f99741f862","evidence/kv-reclamation-20261030/validator/source_workload.py":"aac18d5edc5c9421ea7d28bec475f7089478baa7ae644ec12ddb368a0efe4bc7","evidence/kv-reclamation-20261030/validator/supplemental-freeze.json":"717d7c1a4327f3bcaa9755c9b343a142e0223a651996893d7de3e4065cf6b6f0","evidence/kv-reclamation-20261030/validator/test_benchmark.py":"b3f06f781b5839b666e5b1e51e691bc88351ef9640b2c1e62b93af3364fd45da","pyproject.toml":"1c1c4e482c1577883fb6c4678c96571a86eb819b6fcc834937fe344fb4a94b1f","src/inferrsi/__init__.py":"606691627d4b346e60072208ada8d6b4df2409de190daffb7795c2cc24517e47","src/inferrsi/claim.py":"5f24c4fa91b48f0bb4d4554531607763fc8fc108e90f332816f06cfb520b5cc1","src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/SKILL.md":"4c3f27459df239f9f4fc971c559d4c9ddfbfebd70507a8a93be8eae59a30b69d","src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/metadata.json":"b43fc3a7ccae59c664fcdf9ff77962edcc07c2a3df8b59324db551642bd23e23","src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/references/source-notes.md":"3665b034594fc8cc767c408002f4e16f8803369242405027c7f37ce6c9ecebe9","src/inferrsi/harness/__init__.py":"1006e4713003f44bfb32b2e385a918e1652b9e9833801e4555674a71e57fb0d5","src/inferrsi/harness/evaluation.py":"20d0c52145600a5d6b1e513b61f2fffaa459b83db194876be40726bd4f343a59","src/inferrsi/harness/kv_edit_workbench.py":"bc1f64c8aae39534b8052fdd9a6f3a6b8e63f7a9e9c93cd1077c37ef99cc6f69","src/inferrsi/harness/kv_reclamation.py":"56878c4d1bdda096ac681cadf2d99377bec59d47839802fdf94cfbd9066dc8b9","src/inferrsi/harness/reporting.py":"38c58f02c44e8c8aeacf46527e142be0e3142a83e4b2693246d0ca76c0ddd7c3","src/inferrsi/harness/runner.py":"d290fc87710e77c8e6069e95bb0d347016a393707592d2eed98f07951f2d0865","src/inferrsi/harness/scenarios/__init__.py":"4e97913c806077f0cadfeb29d986d2fa75b55ab8766cb86e6e3e9a195a9a1f5a","src/inferrsi/harness/scenarios/external.py":"f3f8d4fc516ebce81c26d82248e83f79b299eed0fc17ce73e2844995aea9c3f9","src/inferrsi/harness/skill_eval_engineering.py":"c83097f6069dd3e09046080b61c88aebc0922d09aab3fd5f1a03d8f04ec976f2","src/inferrsi/infra/__init__.py":"8779897733779266841b0f79b84b1dc43f0fa57c013855c906c809829ecfc77c","src/inferrsi/infra/artifacts.py":"bcf3adf325e34e10a4ecb4baf0af855ddca3d77eff76de5b66e493baaa3d7eaf","src/inferrsi/infra/executors/__init__.py":"327c3cd5c821c832ff260a41dba0ee1e439cdabefc9edfb9cad1cd38d14bc87a","src/inferrsi/infra/executors/local.py":"37bcbca9f3261c8c58a8bf806c9da40720b18d0676c82e8ed24c3901ac9bbe77","src/inferrsi/infra/processes.py":"3b96ebc7a50913e78e86ff3dcba74c8ee8d03d547d9591ca6488f390ac1e6189","src/inferrsi/infra/targets/__init__.py":"1f3ea75dded7c6cc967754c7b6d4f30ef00da8191a749938aabc3c6bb8f86eb3","src/inferrsi/infra/targets/git.py":"abb93f9332db336298102e28c06b9a7553597d4d03b957358602f60d61b1aa65","src/inferrsi/model/__init__.py":"3c1dba4f5159fd9fbc8d4d96e2a83c8441491c1b1dc7b42031a3fc1782ed9d9f","src/inferrsi/model/anthropic_agent.py":"284d52dd379b2948dd45ede167175010fe26615e5a307f848efcbe90c0ca5646","src/inferrsi/model/anthropic_edit_agent.py":"3db20467b07657c1a083f7968cfd996bbbffbe04f987bf4283a8d1456f679cf2","src/inferrsi/model/protocols/__init__.py":"5a2f54483b164de04e309dbebb87a700b213682b0f44a50d89a752dd71a4ce47","src/inferrsi/model/protocols/metrics.py":"5b32e296a746b6de30cb0c89f