Acceptance evidence

Real acceptance

Use run-real-acceptance.mjs with an explicit target, authorization, private data directory and optional assessor/reporter profiles. The verifier requires real Nmap and ProjectDiscovery HTTPX records. It stores the receipt outside the package.

The verified 2026-10-25 local Faxbot slice used Daybreak Blue xhigh in the assessment container and a separate Daybreak Blue xhigh report container. It produced real Nmap and HTTPX evidence, 40 imported worker records, a complete coverage table, and a report-evidence-v4 draft with four cited observations and zero candidate findings. See managed assessment and real acceptance.

OpenRouter provider acceptance — 2026-10-25

This receipt records live provider checks for the OpenRouter and OpenAI-compatible work. It contains no API key, prompt body, tool arguments, or provider response body.

Documentation basis

Context7 4.1.1 resolved and queried the current upstream documentation sets /openrouterteam/docs, /ggml-org/llama.cpp, and /ollama/ollama on 2026-10-25. The implementation changes driven by that check are:

• omit tool_choice because current Ollama OpenAI compatibility supports tools but does not support tool_choice;
• accept the exact absolute model path that llama.cpp may report when --alias is not used;
• send current OpenRouter provider-routing controls for ZDR, data-collection denial, parameter support, and no provider fallback;
• retain /v1/chat/completions, exact served-model verification, and current tool-call message shapes.

Live OpenRouter draft

• Profile kind: openrouter
• Model: anthropic/claude-haiku-5.5
• Result: exact response OPENROUTER_PROVIDER_OK
• Transport: production runAgentProfile path to https://openrouter.ai/api/v1/chat/completions

Live guarded MCP assessment

The hosted model was given only getEngagement and generateWordlist. It selected and invoked the real Pentest MCP worker through the native MCP SDK.

• Engagement: real-openrouter-1792960695583
• Run: d98b5804-7b6c-47fb-b8fb-54df0ed9ec6e
• Profile: real-openrouter-assessor
• Model: anthropic/claude-haiku-5.5
• Hostile-model protection: disabled, verifying the default-off path
• Status: completed
• Imported records: 12
• Import map: rec-618cedfb-3ad8-4a01-8a6a-22081470b7a3
• Tool counts: one MCP catalog receipt, three guarded turns, two MCP starts, two MCP completions, one allowed scope decision, one successful generateWordlist, and one guarded-run completion

Live quarantined report draft

The separate real-openrouter-reporter draft profile received the bounded report packet and no assessment MCP connection. It produced and saved an analyst-review draft through createAgentReportDraft.

• Draft run: 3ebad94b-a3d8-4757-9ccc-f450cc4ed904
• Draft evidence record: rec-e2f97372-8d52-4b1e-aa44-9333655f4064
• Provider/model: openrouter / anthropic/claude-haiku-5.5
• Source: agent
• Reportable evidence records: 1
• Evidence observations: 2
• Candidate findings: 0
• Status: requires-analyst-review

The preceding run 7f81e0c0-959f-41a9-9973-f450b5ec1019 used a 1,024-output-token cap. It invoked the MCP tools successfully but ended with provider finish_reason: length; the orchestrator rejected it and persisted agentAssessmentFailed after importing 12 worker records. The completed run raised the cap to 4,096. This is retained as failure-path evidence, not counted as a successful acceptance.

Local provider status

Ollama 0.12.3 started on 127.0.0.1:11434 and served /v1/models, but its inventory was empty. No local model result is claimed. No weights were downloaded implicitly. A real local model acceptance still requires an operator-selected installed model or reviewed llama.cpp GGUF.