[kikashy]

Implement the explicitly selected retained-column/1 Base consumer needed by Judgment-Pack/judgment-pack-agent#171. Authenticated selection binds the published Spec RC2 identity through conversion, preparation, persistence, restart, export and offline verification. Ordinary routes retain their existing contracts.

The batch covers all five conversions and eight research types with original raw values and declaration provenance; complete-artifact limits remain enforced. Selected Core 0.3 Base joins the existing Core 0.2 path. Generated inputs require the existing independently approved producer, signature, request commitment and freshness checks. The selected Runtime companion is pinned separately to published v0.39.0-alpha.2 (57c854abb9ace8a91013a75ab688bd70d063797c), preserving historical integration pins.

Published Spec corpus coverage is mapped to its actual consumer owners. Preparation-only synthetic exports are not evidence of enclosing release/chain authenticity. Two actual published-Runtime lifecycle cases cover SQLite first-save/restart, the signed producer fixture, preparation binding, exports 2/3/4/5 and offline verification. These do not establish the future Desk adapter's original-history inventory or grant/stdout bounds. Extended Runner support is outside this selected Base release.

The verify-preparation --capabilities closed interface and report capability declaration advance to version 2 for the explicit conversion selector and selected refusal vocabulary. Ordinary report members, exit statuses and bounds remain unchanged; ordinary extra preparation fields still produce the canonical preparation-differs finding. Unit checks use ten exhaustive, disjoint partitions under the existing race instrumentation and time limits, retaining every test and the mandatory SQLite/memory conformance assertions.

The batch also includes the approved Devkit #57 checklist correction: an incomplete selected checklist cannot return success. CI evidence includes the companion and mandatory selected lifecycle inputs. Release notes prepare the scoped v0.26.0-alpha.1 release; final v0.26.0 scope is preserved.

Review corrections preserve the verified document/frame limits through retained replay and independently bound parsed records, release bytes and run bytes during offline verification. Focused exact-limit/one-over cases, published-Runtime exports and meaningful safeguard-removal failures cover both findings. The original reviewer confirmed both findings CLOSED at the corrected head, with no new material finding. The verbatim confirmation and dispositions preserve the remaining release gates.

Checks

Candidate b68436932be8316ab86a3ad4301e8ef8b93be25d, clean checkout, Go 1.26.5. All tracked bytes match the source snapshot that passed the affected unit check and required full checklist below. The two tested files were committed with sign-off without content changes.

At the previous head a3fb17d5ac31ae9d780ba2d7210c6ddbc073ab11, local full validation passed, but Linux and macOS CI each reached Go's existing 600-second package timeout in the first broad ColumnRole partition. The correction splits that family into six smaller groups, producing ten exhaustive, disjoint main partitions overall. Race instrumentation, package/step timeouts, companion pins and both mandatory store-kind checks are unchanged.

The actual Go selector probe passed all 48 handwritten current/future leaves exactly once, including examples and fuzz seeds. The 15 focused CI-evidence tests passed, including intended omission, overlap, anchor, race, timeout, pin and failure-propagation controls. The affected unit step passed (1 passed, 0 failed; 1,248.415 seconds), followed by the full checklist (16 passed, 0 failed, 0 skipped, 0 not run; 882.288 seconds). Neither run changed source or crossed the disk reserve.

== check summary (dev.json) ==
handover: boundary 2 paths, shared 2 paths; see AGENTS.md "Before you hand over"
PASS     local-paths              0s
PASS     agent-docs               0s
    sync-agent-docs: CLAUDE.md matches AGENTS.md
PASS     diff-check               0s
PASS     go-fmt                   0s
PASS     go-vet                   25s
PASS     build-pinned             35s
PASS     manual-gate              24s
    13 lines matched; last: ok  	github.com/Judgment-Pack/judgment-pack-runner/internal/runner	12.639s
PASS     unit-tests               698s
    60 lines matched; last: ok  	github.com/Judgment-Pack/judgment-pack-runner/internal/runner	218.497s
PASS     chain-tests              13s
    ok  	github.com/Judgment-Pack/judgment-pack-runner/internal/runner	12.474s
    ok  	github.com/Judgment-Pack/judgment-pack-runner/cmd/jpack-runner	5.834s
PASS     openapi                  0s, Python 3.12.11
PASS     build                    3s
PASS     ci-evidence              7s, Python 3.12.11
    Ran 15 tests in 7.119s
    OK
PASS     release-toolchain        0s
PASS     third-party-notices      0s, Python 3.12.11
PASS     release-archives         73s, Python 3.12.11
    Ran 59 tests in 73.509s
    OK
PASS     release-programs         0s
== 16 passed, 0 failed, 0 skipped, 0 not run ==

Focused evidence includes both published-binary lifecycle tests (2 passed, 0 skipped), published RC2 conversion/capability/preparation adapters, original custody and complete-envelope boundaries, and intended safeguard-removal failures with restored positive checks. Exact-head CI run 40535535227 passed on Linux (12 checks), macOS (12 checks) and Release contents (4 checks); DCO passed. The independent combined PR/release review and finding-specific confirmation are complete. Main-branch evidence, signed tagging and all publication gates remain required. The approved Codex review exception is same-vendor, not cross-vendor compliance.