23f730296b0087884c1acc2b6127c1295d4db7a","src/inferrsi/model/request_guard.py":"c2f8a2f822f32785327a2138ee7b0609896b180819d9d1cbe851534aa6198628","src/inferrsi/runspec.py":"dca0952425caf42095ef0c7e2feb8ef4740cc422ecb61e777f203962d9b87709","src/inferrsi/summary_chart.py":"d29023d60f6be804c1b01c9eb73765d4824d6c3bb870fe22c99a24544bc1d8bc","tests/fixtures/kv_reclamation/candidate.py":"a8cdf534deead1d9ba8351e436f1019a4608e15daae61e54a971aec13047928d","tests/fixtures/kv_reclamation/workload.py":"aac18d5edc5c9421ea7d28bec475f7089478baa7ae644ec12ddb368a0efe4bc7","tools/kv_edit_inputs.py":"485a12e198054b683cd2c6dd6abdbce47a76aec094506f930cb27ebed213eeab","tools/kv_edit_pilot.py":"2e1a2baa9586436d27b3004cb68addc16fa0685bc90961406e31d03f69c3fee3","tools/kv_live_job.py":"040b37052aa93737fc8d2818173955ee1125b81024f6690bf6b9ce5f59f0dc36","tools/kv_live_policy.py":"851577df34f48bc36de52d64d95567514d6878a050aaac809f121c7f596ea83b","tools/kv_live_worker.py":"5231dd2fdaa442c20a7e764a4482a5f6b5054776503b18494aa088ec73c3a9f3","tools/kv_pilot_admission.py":"61afc0c901eb0206432e1289b9935d8a000b4dd90f7e5b3aa33d896ecd29eb1e","tools/kv_runtime_qualification.py":"4dc0dfb45f5ed1eb5f31e3843df81d6b5c22c373d69f2d6c47629f202ac03a0c","tools/kv_worker_supervisor.py":"7b2f25148bc0b14536321aa45b434c23424c53858760a708689a28774d585168","tools/qualify_kv_actor_cleanup.py":"45030c6054056c91706701c1c24c5028acac53e8bd2692ae45ab3163941945b0","tools/qualify_kv_edit_actor.py":"f5b5e884f3f4313d39ba8392ba58f1b29c9af0d1ceed9195347092fc7279cd5f","tools/qualify_kv_reclamation.py":"b53efb547ff6dec04492d08f41b9479afee5d07bbee964911a264f453e620815","tools/skill_eval_budget.py":"4c8dd81d585764eb316e0a7e97d32f70c570f0a5e87b638a6f69ca0a089d33a8","tools/skill_eval_budget_git.py":"4d7f24bc62c5884282e9a2db7b34a75626a05c4146bada3a989246533673f48e","tools/skill_eval_pr.py":"156a2b44f98fb30e97bad875efc889855a6ca543000d95a50e7577a487c56a1d"},"execution_id":"kv-reclamation-one-pair-1","expires_at":1793415600,"final_repetitions":8,"final_seconds":180,"final_suite":{"bytes":538674,"path":"evidence/kv-live-pilot-20261031/plan/final-suite.json","sha256":"ff3de0b55e0eb40120b7b3b13559b070f0ea02cd02b08606bfd5d51014352654"},"hidden_traces":{"bytes":109803,"path":"evidence/kv-live-pilot-20261031/plan/hidden-traces.json","sha256":"935058055afeb86a8eab4a4f185ef3c11bccf2f6f58bcb32fb23e6c3624d9284"},"issue":248,"ledger":{"checkpoint_commit":"6ab781845dbbb9c5f566a83134b201d75075821f","genesis_commit":"6ab781845dbbb9c5f566a83134b201d75075821f","ref":"refs/heads/skill-eval-budget/1409534281/275"},"price_evidence":"https://github.com/InferMatrixLab/InferRSI/issues/248#issuecomment-6537375115","prior_owner_group_statement":"https://github.com/InferMatrixLab/InferRSI/issues/253#issuecomment-6518412099; unchanged historical owner statement dated 2026-10-30, not a new account/invoice check or extension of the expired parser plan. Stop if applicability or rates become uncertain.","provider":{"client":{"api_version":"2023-06-01","endpoint":"https://yibuapi.com/v1/messages","kind":"anthropic-messages-edit-check-v1","max_output_tokens":4096,"max_read_bytes":65536,"max_request_bytes":262144,"max_requests":8,"max_tool_calls":32,"provider":"yibuapi"},"model":{"effort":"not_applicable","name":"claude-opus-5-5"},"pricing":{"currency":"USD","expires_at":1793415600,"group":"owner-attested