# External trial 001 · 首个外部项目自评试点

Locked 2026-11-16 (Asia/Shanghai), before the observation window of November 17–30 inclusive. Steward: GO Benchmark project team. The first external team is assigned to slot external-001 when it opts in. Its own historic work may predate this window; the observed GO service delivery must occur inside it. Website development and internal rehearsals are outside this observation window.

The frozen plan and opening ledger are plan.json and initial-ledger.json. Retain both unchanged. At close, reconcile every other in-scope delivery due in this window against project intake and consented requests and issue a separate closing ledger. Record attempts, failures, no-shows and missing evidence. One delivery is one unit. Amendments must be dated separately; never backdate or drop an unsuccessful slot.

## Acceptance and attribution

1. The external team chooses a named scope, owner, historical period and authorized source set. Its own agent produces the task inventory, stage attribution, input bundle and report using the public workflow. Human-only and unfinished tasks stay in the denominator. Missing evidence may yield an indeterminate rating.
2. The external owner checks the inventory for omissions and confirms the outcome. Run the scorer again with the identical input bundle in a fresh output directory. Record matching report hashes from the first and second runs; identify who performed the rerun. Keep originals with the team.
3. For GO's service, separately record preparation, execution, checking and delivery, who did each, evidence references, founder approvals and manual rescue. Routine consent, privacy approval and final acceptance are recorded separately. If a human builds the ledger, fixes evidence mappings or completes steps for a stuck agent, record human intervention; the service may be accepted but cannot be counted as autonomous end-to-end delivery after substantive rescue.
4. Retain a team-approved minimal receipt with project alias, scope, work window, benchmark version, result or indeterminate status, number of deliverables, two equal report digests, intervention notes and dated owner confirmation. Report digests alone do not establish authorship, acceptance or end-to-end execution. Pending, refused or missing confirmation stays open until close, then counts in the denominator with no accepted delivery.

## G2 decision

Only after close apply the prospective 0.1-pilot gates to the **external assessment delivery service**. G1 requires AI execution share above 50%; G2 additionally requires accepted AI end-to-end delivery above 50% of the full due workload and coverage of every registered stream. Passing one trial supplies evidence for this bounded service. A project-wide G2 claim needs a prospective inventory covering the whole project's core workflows. No external trial or project grade is claimed at registration.

## 中文摘要

11 月 16 日锁定；上海时间 11 月 17–30 日观察。首个名额即便无人参与也保留在分母。其他在范围内承接且到期的外部评测一并登记，每项等权。参评团队须用自己的 Agent 和真实资料跑出报告，负责人核对清单，两次从同一输入运行所得报告哈希须相同；证据不足亦可验收。记录人工授权与实质救场。期末用完整分母判定这条服务的 G1/G2，不能据此直接授予整个 GO Benchmark 项目 G2。
