ADR-0218: Release-specific compatibility for checkpointed recall¶
- Status: Accepted scope under the owner-approved bounded rollout; verification pending
- Date: 2026-10-18
- Related: ADR-0178, ADR-0199, ADR-0212, ADR-0214, ADR-0215
- Scope: The completed checkpoint comparison and one explicitly verified release
Problem¶
The owner authorized completing the recall release, opening its pull request when
ready, and following it through production deployment. The separately evaluated
runtime remains pinned to 16153a5f, with the approved formation repair recorded
under its own identity. Production has meanwhile advanced to 9456e67c without
waiting for this isolated evaluation, as the owner directed.
The complete ADR-0199 input census finds 109 changed included files between those commits, including 29 application source files. These are real differences, not just a merge commit label. In particular, retained-recall validation now checks semantic authority only for semantic receipts and withholds unavailable source recovery. The run loop reserves time for synthesis and rechecks resumed tool admission. Browser and calling composition also changed. The proposed release adds the separately verified checkpoint collector and formation deadline repair.
ADR-0199 currently requires new matching measurements for any changed evaluated input. The checkpoint input-binding implementation correctly refuses this target; its publication also explicitly grants no activation authority. Passing the current comparison cannot establish complete byte equality with production.
Bounded compatibility exception¶
The owner authorized the compatibility route on 2026-10-19, alongside the release-specific acceptance revision in ADR-0220. Admit the completed 1,600-row comparison through a separately versioned compatibility receipt for one final release tree. Preserve the existing exact-input route and its refusal tests unchanged. This exception must never masquerade as an ADR-0199 equality receipt.
The new route requires all of the following before activation:
- Retain and mechanically reconstruct every original observation, bank, event, segment and charge. Preserve the original identities and separately named formation repair. Complete all 400 scenarios and both paths without favorable resampling, unknown-work retries or invented timing or usage.
- Apply the explicitly revised bounded-rollout criteria in ADR-0220 while retaining every original numerical result and failed gate. Obtain a genuine independent human review of difficult answers, interruptions, repairs and uncertain costs, including explicit residual-risk findings. No human findings or successful compatibility assessment are implied by scope approval.
- Inventory every included file in the measured, collector and final release trees. Enumerate every changed file and its exact hashes, behavior disposition and verification evidence. No caller-supplied wildcard, ignored source subtree, aggregate pass flag or inherited branch name establishes compatibility.
- Verify the production-shaped effective configuration, model and embedding identities, schema, formation requests, context limits and recall budgets. Keep the actual runtime changes visible. Specifically exercise retained recall, semantic-authority failures, source-recovery failures, synthesis time reserves, resumed approval handling, and unchanged successful reader/tool requests. Any material model-visible or policy difference not established compatible by these checks requires matching additional measurements before admission.
- Bind the accepted compatibility assessment and owner decision to the exact final reviewed source, publication bytes and actual deployment tree. Recheck source and configuration at deployment, and revoke admission on mutation. A later release cannot inherit this exception merely because it descends from the admitted commit. Complete CodeRabbit, final-head CI and ordinary production delivery verification for this release.
This is an authorized change to the evidence-transfer rule, not an implementation or
a claim that compatibility is already established. The existing owner instruction
to ship does not silently turn changed inputs into identical inputs. The dev
push hold and unrelated production work remain unaffected.
Alternative¶
Keep ADR-0199 unchanged and measure a complete new comparison against the final integrated runtime, retaining the current completed evidence separately. This avoids the compatibility exception but requires another full collection and human review. It cannot be represented as merely continuing the current measurement.