default/claude-code/claude-code-pro equal-price class","max_input_tokens":1000000,"model":"claude-opus-5-5","rate_evidence":"https://github.com/InferMatrixLab/InferRSI/issues/248#issuecomment-6537375115","rates":{"cache_read":"0.4","cache_write":"10","input":"8","output":"40"},"request_fee":"0","unit":"per_million_tokens"},"spend_limit":{"amount":"162.72","currency":"USD"}},"qualification":{"archive":{"bytes":188377,"path":"evidence/kv-actor-qualification-20261030/results/hosted-41154288295/qualification-artifact.zip","sha256":"6a8f975a4873ac50117db693bfddd8d91cd95f32a9794205a439433c92d418aa"},"artifact_id":13024202043,"dependencies_sha256":"0231e0d789b774a5952d1a458d3d8bedbb60b84c75a86ef6a4c0f18e24ca5666","head":"68fa4f405c6b010c0626a4e4c9d000906673e739","receipt":{"bytes":82259,"path":"evidence/kv-actor-qualification-20261030/results/hosted-41154288295/qualification.json","sha256":"007065b8e0a4526d4df3818dbcf99f7e712eebda49512fef8a2340dfe51a0f6e"},"revision":"01e2d6f35b67cb32c2cf3bcaec1664984f1ef037","run_attempt":1,"run_id":41154288295},"repository_id":1409534281,"schema":"kv-live-plan-v1","scope":"One exploratory synthetic CPU editing feasibility pair; A2 common workbench, A3 identical common bytes plus complete pinned K. Knowledge-aware near-transfer, not a historical holdout, vLLM conformance, independent K efficacy or service claim. Final timing is nonheldout. No retries, replan, refund or relaunch; failed and incomplete first attempts are retained.","suite":{"bytes":538671,"path":"evidence/kv-live-pilot-20261031/plan/visible-suite.json","sha256":"812ac3e5dd686271d40598e8b8d35faa6f039888f1d36852e571b11b36c600bc"}},"plan_sha256":"b66130352cce6fb6201561d3a026a7a78d800ac0f84ecb899f57a0906ffa4767","schema":"kv-live-authorization-v1"}Dong1017 commentedon Oct 31, 2026
Second-attempt price and accounting evidence
Read-only snapshot: 2026-10-31 08:17–08:19 UTC. This supports the already scoped one additional A2/A3 pair (8+8 requests), not a third attempt, refund, or automatic retry. PR275 remains the sole charge identity.
a42d372ccf0b5dd13ecf71203521f9d2, is unchanged forclaude-opus-5-5: USD per million tokens: input 8; output 40; cache read 0.4; cache write 10. Thedefault,claude-code, andclaude-code-progroup ratios are all 2. The public token record advertises no separate request fee; the bound uses 0. Higher-priced groups are not covered.23bae898…retains the first pair's $162.72 unknown-cost claim. Daily checkpoint936569a5…retains that plus $491.35 historical carryover: $654.07, all claimed with unknown actual cost and no closure. No day/PR reset or refund; these are conservative commitments, not actual billed totals.Worst case: per request, 1,000,000 input tokens at the highest input-class rate ($10/M) plus 4,096 output tokens at $40/M = $10.16384, rounded up to $10.17; 16 requests = $162.72.
The reviewed prospective exception limits the lifetime-cap change to PR275: $350, with the Shanghai-day and pilot-total ceilings unchanged at $1,000 each. Retaining every old claim and adding the one new pair would total PR275/pilot $325.44; day $816.79, leaving $24.56 / $183.21 / $674.56 respectively. Other PRs retain the default $200 cap. No second reservation has been made by this publication.