[kikashy]

Runner PR183 independent combined PR and release review

Decision: request changes for R183-01 and R183-02. Both are MEDIUM contract-enforcement findings demonstrated through the existing candidate snapshot CLI. No HIGH finding was established. This report does not approve merge, tag or publication.

Reviewed candidate d1054cdfbc68350d09f85aea95cc90abe16632eb; PR base 530afd545c9a19ecd2096ce8600688a4238c1e7d; previous release v0.25.0 source 604c1b7dad897374d4872722bd11ace592fe5be9. The checkout matched the candidate and remained clean. The release range contains 127 changed files; their hashes are in source-inventory.json. Spec RC2 checkout matched bcac15b7537ce21879201b5e5dfe795496225121.

The review worker was selected explicitly as gpt-6.1-sol/high for contract, security, custody and replay boundaries. This is an independent OpenAI review under the authorized same-vendor exception, not different-vendor compliance. The reviewer read the brief, repository AGENTS.md, release procedure, PR/issue, published Spec contract/source, implementation/tests and check artifacts. Builder handoffs/reasoning were not read; no descendants were launched.

R183-01 — MEDIUM: disposition replay discards effective Runtime limits

Location: internal/runner/mapping_v2_export.go:198–202; related enforcement at internal/runner/runtime.go:392–398 and internal/runner/column_role_v1_capabilities.go (runtimeConversionInput).

VerifiedRun.Disposition calls runtimeConversionAdmission, which validates the Runtime tuple but throws away its minimized document/frame maxima. It then constructs an input containing only Facts/Evidence for evaluateWith. Consequently inputConversion(input) is nil, runtimeConversionLimits returns nil, and the ordinary evaluation helper skips runtimeConversionInput. The replay can evaluate an artifact larger than the exact supporting Runtime's positive declared capacity.

Public-path reproduction: python3 probes/cli_replay_limits.py. The probe uses the unmodified candidate snapshot CLI and a disposable script Runtime, whose executable hash is frozen in the disposable release and whose release digest is recomputed independently. It preserves the preparation, acquisition and audit-binding bytes. This synthetic Runtime isolates admission behavior; it is not claimed as published Runtime interoperability.

• Exact installed tuple, documentBytes 1048576: exit 0, verified-disposition, one evaluation call.
• Same tuple, documentBytes 1, retained pack 7307 bytes, bound mapping 1063 bytes: exit 0, verified-disposition, one evaluation call. Expected refusal before evaluation.
• Unsupported Runtime revision: exit 1, no evaluation call. The identity safeguard remains effective.

Evidence: probes/replay-limit-result.json, the three *-report.json files, stdout/stderr, runtime-fixture, replay-output.json, replay-export-v2.json and evaluation-calls.txt. An initial setup-only attempt reordered the audit binding's members and stopped at binding-mismatch; its result is preserved in replay-setup-incomplete-result.json and is not counted as control evidence. The final fixture preserves those bytes and reaches replay.

Required correction: retain/recheck the effective Runtime document/frame limits against the verified complete input and retained release pack before replay evaluation. Passing the verified prepared input through the common evaluation helper, or explicitly applying its negotiated admission before stripping transport data, would preserve the boundary. Verify exact/one-over lower document and frame limits on the public replay path, including zero evaluation effects on refusal.

R183-02 — MEDIUM: offline exports bypass independent record and storage ceilings

Location: internal/runner/mapping_v2_export.go:247–255 and internal/runner/audit_bytes.go:122–135; compare store guards at internal/runner/store_contract.go:247–262.

The offline reader applies export wire/base limits but does not independently apply the selected embedded release/run storage ceilings. Its record check measures only Run.AuditBytes for version 3 and later, leaving the parsed version-2 Run.Audit unconstrained by recordBytes. The larger export budget therefore admits bundles that could not pass the selected artifact boundaries. The Spec explicitly requires preparation-verification support for recordBytes, releaseStoredBytes and runStoredBytes independently of export size; this is not waived by an export that retains no exact record bytes.

