Review paired agent trajectories, side by side
Each case is one benchmark task attempted by two models under the same harness: one failed the graded check, one passed it. Failing runs carry expert annotations (failure category, rationale, causal chain, and step-ranged workblocks) that the dashboard turns into a readable case file.
Nothing ships with this app and nothing leaves your machine: sets are parsed and stored only in your own browser.
How ingestion works
rule-based parsing, no model in the loopUpload and name the set
Drop a ZIP of trajectory JSONs, then say which trajectory type and benchmark it holds. That choice selects the reader written for that export and the taxonomy that governs it — the SWE taxonomy for Agentic Coding, the consolidated one for Agentic and Non-Agentic sets. Files are unzipped right here in the browser; nothing is sent anywhere.
Deterministic parse
Filenames are matched by pattern, JSON fields are mapped directly, and annotations are split with fixed rules. Every check is reproducible; no AI touches the data.
Review in the dashboard
Matched pairs get the full case-file treatment: verdicts, causal chains, workblocks, trace ribbons, and test results. Raw exports without expert annotations still open, showing the trace and test results. Anything odd is listed before you start.
Your data stays with you
what protects an uploadYour trajectories never leave your computer. When you upload a ZIP, it is opened and read inside your browser, and there is no server or database behind this page, so nothing is ever sent anywhere. The page itself is locked down and cannot make outside connections, even if a file tried to make it. Text from your files is always shown as plain text, so a malformed or malicious file cannot run code or change the page. When you are done, Clear loaded set wipes everything from your browser.
Upload a trajectory set
ZIP of failing and passing trajectory JSONsFolder prefixes inside the ZIP are ignored; non-matching files (for example __MACOSX entries) are skipped and reported.