New prospective evidence window ends 2026-10-31 10:00 UTC (
expires_at=1793440800). This is not a provider price guarantee or an extension of the expired first-run freeze. The required 1,320-second runway means actual admission no later than 09:38 UTC, allowing for queue/preflight time. Fresh prices/accounting, exact new v2 freeze/locator, reviewed source-bound qualification and durable admission remain required before any call. No new credential/account access, broader acceptance, or efficacy claim is implied.Dong1017 commentedon Oct 31, 2026
Refreshed rates/accounting — same unused additional pair
Fresh public read: 2026-10-31 10:22–10:23 UTC. New explicit evidence expiry: 12:00 UTC (
1793448000); latest actual admission: 11:38 UTC, retaining the required 1,320-second runway. The previous 10:00 window expired without a second admission and remains unchanged. This is the same one additional 8+8 pair, not another attempt or budget reset.a42d372ccf0b5dd13ecf71203521f9d2. Forclaude-opus-5-5, USD/M rates remain input 8, output 40, cache read 0.4, cache write 10.default,claude-codeandclaude-code-proeach retain group ratio 2. No separate request fee is advertised; the bound uses 0. Other higher-priced groups are not covered.23bae898…retains the first unknown-cost $162.72 claim; daily936569a5…retains it plus $491.35 older carryover: $654.07. No second reservation, settlement or refund is observed. These are retained bounds, not actual billed totals; midnight does not release them.Bound:
(1,000,000 × $10/M) + (4,096 × $40/M) = $10.16384, rounded up to $10.17/request × 16 = $162.72. The exact prospective exception remains PR275 lifetime $350 only; day/pilot ceilings $1,000 each; exactly one additional pair, with all old claims retained and other PRs' default $200 unchanged. Prospective totals remain PR275/pilot $325.44; day $816.79, leaving $24.56 / $183.21 / $674.56 respectively. No third attempt or automatic retry.The new expiry is an operational evidence window, not a provider price guarantee. Exact refreshed v2 freeze/locator, reviewed source-bound qualification, current-host preflight, fresh action-time price/accounting checks and durable admission must still pass; queue/preflight time counts toward the 11:38 admission cutoff. This comment makes no reservation or provider call and grants no new credential/account access or broader efficacy claim.
Dong1017 commentedon Oct 31, 2026
{ "plan": { "actors": { "A2": { "config": { "bytes": 3527, "path": "evidence/kv-live-pilot-20261031/plan/second-attempt/A2.json", "sha256": "4abdd77aecd9c654000226bf80cdb5e77c4310d2d4aab0a41e97b6c3bebb1bdb" }, "workspace": [ { "path": "CONTRACT.md", "source": { "bytes": 5608, "path": "evidence/kv-reclamation-20261030/validator/independent/CONTRACT.md", "sha256": "85714ad11d360ca885c417cb4d56ef4ecf289e3a0c24696e94808e45661d2ccd" } }, { "path": "candidate.py", "source": { "bytes": 3386, "path": "tests/fixtures/kv_reclamation/candidate.py", "sha256": "a8cdf534deead1d9ba8351e436f1019a4608e15daae61e54a971aec13047928d" } }, { "path": "development.json", "source": { "bytes": 11315, "path": "evidence/kv-live-pilot-20261031/plan/public/development.json", "sha256": "e6da6e89af6a2956cc8f7dfe494b61fe9919900eefac5cb724bb508947dcb56c" } }, { "path": "task.json", "source": { "bytes": 214, "path": "evidence/kv-live-pilot-20261031/plan/public/task.json", "sha256": "8ee9cc95f07c8c36550f25008862f381a0a0a457bf2f62d4e644c6e38cfad44b" } }, { "path": "workbench.md", "source": { "bytes": 