Eval-harness on the prod loop

Parent: agent-core

The eval-harness imports only the public agentcore facade (never internal/) and runs the byte-for-byte same agent loop as prod — so a bug found in eval is a real bug, not an artifact of a re-implementation. Fixtures are injected as legitimate alternative data sources, not mocks (the same ports the postgres repos satisfy).

It tests agent quality multi-dimensionally: grounding, privacy canaries (a secret that must never leak — interview.py), tool-use mechanics, context compaction (>32K — compaction-test.sh), cross-conversation coherence (cross-conversation-test.sh) — split into deterministic scenario runs (scenarios/*.yml via --scenarios, gold such as expect_ghost; scenario.go/runner.go) and live runs judged on a real model (*_live_test.go, and an interviewer agent driving --ask turn by turn).

This facade is also the seed of agent-as-injectable-driver — "inject fixtures and launch the real loop" generalizes into "inject capabilities + prompts and launch many in parallel."

Ask this entry a question.

The AI answers in the owner’s voice, grounded in this corpus. Bring your own key and start — it never leaves your browser.

no code?·BYOAI

Bring your own AI.

Use your own key against the owner’s public corpus. Pick any OpenAI-compatible provider, or point at your own self-hosted endpoint. Private topics return “need a code”.

· your key never touches our server — stored encrypted in your browser only
· owner pays for retrieval, you pay for inference
· private topics return “ask for a code”