Two public-path reproductions use the original valid Core 0.2 version-2 lifecycle export, its unchanged trusted release digest, profiles, selector, acquisitions and preparation:

• python3 probes/cli_record_bound.py: original export exits 0. Adding an opaque audit member produces an exact audit object token of 9438906 bytes, above 9437184 (9 MiB), inside a 9475550-byte export. It still exits 0 with verified-inputs and matching preparation binding.
• python3 probes/cli_storage_bound.py: adding a problem string of 41943040 bytes yields a complete run token of 41957382 bytes, above 41943040 (40 MiB), inside a 41981402-byte export. It also exits 0 with verified-inputs and matching binding.

Evidence: probes/record-bound-result.json, probes/storage-bound-result.json, their exact input exports and stdout/stderr. These witnesses demonstrate bounds enforcement, not forged signatures or changed source authority. The record fixture follows the existing boundary test's opaque audit-padding approach, but covers the version-2 path omitted by that test.

Required correction: independently validate selected complete release/run artifacts and the parsed record representation on offline entry, across export versions 2/3/4/5, while retaining independent original-record, chain and sidecar limits. Keep exact-bound positive controls and reject one-over artifacts under an otherwise admissible export envelope.

Scope and evidence assessment

Reviewed the full PR and previous-release range: all five conversions, eight research types, ordered diagnostics, exact raw value/declaration custody, lineage references and occurrence binding, selected identity/roles, trusted producer tuple and capability limits, signature/request/freshness verification, SQLite and durable-operation custody/restart, exports and public verifiers, Core 0.2/0.3 evaluator matching, discovery v2, OpenAPI/docs, checklist correction and the main-CI release-evidence workflow.

The code keeps claim canonicalization separate from raw custody, recomputes transformations, checks the approved producer/profile and original snapshot relation, and refuses unsupported selected execution admission. Test-only synthetic installation/acquisition seams remain private; public verification retains the full acquisition reader. The release notes correctly limit the claim to selected Base at draft maturity and distinguish producer consistency from future Desk original-history/grant/stdout compliance. No additional material finding was established in the remaining reviewed changes.

Reused exact-head local evidence: round3-final-check/result.json, full.log and per-step logs: Go 1.26.5, affected unit step passed, followed by 16 passed, 0 failed, 0 skipped, 0 not run. Reused required CI run 40494251522: Ubuntu Runner, macOS Runner, Release contents and DCO all succeeded at the exact candidate. The auxiliary update check was skipped. These are PR checks, not push-to-main tag-time release evidence.

Reused the four-archive snapshot evidence and the actual published Runtime alpha2 lifecycle/CLI artifacts for both Core versions, including signed supplied-entry verification, first-save/restart, exports 2/3/4/5 and preparation recomputation. Runner is a local snapshot, not a published release. These fixtures do not establish sealed-session completeness, independently witnessed history or live Desk producer acceptance. The focused reviewer probes add no such claim.

Only the three bounded no-build CLI probes above were executed by this reviewer. No broad suites, builds, network mutations, repository writes, cleanup, safeguard removals or heavy concurrent probes were performed. TMPDIR stayed under this review root; disk free before the largest probe was 4906672128 bytes, above the 2576980378-byte reserve. Routine setup failure is recorded separately; no permission/safeguard refusal occurred.

Remaining gates

1. Builder fixes R183-01/R183-02 together; focused regression checks, repository-required integrated validation and exact-head CI on the changed candidate. This reviewer confirms the findings against the final fixes; existing green evidence does not qualify a new head.
2. Parent posts this record and explicit finding dispositions. Required maintainer merge/release approval remains separate.
3. After merge, qualifying push-to-main CI/evidence at the exact release commit; signed annotated new tag, tag admission and 48-hour evidence gate, release-toolchain packaging, archive/native smoke checks, checksums and attestations, production approval and published asset verification under docs/releasing.md.
4. Preserve the declared alpha/Base/draft scope and separate future Desk producer/history and complete production compatibility gates. Do not describe this review as cross-vendor or these fixtures as whole Desk/live producer acceptance.

The evidence manifest hashes this report, probe artifacts, immutable source inventory, consumed check records, published contract inputs and the reused executable/evidence identities. It grants no publication authority.

Coordinator dispositions

Finding	Disposition
R183-01	Open. The original builder is implementing effective Runtime limit enforcement on replay, with exact-bound and refusal-before-evaluation regression checks. Merge and release remain held.
R183-02	Open. The original builder is implementing independent record/release/run limits across selected export versions, with boundary regressions. Merge and release remain held.

