Register runtime comparison protocols and native readiness evidence

[seathatflowsinourveins]

Scope

Register the runtime comparison protocols and retained native readiness evidence as the first half of the PR633 split. The controlled network fixture now passed its frozen 51-command run, including allow/deny/allow behavior and independent cleanup observations. Coding bounds remain a prospective source proposal; all selected modes remain null, full_usage is unchanged, and no measured default or comparison is admitted.

• Base commit: cac8700ba914950266272347468bff7ad630a4bf
• Reviewed head requested: cfbb6bb8be46b29687c8e2101b1d54b01086487f
• Lane: lane:foundation
• Owned paths: runtime comparison blueprints/source receipts and their evidence registry rows. The five catalog/freshness/decision files remain in a separate dependent change.

SOTA sources

• Harbor 0.23.0: original trial/timeout contracts and unchanged Docker egress-sidecar sources.
• Pinned SWE-bench Verified task originals: all 24 selected task.toml files declare agent/verifier3000s and build1800s.
• Inspect 0.3.273: cooperative reported-usage limits, separate scoring allowance and model/sample retry scopes; no genuine aggregate token-cap qualification.
• Docker CLI and Compose: isolated empty client configuration, native CIDfiles, exact inspect LF formatter and environment-free render flags. Delivered image digests and first-party source pins are retained in the receipts.
• OpenHands SDK 1.50.1, Codex 0.160.0 and Pi 1.0.0: native coding bounds and distinct retry scopes. The proposal covers one future coding M9 checkpoint; actual configuration and delivery remain unqualified.
• OmniRoute base23a1148, releasec1e30b7 and PR15167 read at0585aba: native catalog/alias precedence, token switches and mechanical assertions; no new feature or route acceptance.
• Codex Action v1.12 and GitHub Agentic Workflows0.89.21: prospective public-marker inputs and supported controls, with actual fresh-home/keyless delivery unqualified.
• Martian reviewer oracle, RD-Agent v1.0.0 and Qlib v0.9.7: scoped source proposals, no oracle/default adoption or trading experiment.
• The remaining pinned candidate/extension originals and failure history are inventoried under evidence/artifacts/runtime-jobs-source-intake-20261003/; historical acceptance is not relabeled as a new run.

Evidence-class table

Claim	Evidence class	Command / receipt
24 original task budgets and prospective method	source_review	coding-original-budget-bindings-20261003.json; quality-time-bound-proposal/
Three digest pulls, unchanged sidecar build, four image observations	native_proven	network-policy-native-dependency-install-20261003.json; eight native terminal exits0
Render and eleven no-network metadata helpers, independent name/ID cleanup	local_integration	network-policy-native-metadata-20261003.json; 46 native commands, no policy/model/task run
Frozen controlled allow/deny/allow, owner checks and independent cleanup	local_integration	network-policy-native-controlled-fixture-20261003.json; 51 native commands: 35 exit0, 15 expected absence exit1, one expected denial exit52; no services remain
Native image-volume projection rejects optional missing key before UP	local_integration	network-policy-native-volume-first-precheck-20261003.json; original exit1 retained; prospective source repair re-reviewed before run
Coding bounds and SDK role evidence	source_review	coding-native-bounds-source-proposal-20261003.json and sdk-role-source-carry-20261003.json; their earlier specifically checked-source locator gap is historical; actual roles remain unqualified
SDK48 original-source reconciliation	unchanged upstream historical tests	codex-sdk48-originals-reconciliation-20261003.json; original full-suite nativec41c72 exit0/48PASS/5suites, source/binary/ten-stream identities independently matched; no root rerun or native fixture regrade
First absence predicate rejects native LF	local_integration	network-policy-native-helper-first-precheck-20261003.json; original FAILED predicate retained
Native Ultra, carrier inputs and gatherer timer scopes	source_review	native-ultra-gateway-source-qualification-20261003.json; coding-carrier-native-inputs-source-proposal-20261003.json; research-native-bounds-source-proposal-20261003.json; actual delivery remains unknown
Unchanged selected SDK spawn-mock unit tests	unchanged upstream selected tests	codex-sdk-selected-exec-unit-20261003.json; native exit0/14PASS, initial driver127 retained; matching earlier14PASS receipt also retained; no native CLI/provider role proof
Native suffixless Ultra input, GitHub marker, reviewer oracle and trading R&D	source_review	seven new scoped source payloads preserve zero defaults/false freeze eligibility, original failures and prior drafts; actual routes/common tasks remain unqualified
Gateway token delta originals	source_review	gateway-token-native-source-proposal-20261003.json;33 exact source bodies/13 unchanged helpers and six reused cache controls;95% phrase/3% mean/2% median assertions are mechanical, not provider/useful-task results
Network original-stream and observation map	local controlled fixture projection	network-policy-native-original-stream-map-20261003.json;53 retained command records/106 streams/53 completion JSONs,61+2 ordered tool returns including10 missing exits, eleven final observations and the failed precheck preserved;216404B/SHA2aed4c0b; no controls rerun
Eight job defaults and comparative results	source_review	0/8 measured defaults; no comparative provider/model run; all actual execution gates remain pending

