# Current merged-custody baseline: two genuine human-chat controls

This is a partial current-production campaign on production `39e811c9c5226979fce42e052a27cfb28301018a` (including merged credential-custody PR17949), evaluated from source `33f99c6c5a7d6faae19cd8e47afd0d2f78882967`. It uses Paperclip Runner OpenCode with OpenRouter model `openrouter/deepseek/deepseek-v4.1-flash-1017` on actual AWS on-demand hosts with embedded PostgreSQL.

**Current baseline: 2 measured scenarios, 1 qualified end-to-end pass, 1 failed, 242 not measured on this baseline; denominator 244.** These two are explicitly requested comparisons of formerly passing controls, not two unused initial scenarios or newly granted correction budgets. Three unused project initials are selected in this same run; their pending jobs are not measurements. No prior pass is carried into these counts.

[Actual Actions run 40620786554](https://github.com/paperclipai/paperclip/actions/runs/40620786554). Source is frozen; automatic retries are zero. Both controls sent the actual canonical human query once and used one actual native run each. Both cleanup and packaged evidence privacy checks passed. Usage is metered but unpriced in the retained receipts; unknown cost is not zero cost.

| Human journey | Actual result | What happened |
| --- | --- | --- |
| **agents.change-manager** — “Have Morgan report to Riley from now on.” | **PASS, 9/9** | Actual saved manager relationship and org-view checks passed, with task/permission preservation. Native execution succeeded. Exact frozen definition replay and authenticated same-run source/build/native restore proof qualify this as current-baseline evidence. |
| **cli-operations.create-from-cli** — “Give me the CLI command to create a website task for Morgan in Launch.” | **FAIL** | Native execution succeeded and the returned supported CLI command created exactly one correctly scoped Morgan/Launch task as TODO. Later `recovery.reconcile_stranded_assigned_issue` moved the manually paused worker's task to BLOCKED. Two final-status equality assertions failed. This is a localized lifecycle/status assertion mismatch; the original failed result is preserved, not relabeled as a pass or provider outage. |

The runtime uses the product's saved `timeoutSec: 0` policy, with no artificial active-turn deadline. Original outer attempt, run-count, human-turn, spend, privacy and cleanup bounds remain in force. The legacy authored 120-second descriptor is retained as historical metadata, not represented as the effective native deadline.

## Still pending in this same pilot

- **projects.create-project** — “Create a project called Launch for the new website.” Actual project creation and duplicate-safe retry must be verified.
- **projects.project-work** — “What is left in Launch, and create a task for Morgan to fix the mobile layout there.” Actual unfinished-work summary and one correctly assigned task must be verified; a completed website is not required.
- **projects.archive-project** — “Archive the Launch project but keep its history.” Actual archive state and retained readable history must be verified.

These have no outcome credit until authenticated actual results arrive. The other 239 catalog scenarios have no new-baseline measurement in this report; historical failures and explicit external/core/public-path blockers are not silently reset or declared runnable.

## Preserved earlier evidence

[Previous OpenCode campaign and full 244-scenario coverage disposition](https://d1p6rlowie26tp.cloudfront.net/runner-e2e/campaigns/gha-40584967834-1-final34-report/) records **57 distinct measured scenarios, 6 passes, 51 failures, 187 unrun and 62 attempts** on the older production baseline. Those measurements and the three credential-scan failures remain unchanged. That run was intentionally cancelled after shared key-custody findings, not fleet queuing. Its 6 passes are historical evidence, not current-baseline credit.

[Historical local Codex partial campaign](https://d1p6rlowie26tp.cloudfront.net/runner-e2e/campaigns/historical-local-codex-chat-catalog-20261027-158measured-86unrun/) separately preserves 17 historical passes, 141 failures and 86 unrun. Six catalog journeys remain incomplete or blocked by a missing public path; implementation coverage alone is not proof of successful end-to-end behavior.

## Campaign feedback review

0 distinct feedback items await review. Collection gaps and failed submissions remain separate findings.