The fixes will be grouped, checked on the changed candidate, and returned to this independent reviewer after exact-head CI is green. Standing maintainer merge/release authorization does not waive these gates.

[kikashy]

Runner PR183 finding-specific confirmation

R183-01 CLOSED. R183-02 CLOSED. No new material finding on the changed code. The review recommendation is to proceed to the maintainer's merge/release decision for the scoped alpha candidate, subject to the remaining gates below. This report does not itself authorize merge, tagging or publication.

Reviewed head: b68436932be8316ab86a3ad4301e8ef8b93be25d, clean checkout. Compared all six changed files against the original reviewed head d1054cdfbc68350d09f85aea95cc90abe16632eb. The original combined PR/whole-release report remains the review of unchanged material; it was not restarted. The original PR base and v0.25.0 release-base identities remain unchanged.

Reviewer selection remains explicitly gpt-6.1-sol/high, retained from the original review for replay/custody and independent complete-artifact boundaries. This is independent same-vendor OpenAI review under the approved exception, not cross-vendor compliance. No descendants were launched, and no builder reasoning or handoff prose was used.

R183-01 — CLOSED

VerifiedRun.Disposition now passes the complete verified prepared input into the common evaluation helper (internal/runner/mapping_v2_export.go:201). The helper therefore sees the selector and retained documents/frames and applies the actual Runtime's minimized capabilities before evaluation. Runtime executable pinning and rehearsal checks remain intact.

Executed the adapted original public CLI reproduction using the exact-head snapshot:

TMPDIR=<confirmation-root>/tmp python3 <confirmation-root>/probes/cli_replay_limits.py

The fixture uses a disposable script Runtime frozen by its executable digest in an independently re-digested release. It preserves source, preparation and audit-binding bytes; it isolates admission rather than claiming production Runtime behavior.

Case	Exit	Evaluation calls
Full reported capacity	0, verified-disposition	1
Exact lower document/frame maxima: 7307 / 1629 bytes	0, verified-disposition	1
Original witness, documentBytes 1	1, documentBytes limit	0
Document one-over: actual 7307, maximum 7306	1, documentBytes limit	0
Frame one-over: actual 1629, maximum 1628	1, frameBytes limit	0
Unsupported Runtime tuple	1, unsupported selection	0

Evidence: probes/replay-limit-result.json, report fixtures, exact export, script, outputs and evaluation marker. The committed regression test also covers asserted/generated frames, mapping/pack-dominant documents, exact/one-over limits and unsupported identity through VerifyDisposition.

R183-02 — CLOSED

exportWithinLimits now independently measures the parsed audit token and complete release/run owner encodings for every selected export version (internal/runner/audit_bytes.go:112–127). Version 3–5 original audit bytes, sidecar and chain guards remain separately enforced. The correction is selected-only and does not enlarge ordinary limits.

Executed the original public version-2 witnesses, without rewriting their inputs:

TMPDIR=<confirmation-root>/tmp python3 <confirmation-root>/probes/confirm_artifacts.py

• Original valid export: exit 0, verified-inputs, matching preparation binding.
• Original 9438906-byte parsed audit token, above 9437184: exit 1, the parsed audit record exceeds selected record limit.
• Original 41957382-byte complete run token, above 41943040: exit 1, retained run exceeds selected storage limit.

Evidence: probes/artifact-result.json, stdout/stderr and the hash-bound original witness inputs. Inspected the committed exact/one-over tests for parsed record, original record, complete release and complete run across exports 2/3/4/5; original-record tests correctly exclude version 2, which carries none. The tests keep the other artifact budgets admissible and verify signed version-5 positives. Reused the actual full/check and CI execution of those tests rather than rerunning the expensive family.

Inspected all four safeguard-removal logs: removing replay admission, parsed-record, release or run guards caused the intended changed-behavior failure, not setup/compilation failure. The initial unknown-tuple refusal-substring mismatch in focused.log is historical expectation failure and is not counted as an acceptance pass. The restored source is covered by the later successful controls/full validation and current exact-head CI.

Partition and failure-semantics assessment

The ten primary selections in scripts/checks/unit-tests.sh:26–35 divide the costly release/run children, their exact complement, other costly ColumnRole cases, remaining retained cases, Workflow, ordinary A–M and the final complement. The anchored exclusions preserve future similarly named siblings. Examples and fuzz seeds remain in the final complement. Parent setup can execute for the split subtest groups; the claim of disjointness concerns test leaves. The separate pre-existing eleventh invocation intentionally reasserts both store kinds.