950, "path": "evidence/kv-live-pilot-20261031/plan/public/workbench.md", "sha256": "97596737c04f1904941e6eddd0090107af762d2743625260bf2185a35d85df5f" } } ] }, "A3": { "config": { "bytes": 3907, "path": "evidence/kv-live-pilot-20261031/plan/second-attempt/A3.json", "sha256": "ab4b0839c5ca6b093b6ea7dd34de2ea4021baadc704dbe993d5d6eb28c57bf95" }, "workspace": [ { "path": "CONTRACT.md", "source": { "bytes": 5608, "path": "evidence/kv-reclamation-20261030/validator/independent/CONTRACT.md", "sha256": "85714ad11d360ca885c417cb4d56ef4ecf289e3a0c24696e94808e45661d2ccd" } }, { "path": "candidate.py", "source": { "bytes": 3386, "path": "tests/fixtures/kv_reclamation/candidate.py", "sha256": "a8cdf534deead1d9ba8351e436f1019a4608e15daae61e54a971aec13047928d" } }, { "path": "development.json", "source": { "bytes": 11315, "path": "evidence/kv-live-pilot-20261031/plan/public/development.json", "sha256": "e6da6e89af6a2956cc8f7dfe494b61fe9919900eefac5cb724bb508947dcb56c" } }, { "path": "knowledge/SKILL.md", "source": { "bytes": 7139, "path": "src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/SKILL.md", "sha256": "4c3f27459df239f9f4fc971c559d4c9ddfbfebd70507a8a93be8eae59a30b69d" } }, { "path": "knowledge/metadata.json", "source": { "bytes": 781, "path": "src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/metadata.json", "sha256": "b43fc3a7ccae59c664fcdf9ff77962edcc07c2a3df8b59324db551642bd23e23" } }, { "path": "knowledge/references/source-notes.md", "source": { "bytes": 13853, "path": "src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/references/source-notes.md", "sha256": "3665b034594fc8cc767c408002f4e16f8803369242405027c7f37ce6c9ecebe9" } }, { "path": "task.json", "source": { "bytes": 214, "path": "evidence/kv-live-pilot-20261031/plan/public/task.json", "sha256": "8ee9cc95f07c8c36550f25008862f381a0a0a457bf2f62d4e644c6e38cfad44b" } }, { "path": "workbench.md", "source": { "bytes": 950, "path": "evidence/kv-live-pilot-20261031/plan/public/workbench.md", "sha256": "97596737c04f1904941e6eddd0090107af762d2743625260bf2185a35d85df5f" } } ] } }, "arm_order": [ "A2", "A3" ], "budget_exception": { "additional_attempts": 1, "campaign_id": "kv-reclamation-feasibility-20261031-second", "daily_checkpoint": "936569a520083cb359a375d3d474a2391a15edae", "execution_id": "kv-reclamation-one-pair-2", "key": { "pr_number": 275, "repository_id": 1409534281 }, "pilot_lifetime_cap_microusd": 1000000000, "pr_checkpoint": "23bae89818a0eb4f1ed9ccda2361d10cbf3908ed", "pr_lifetime_cap_microusd": 350000000, "prior_binding": { "plan_digest": "1f7051e0872e8b790c8d1a7c0aa1e03a6bcd08385af510c3d0b1621f54adac6c", "run_id": "kv-reclamation-one-pair-v1:kv-reclamation-feasibility-20261031", "worst_case": 162720000 }, "schema": "kv-one-additional-attempt-v1" }, "campaign_id": "kv-reclamation-feasibility-20261031-second", "charge_head": "5af3553683137582d624379d26c2c5f7cc9e2785", "charge_pr": 275, "daily_ledger": { "checkpoint_commit": "936569a520083cb359a375d3d474a2391a15edae", "genesis_commit": "1d3df4ea3924c2b4ce92f051becf566839d66c9c", "ref": "refs/heads/skill-eval-budget/1409534281/daily" }, "dependencies": { ".github/workflows/kv-edit-pilot.yml": "421827ae94e0b5308584c581f4be5ab5f5b155b622e8bae8163009cf73a203a8", "evidence/kv-actor-qualification-20261030/plan/qualification-plan.json": "b2f2d4fa301573b1bfb451fb0ce7f1777bfcd8c8088a61576af8703c45cd7949", "evidence/kv-live-pilot-20261031/plan/second-attempt-budget.json": "3146