Repository navigation
CASSIA 1.4.0 release notes
CASSIA 1.4.0 makes the command-line interface a supported package entry point
and packages the benchmark-tested agent workflows that were previously kept in
the research workspace.
Highlights
cassia annotatesupports API providers and local Codex, Claude Code,
Cursor, OpenCode, or custom-shell agent backends.- Three explicit annotation modes:
one-shot,validated, andfused-boost. v2-compactis the Fused Boost release default. Historicalv2remains
available for exact legacy-prompt reproduction.- Stable
cassia.annotation.v1result normalization across standard and Fused
Boost outputs. cassia judgeimplements the blindedstable-judge-v1.1protocol with
rank-1 subtype determining core credit.cassia guide, nestedcassia help,doctor,validate,examples,
report, andresumecover setup, preflight, and run recovery.cassia boost,subcluster, and deterministicconsensusworkflows are
available from the installed executable.cassia agentprovides an optional long-lived R/Seurat bridge with marker
queries, cluster operations, labels, QA, finalization, and audit history.
Benchmark evidence
The release snapshot under Benchmark/cli_release/ contains fixed 50-, 100-,
and 392-case manifests and compact result records.
- Boost-100: one-shot scored 50/100 exact and review Boost scored 71/100 exact
on an intentionally hard, boost-enriched diagnostic set. There were 26 strict
rescues and five strict regressions, supporting selective rather than
unconditional Boost. - Full392 prompt ablation: legacy Fused Boost v2 and the compact treatment both
scored 296/392 exact. Correct discordance was 19 versus 19 (exact McNemar
p=1.0); estimated annotation cost was 16.3% lower for the compact arm. - The 50-case cross-model card is included as a capability profile. It is a
single-run, fixed-referee comparison and should not be read as a permanent
leaderboard.
Compatibility and packaging
- Python 3.9 through 3.13 remain supported.
- The wheel and source distribution include the CLI guide, agent system prompt,
and R daemon resource. - Existing Python APIs remain available; the CLI is additive.
Release verification
Before upload, the release artifacts must pass:
- the full CLI pytest suite;
- wheel and source-distribution metadata/content inspection;
- installation into a fresh virtual environment;
- offline example, validate, one-shot/validated shell-agent, Fused Boost
dry-run, Judge help, and packaged-resource smoke tests; - one low-cost live OpenRouter annotation smoke test;
- one-cluster live annotation smoke tests through Codex CLI, Claude Code,
Cursor Agent, and OpenCode.