Local commands run

rtk python3 scripts/validate.py
exit0: 69 components, 9665 hashed files, 4 profiles, 193 receipts
rtk python3 -m unittest tests.test_runtime_worker_skills
exit0: 22 tests
rtk git diff --check
exit0

The sealed evidence-base and catalog-split trees separately passed the existing 186-test suite (two skips); the unchanged22-test runtime-skills check is reused with matching inputs. These are repository integration checks. Native source/install/helper/controlled-fixture checks are distinct from unchanged upstream evaluator acceptance. The seven source payloads were previously validated at14c6e367. The new original-stream map and registry passed native validator0/9665; all9664 prior full rows and their order relative to14c6e367 are preserved. Across the whole PR from cac8700, all9541 foreign file rows, their order and all non-file metadata are preserved. Earlier matching test evidence is retained with its original inputs; no control or model comparison was rerun to publish the map.

Decision record

blueprints/runtime-workers/comparisons/quality-only-source-proposal-20261003.md and evidence/artifacts/runtime-jobs-source-intake-20261003/quality-time-bound-proposal/ retain the dated prospective method and source correction. coding-native-bounds-source-proposal-20261003.json adds the prospective native-bounds method and the corrected Pi factory citation at line216. No comparative runtime decision is made here. Exact-head Claude-family read remains required alongside the consequential source reads before any new freeze/merge.

Host evidence

No files under evidence/hosts/ change. Existing 193 receipts and29 convergence records are preserved. Source/installation metadata does not change platform status.

Checklist

• No GitHub Actions change in this evidence-only PR.
• No credential/auth values or private host paths are committed; native sign-ins stay native.
• No new required secret, paid hosting, subscription or billing surface.
• Peer-owned worktrees and Claude-lane paths are preserved.

[seathatflowsinourveins]

Exact-head Codex SOURCE ACCEPT WITH RESIDUALS on e4d95062be3822c670c74dbd61d94a8eca97e20d (#669). This is prospective method acceptance; selected modes remain null, the default remains full_usage, and freeze_eligible remains false. The separate D80 catalog head d80a089ce2e61df8a2523473f730aa5c932b4625 carries the five previously reviewed catalog payloads byte-identically; it is not current-main acceptance.

Consequential trigger: this revision proposes omitting the otherwise mandatory gross-200000 token envelope in a future, explicitly selected quality-only study. One native Codex reader requested gpt-6-astra/max through the existing OmniRoute pool and completed with exit 0, source acceptance and residuals. I independently checked the 129 mapped repository bodies and all 52 pinned primary bodies against complete Git blob/SHA256/byte bindings. Requested route/model is recorded; served family and complete provider usage remain unknown. This read ran no comparison, installation or role qualification.

