$ the-wire · showcase
End-to-end runs now come back without OverflowError
By RepoJournal · Filed · About Safety Evals · Composed from the cited sources · methodology
inspect_ai fixed the crash that killed whole runs when any single sample carried an integer past int64, and cleared the compromised emSign roots out of both the core package and its sandbox tool build.
fix: samples_df and evals_df keep integers beyond int64 as exact strings instead of raising OverflowError (#6090) UKGovernmentBEIS/inspect_ai
An integer beyond int64 in a log used to raise OverflowError out of samples_df and evals_df, taking the whole results frame with it; such values now land as exact strings, like the existing mixed-type fallback, and the fallback keeps missing values as null instead of the string 'nan'. Ints are detected from the records rather than inferring from pandas dtypes, so it holds even when the frame's ...
feat(sentinel): the log records where each model call was made, and generate steps keep their ids in replay (#5990) UKGovernmentBEIS/inspect_ai
The sentinel log now records where each model call was made, and generate steps keep their ids in replay. Every stage reads one per-conversation record for step.agent_messages, so a scaffold that rebuilds or folds its input no longer changes what that field holds at AfterGenerate; tool stages no longer read the list passed to execute_tools().
feat(sentinel): a live sentinel follows the span a bridged agent's event sink gives each generate (#6004) UKGovernmentBEIS/inspect_ai
A live sentinel now follows the span that a bridged agent's event sink gives each generate, extending the same per-conversation record to agents running behind a bridge. Operators tracing bridged runs get the call site attached to the span rather than reconstructed after the fact.
chore(deps): bump certifi to 2026.10.23 to drop the remaining emSign roots (#6113) UKGovernmentBEIS/inspect_ai
certifi moves to 2026.10.23, which drops the remaining emSign roots after the sandbox tools build lock was bumped past 2026.7.22's ECC G3, ECC C3 and C1 anchors. The lock's documented command passes --no-config, skipping pyproject.toml's certifi cooldown exemption, so the lock now repeats that exemption explicitly.
Migrate TheAgentCompany browser to sandboxed Playwright MCP UKGovernmentBEIS/inspect_evals
TheAgentCompany swaps the deprecated web_browser() for Playwright MCP 0.0.83 with Chromium inside each agent sandbox, installing the pinned Node.js/MCP browser across all 45 task images and replacing aisiuk/inspect-tool-support with digest-pinned python:3.12.14-bookworm. The task version moves from 3-B to 4-B, so results recorded under the old version are not directly comparable; scoring itself...