b1daf7dc59710da6346844f7768cc74333de51c6d4292d2731ea7c6458a3", "evidence/kv-reclamation-20261030/plan/qualification-plan.json": "60cc479686aa2ba3772f627cc5179a3c0dc6d215c0fdeb57547fd58dc30859ed", "evidence/kv-reclamation-20261030/validator/BENCHMARK_PROVENANCE.md": "54ba35ab354e0305f101bfb78d96eee016d1c3bad50b7a97d9452b8f254f2111", "evidence/kv-reclamation-20261030/validator/benchmark.py": "c9fd5bdb12ad82ea4bdf9b1cace13e56053d7294eceb9d9357ffdae304984c69", "evidence/kv-reclamation-20261030/validator/independent/CONTRACT.md": "85714ad11d360ca885c417cb4d56ef4ecf289e3a0c24696e94808e45661d2ccd", "evidence/kv-reclamation-20261030/validator/independent/FREEZE.json": "163c21134a8155da2a6469d9dbddb6ea71cb162e14f53e1c4266fe8ce75126f7", "evidence/kv-reclamation-20261030/validator/independent/INDEPENDENCE.md": "94acf341905fb171ef34aedbefb6f4d13fd485d0fc8013d9dc3027b64d97e6e6", "evidence/kv-reclamation-20261030/validator/independent/build_fixtures.py": "d1d3f77b96d4187a4e12f5943703daa2de7507175c37faa34e96c57451c7bb46", "evidence/kv-reclamation-20261030/validator/independent/control_spec.json": "8db4fece4e266bd1c41d3ae56ecaad3be3fcdcc6ef7a970bb574749cf20d7de5", "evidence/kv-reclamation-20261030/validator/independent/fixtures.json": "5f00b3782b85d0f0c5294ab7f191f2b4e58996a14ab609d3de22c9a291c0e34c", "evidence/kv-reclamation-20261030/validator/independent/oracle.py": "1521b278a258d6579d000e640068a981b697ab679b3d967593ed62dcf6af2712", "evidence/kv-reclamation-20261030/validator/independent/test_verifier.py": "88ef532797ebb502d5a9e0f0a5aea22080a1aec0f3a429bf23f260d0cd64e30f", "evidence/kv-reclamation-20261030/validator/independent/verifier.py": "e641fa6088032e15568fe17e87385a9f97e7e1d8d702a65fb62084f99741f862", "evidence/kv-reclamation-20261030/validator/source_workload.py": "aac18d5edc5c9421ea7d28bec475f7089478baa7ae644ec12ddb368a0efe4bc7", "evidence/kv-reclamation-20261030/validator/supplemental-freeze.json": "717d7c1a4327f3bcaa9755c9b343a142e0223a651996893d7de3e4065cf6b6f0", "evidence/kv-reclamation-20261030/validator/test_benchmark.py": "b3f06f781b5839b666e5b1e51e691bc88351ef9640b2c1e62b93af3364fd45da", "pyproject.toml": "1c1c4e482c1577883fb6c4678c96571a86eb819b6fcc834937fe344fb4a94b1f", "src/inferrsi/__init__.py": "606691627d4b346e60072208ada8d6b4df2409de190daffb7795c2cc24517e47", "src/inferrsi/claim.py": "5f24c4fa91b48f0bb4d4554531607763fc8fc108e90f332816f06cfb520b5cc1", "src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/SKILL.md": "4c3f27459df239f9f4fc971c559d4c9ddfbfebd70507a8a93be8eae59a30b69d", "src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/metadata.json": "b43fc3a7ccae59c664fcdf9ff77962edcc07c2a3df8b59324db551642bd23e23", "src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/references/source-notes.md": "3665b034594fc8cc767c408002f4e16f8803369242405027c7f37ce6c9ecebe9", "src/inferrsi/harness/__init__.py": "1006e4713003f44bfb32b2e385a918e1652b9e9833801e4555674a71e57fb0d5", "src/inferrsi/harness/evaluation.py": "20d0c52145600a5d6b1e513b61f2fffaa459b83db194876be40726bd4f343a59", "src/inferrsi/harness/kv_edit_workbench.py": "bc1f64c8aae39534b8052fdd9a6f3a6b8e63f7a9e9c93cd1077c37ef99cc6f69", "src/inferrsi/harness/kv_reclamation.py": "56878c4d1bdda096ac681cadf2d99377bec59d47839802fdf94cfbd9066dc8b9", "src/inferrsi/harness/reporting.py": "38c58f02c44e8c8aeacf46527e142be0e3142a83e4b2693246d0ca76c0ddd7c3", "src/inferrsi/harness/runner.py": "d290fc87710e77c8e6069e95bb0d347016a393707592d2eed98f07951f2d0865", "