The extra-context layer for AI agents — L1 eviction, sticky L2/L3 context pools, context paging, cold-tier compression.
Registered before these solver and review calls. Two additional open issues, complete discussion snapshots, fixed checks and randomized paired order are bound by the manifest, SHA-256 aca9a7327a630a7825ef40faafc7ddd9f0d845201adbc6600097a5f90774d506. Selection followed prior results; this is not one pre-registered 30–50-task independent cohort.
| Issue | Frozen acceptance | Source-only reference |
|---|---|---|
| #3644 | Splitting one- and two-meter units into feet and inches through the expression parser returns the specified two components; direct unit splitting and conversion remain valid. | PR #3671 |
| #1905 | help(not) evaluates and returns help text; ordinary logical negation and help(and) remain valid. |
Closed PR #3470 |
The qualification ledger records baseline failures and reference passes for public and acceptance checks, with controls passing on both. The #3644 test was corrected during candidate qualification to handle an array result before freezing. These are partial checks, not upstream acceptance. The two issues touch different behaviors but share a repository and source areas with earlier tasks; independence cannot be assumed. Public comments and patches can disclose solutions or appear in model training.
The VM, runtime, paired randomization, tools, rubric, model budgets and 503-only retry policy follow the previous registration. No timeout is retried; provider failure or missing usage stops each stage. The estimated cost target is below $2, not a hard cap. Same-proxy AI reviews and short seeded issue discussions do not establish production effectiveness, savings or autonomous memory management.