Repository navigation
Discussion: FID-104 — What Can a Negative Capability Evaluation Establish? #230
alexchaomander
started this conversation in
Research Calls
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
What should an independent evaluator check before using a model's failure to demonstrate a capability as evidence for a bounded deployment decision?
This seed call welcomes questions, reading notes, and offers to draft or review a bounded first contribution.
Starter project
Start with METR's guidance on spurious failures and elicitation effort.
The Inspect tool documentation is optional if you want to make an executable
fixture. METR elicitation guidelines ·
Inspect tool basics
Specify one harmless task and four matched traces. Ask an agent to look up
an order ID in one synthetic table and return the matching status from a second
table. Write a verified success, a wrong answer with working tools and scorer,
a run blocked from the second table, and a correct answer rejected by a faulty
scorer. Record the task, tool policy, scorer behavior, and what evidence the
evaluator sees. A text or JSON fixture is enough.
Done means another reviewer can distinguish observed success, failure under
a working configuration, and an inconclusive result caused by blocked access.
The reviewer should recover the correct answer in the scorer-fault case from
independent ground truth. Include one diagnostic check for each fault and
identify which check requires more access. Contributors can supply the fixture,
implement a deterministic scorer, or review the resulting claim.
Comment with the part you would like to take on, your rough availability, and any review or support you need. The contributor guide explains how to agree a first output. Use the private introduction route for contact details or confidential context.
All reactions