src/inferrsi/harness/scenarios/__init__.py": "4e97913c806077f0cadfeb29d986d2fa75b55ab8766cb86e6e3e9a195a9a1f5a", "src/inferrsi/harness/scenarios/external.py": "f3f8d4fc516ebce81c26d82248e83f79b299eed0fc17ce73e2844995aea9c3f9", "src/inferrsi/harness/skill_eval_engineering.py": "c83097f6069dd3e09046080b61c88aebc0922d09aab3fd5f1a03d8f04ec976f2", "src/inferrsi/infra/__init__.py": "8779897733779266841b0f79b84b1dc43f0fa57c013855c906c809829ecfc77c", "src/inferrsi/infra/artifacts.py": "bcf3adf325e34e10a4ecb4baf0af855ddca3d77eff76de5b66e493baaa3d7eaf", "src/inferrsi/infra/executors/__init__.py": "327c3cd5c821c832ff260a41dba0ee1e439cdabefc9edfb9cad1cd38d14bc87a", "src/inferrsi/infra/executors/local.py": "37bcbca9f3261c8c58a8bf806c9da40720b18d0676c82e8ed24c3901ac9bbe77", "src/inferrsi/infra/processes.py": "3b96ebc7a50913e78e86ff3dcba74c8ee8d03d547d9591ca6488f390ac1e6189", "src/inferrsi/infra/targets/__init__.py": "1f3ea75dded7c6cc967754c7b6d4f30ef00da8191a749938aabc3c6bb8f86eb3", "src/inferrsi/infra/targets/git.py": "abb93f9332db336298102e28c06b9a7553597d4d03b957358602f60d61b1aa65", "src/inferrsi/model/__init__.py": "3c1dba4f5159fd9fbc8d4d96e2a83c8441491c1b1dc7b42031a3fc1782ed9d9f", "src/inferrsi/model/anthropic_agent.py": "284d52dd379b2948dd45ede167175010fe26615e5a307f848efcbe90c0ca5646", "src/inferrsi/model/anthropic_edit_agent.py": "3db20467b07657c1a083f7968cfd996bbbffbe04f987bf4283a8d1456f679cf2", "src/inferrsi/model/protocols/__init__.py": "5a2f54483b164de04e309dbebb87a700b213682b0f44a50d89a752dd71a4ce47", "src/inferrsi/model/protocols/metrics.py": "5b32e296a746b6de30cb0c89f23f730296b0087884c1acc2b6127c1295d4db7a", "src/inferrsi/model/request_guard.py": "c2f8a2f822f32785327a2138ee7b0609896b180819d9d1cbe851534aa6198628", "src/inferrsi/runspec.py": "dca0952425caf42095ef0c7e2feb8ef4740cc422ecb61e777f203962d9b87709", "src/inferrsi/summary_chart.py": "d29023d60f6be804c1b01c9eb73765d4824d6c3bb870fe22c99a24544bc1d8bc", "tests/fixtures/kv_reclamation/candidate.py": "a8cdf534deead1d9ba8351e436f1019a4608e15daae61e54a971aec13047928d", "tests/fixtures/kv_reclamation/workload.py": "aac18d5edc5c9421ea7d28bec475f7089478baa7ae644ec12ddb368a0efe4bc7", "tools/kv_edit_inputs.py": "485a12e198054b683cd2c6dd6abdbce47a76aec094506f930cb27ebed213eeab", "tools/kv_edit_pilot.py": "2e1a2baa9586436d27b3004cb68addc16fa0685bc90961406e31d03f69c3fee3", "tools/kv_live_job.py": "f58e9f2d8a714a7d010aa50fa134ce9295fa88d1c1f8a4dc093895085f5863fc", "tools/kv_live_policy.py": "a2537e0dc5ba111781fdd9adecbb6b1f6766d573ac3910d13c0b0a9187ff005a", "tools/kv_live_worker.py": "19b0b3765a56b1ae33eef62cac496a793f5844ff42e82bdc6cf21aa76617909c", "tools/kv_pilot_admission.py": "1538a51f3a7e116cd2039006b874f7a9211a0344b30f587d707fc4f7dfb34949", "tools/kv_runtime_qualification.py": "4dc0dfb45f5ed1eb5f31e3843df81d6b5c22c373d69f2d6c47629f202ac03a0c", "tools/kv_worker_supervisor.py": "7b2f25148bc0b14536321aa45b434c23424c53858760a708689a28774d585168", "tools/qualify_kv_actor_cleanup.py": "45030c6054056c91706701c1c24c5028acac53e8bd2692ae45ab3163941945b0", "tools/qualify_kv_edit_actor.py": "f5b5e884f3f4313d39ba8392ba58f1b29c9af0d1ceed9195347092fc7279cd5f", "tools/qualify_kv_reclamation.py": "b53efb547ff6dec04492d08f41b9479afee5d07bbee964911a264f453e620815", "tools/skill_eval_budget.py": "4e88ad86ab7f7e2be22e3966c825a217faf166df213e5e4b5449a72bfdb1b88f", "tools/skill_eval_budget_git.py": "488b5c963c77021fbb25802909b599e72a21b44c2cd7dcdc4dbab8d712863e30", "tools/skill_eval_pr.py": "156a2b44f98fb30e97bad875efc889855a6ca543000d95a50e7577a487c56a1d" }, "execution_id": "kv-reclamation-one-pair-2", "expires_at": 