Independently parsed the ten recorded actual-Go JSONL traces, checked them against their report and current selector source hashes, and confirmed 48 distinct current/future representative leaves, each selected once, with no failed/skipped event. Reviewed the selector fixture and the committed test's enumeration of current top-level tests plus future/alphabet/example/fuzz representatives. No omitted or overlapping family was found.

All primary invocations retain go test -race, ./..., all six pinned companion environment variables, default Go package timeout and ordinary caching semantics. No timeout increase or test omission was introduced. set -e and the dedicated conformance command still fail on nonzero Go exit; explicit PASS assertions still require both sqlite and memory. The committed controls exercise failure at each of eleven calls, omissions/overlaps, removed race/pin/default-timeout guarantees and either removed store assertion. Current CI's 15-test ci-evidence check passed on both platforms; no wiring-only result is substituted for actual Go execution.

Validation and identity

• Required exact-head CI run 40535535227: Linux 12 passed, macOS 12 passed, Release contents 4 passed; each has 0 failed/0 skipped/0 not run. Raw check-runs evidence additionally confirms DCO success, all at b684369; auxiliary update was skipped. Main-only evidence-capture steps are appropriately skipped in this PR run.
• Reused local affected-unit success (1248.415 s) and full checklist 16 passed, 0 failed, 0 skipped, 0 not run (882.288 s), Go 1.26.5. These ran with captured HEAD a3fb17d... before the final sign-off commit. Independently verified that before/after snapshots match and all 1228 tracked-file hashes match the current clean b684369 checkout. This establishes source equivalence without relabeling the earlier captured commit as b684369. Current CI independently ran at b684369.
• Used the supplied b684369 snapshot executable only: SHA256 a2a4d21ec6775e52b87518538a63b27d01a812d7f6bee83ead7304baa7fb57dd. Reused the exact-head four-archive snapshot/checksum/verifier and native Linux smoke record. It is a local snapshot, not a published Runner release or four-platform native certification.
• Additionally executed current snapshot verify-run with --require-signed --require-bound --runtime on both existing Core 0.2 and Core 0.3 version-5 exports. Both exited 0 with verified-disposition, matching binding, signed record and zero findings for the supplied chain entry. Runtime was published alpha2, SHA256 bf063946dc1fbcf0dfb445c9e57db7c0520551655849c8df0b19284e9d934f53. Complete commands/outputs are in probes/published-runtime-result.json. This establishes replay of those retained fixtures, not sealed-session completeness, independently witnessed history or actual Desk producer acceptance.

The confirmation root was initialized with explicit jpdev clean init --root ... --session ...; TMPDIR remained inside it. Only bounded no-build CLI probes and read/hash checks ran. No broad suite, build, CI watch, repository/Git/public write, cleanup/seal or descendant work occurred. Free disk before probes was 3878305792 bytes, above the required 2576980378-byte reserve. No safeguard refusal occurred in this confirmation.

Remaining gates

Parent posts this confirmation and dispositions; maintainer merge approval remains separate. After merge, qualifying push-to-main CI/evidence must cover the exact release commit. Signed annotated new-tag admission, the 48-hour evidence gate, release-toolchain packaging, archive/native smoke, checksums/attestations, production approval and publication/asset verification remain required under docs/releasing.md. Any subsequent substantive candidate change needs its applicable checks and review.

Preserve selected Base/draft/alpha scope, all five conversions/eight original types and ordinary compatibility. Future Desk original-history/producer grant/stdout implementation and corresponding complete production compatibility evidence remain separate obligations. This confirmation does not complete Agent #171, establish cross-vendor compliance or authorize a broader product-support claim.

Coordinator dispositions

Finding	Disposition	Evidence
R183-01, MEDIUM	Fixed in this PR; original reviewer confirms CLOSED at b68436932be8316ab86a3ad4301e8ef8b93be25d.	Complete prepared input reaches capability enforcement; exact-limit replay passes, one-over document/frame limits refuse with zero evaluation.
R183-02, MEDIUM	Fixed in this PR; original reviewer confirms CLOSED at the same head.	Independent parsed-record and release/run storage limits apply across selected exports; original oversized witnesses refuse and positives pass.

The coordinator verified all 111 evidence-manifest entries and the raw focused probe outcomes. Confirmation report SHA-256: 810cce7770b96c6593e3e2e670a64cbcccdb51a04aae358e8ceea869a6db5d0d; manifest SHA-256: b3ab5d826c10db7025ef3511278ae11ba147d474c15175b0b67b8789aed126bc. Exact-head CI and DCO are green. The maintainer has provided standing merge and release authorization for this batch; all main-branch, tag, production-approval and published-artifact gates remain. No tag or publication is claimed by this record.