Claude + Codex cloud tasks pass with Snipara
vs 32/300 cold baseline across the same continuity protocol
Method and limitsSnipara evals ask whether the agent understood the project before editing: accepted decisions, impact, handoff state, source authority, and verification.
Report published June 29, 2026. These figures describe the scenarios and scoring in that report; they do not qualify a new independent campaign or measure every coding task.
vs 32/300 cold baseline across the same continuity protocol
Method and limitsGPT-OSS, Qwen3-Coder, Devstral across hosted retrieval
Method and limits6.3K selected tokens per query vs 32K raw baseline
Method and limitscontradictions only, omissions tracked separately
Method and limitsThe baseline starts with the repo and task only. The Snipara condition starts with retrieved decisions, handoff context, impact hints, and verification guidance.
Sonnet and Opus both reached 60/60 with hosted Snipara retrieval.
Three Codex CLI model runs, same repo tasks, same scoring.
GPT-OSS 20B, Qwen3-Coder, and Devstral moved from cold starts to usable continuity.
A project-aware eval separates two agents that both compile: one respects the current project truth, the other reintroduces an old decision.
Did the agent respect the accepted product decision?
Did it see callers, files, routes, and tests before editing?
Did it resume from the right point after a handoff?
Did it identify the checks that make the result trusted?
Did unsupported facts disappear from the answer pack?
We lead with checks that can be rerun: source authority, stale handling, continuity contracts, Code Graph structure, and answer-pack grounding.
Source authority, stale caveats, routing.
Resume scope, supersession, next action.
Callers, imports, symbol cards, test hints.
Unsupported claims removed from grounded packs.
We keep the public claim simple: Snipara improves continuity and project grounding before coding agents act.
These are controlled continuity scenarios, not a global ranking of models.
Model-graded scores carry a run date and should be read as directional.
Retrieval precision depends on the project corpus, chunk freshness, and source authority.
The strongest claim is not smarter models. It is better project continuity before the model acts.