1793448000, "final_repetitions": 8, "final_seconds": 180, "final_suite": { "bytes": 538674, "path": "evidence/kv-live-pilot-20261031/plan/final-suite.json", "sha256": "ff3de0b55e0eb40120b7b3b13559b070f0ea02cd02b08606bfd5d51014352654" }, "hidden_traces": { "bytes": 109803, "path": "evidence/kv-live-pilot-20261031/plan/hidden-traces.json", "sha256": "935058055afeb86a8eab4a4f185ef3c11bccf2f6f58bcb32fb23e6c3624d9284" }, "issue": 248, "ledger": { "checkpoint_commit": "23bae89818a0eb4f1ed9ccda2361d10cbf3908ed", "genesis_commit": "6ab781845dbbb9c5f566a83134b201d75075821f", "ref": "refs/heads/skill-eval-budget/1409534281/275" }, "price_evidence": "https://github.com/InferMatrixLab/InferRSI/issues/248#issuecomment-6542191862", "prior_owner_group_statement": "https://github.com/InferMatrixLab/InferRSI/issues/253#issuecomment-6518412099; unchanged historical owner statement dated 2026-10-30, not a new account/invoice check or extension of the expired parser plan. Stop if applicability or rates become uncertain.", "provider": { "client": { "api_version": "2023-06-01", "endpoint": "https://yibuapi.com/v1/messages", "kind": "anthropic-messages-edit-check-v1", "max_output_tokens": 4096, "max_read_bytes": 65536, "max_request_bytes": 262144, "max_requests": 8, "max_tool_calls": 32, "provider": "yibuapi" }, "model": { "effort": "not_applicable", "name": "claude-opus-5-5" }, "pricing": { "currency": "USD", "expires_at": 1793448000, "group": "owner-attested default/claude-code/claude-code-pro equal-price class", "max_input_tokens": 1000000, "model": "claude-opus-5-5", "rate_evidence": "https://github.com/InferMatrixLab/InferRSI/issues/248#issuecomment-6542191862", "rates": { "cache_read": "0.4", "cache_write": "10", "input": "8", "output": "40" }, "request_fee": "0", "unit": "per_million_tokens" }, "spend_limit": { "amount": "162.72", "currency": "USD" } }, "qualification": { "archive": { "bytes": 186706, "path": "evidence/kv-actor-qualification-20261030/results/hosted-41182807900/qualification-artifact.zip", "sha256": "93baf16bdf448abaddbec567f7bbc507f4799a44ee583f16c3135c2a5029b005" }, "artifact_id": 13034764841, "dependencies_sha256": "d6d1103126fac4332a4eef8e9ecadd4e2de5b79b56d51b340abc21fc16e6541a", "head": "1566c3700079c4eab133058083bdb48c12e88333", "receipt": { "bytes": 82329, "path": "evidence/kv-actor-qualification-20261030/results/hosted-41182807900/qualification.json", "sha256": "f5228b65e714d82c5f38fa080cfb4224bad813940d07bdb4a9c2497c669da25f" }, "revision": "e4de8d8b271a9f87c8fe26e5417bbaf1ca115360", "run_attempt": 1, "run_id": 41182807900 }, "repository_id": 1409534281, "schema": "kv-live-plan-v2", "scope": "Exactly one authorized additional exploratory synthetic CPU editing pair; same original task/K/controls/verifier/metrics and equal 8+8 budgets. Prior full claim and all first-run evidence remain immutable. No third attempt, automatic retry, refund, replan or relaunch. Near-transfer feasibility only; timing is nonheldout, not a vLLM-conformance, independent knowledge-efficacy or service claim.", "suite": { "bytes": 538671, "path": "evidence/kv-live-pilot-20261031/plan/visible-suite.json", "sha256": "812ac3e5dd686271d40598e8b8d35faa6f039888f1d36852e571b11b36c600bc" } }, "plan_sha256": "d79e49f96e5373fa5305c38c9c86ad4b7dd4e9daba7cc7cfc441f5c1d508f8a7", "schema": "kv-live-authorization-v1" }