The prospective proposal is coherent as an inactive conditional exception: exact accepted/reconciled contracts and a new preregistration must explicitly choose quality_only before candidate output. It changes only the named token gate clauses. Historical/full_usage fields and failures remain intact. Original partial/native counters retain their values, types, presence and attempt/role coverage; unknown complete tokens and cost remain null. Tokens, elapsed time and resource footprint do not rerank quality ties. Numeric quality rules, task/question denominators, oracles, incumbent ties, failure retention and challenger ordering remain unchanged; evaluation's unbound numeric gain/order still block freeze.

Execution remains HOLD on the retained prerequisites. In particular:

• Research still requires 48 valid reports while its ceiling is 48 attempts; invalid attempts plus mandatory whole-block recovery have no headroom. Resolve that scope before execution without silently increasing it to 54, dropping questions or counting incomplete coverage as a winner.
• Coding still lacks qualified clock delivery and finite tool/internal-retry/overhead bindings. The selected 24 task originals each specify agent/verifier 3000 seconds and build 1800 seconds: pinned task source 86723674f04e4209ac479d0fb75d9d9f44b4377e. Harbor's framework default 600 seconds does not establish those selected task build limits. The historical 195-minute block remains failed against its 180-minute reservation.
• Inspect's cooperative token limiter depends on supplied usage and supports no complete all-role gross-token guarantee. Its additional scoring allowance and model retry scope must be separately bound. Message/turn limits do not establish search/fetch quotas. Scoped source inspection of four Harbor files is not a client/ecosystem absence claim.
• Actual positive provider-role identities, exact treatment/task/oracle/attempt hashes, native time/tool/retry/phase/breach evidence, image/network/isolation controls, independent lifecycle/cleanup and a sufficient exclusive owner window remain required. GOST/network fixture holds and failed conditions remain retained.
• R13 EOF/outer custody and full SDK/native-host qualification remain open. Codex SDK exec.ts at a956835d020762cb2b570053af06f643a11c0ecc drains stdout before awaiting exit; thread.ts keeps consuming iteration after turn.completed. A source verdict and the owner's separate 48-test scoped SDK prerequisite do not qualify this lifecycle or real provider roles.

No mode, actual budget, freeze, route/default, pool/window or service configuration is selected by this comment. Existing user authorization and runtime-owner scope govern admission; no new approval flow is introduced. The runtime goal retains both comparisons and SDK role custody. Source agreement does not complete either measurement or the root goal.

[seathatflowsinourveins]

SOURCE ACCEPT WITH RESIDUALS for #669 head e0c327f44fc8890d96dc2f9054f4ac3f5732b43f; native execution, delivery and role qualification remain HOLD. This carries the earlier e4d9506… source verdict through 454e4d… and this exact head.

Root independently checked all seven present old/new/main body bindings for the latest delta, all 117 head membership tuples and all 71 named pinned primary bodies (2,920,806 bytes). The latest delta adds four JSON records and updates the registry: all 112 existing owned payloads, 10,354 foreign old Git memberships, 9,653 prior full ordered registration rows and all non-files registry fields remain exact. Earlier e4→454 five-record additions were independently checked separately. These counts describe source carry, not a prospective merge against current main.

One consequential requested Astra/max source read through OmniRoute completed, outer exit0. Root re-read the operative primary sources and accepts these distinctions:

• Native Ultra orchestration is conditional. Inference effort normalization is separate; the bundled Sol override is supported xhigh. Actual mode/catalog/model/effort remain unobserved.
• Tighten the receipt's HTTP sentence and my earlier shorthand: resolve precedence, clamp for the model, then rewrite any remaining Ultra to Max. App-server dispatch precedes that conversion; its suffix overrides body effort, and the selected string reaches native turn/start.
• The stock sidecar selects the base image, while the explicit Codex installation is in runner-cli. Keep stock-profile HOLD. Source discrepancy is not an observed container failure. The text-flattening/new-thread bridge and health/account-unknown responses do not qualify runtime preservation or authenticated delivery.

The carrier/research bound records remain inactive proposals: selected modes/complete usage are null, with no new freeze, admission, comparative call or winner. The named research timeout sources cover different scopes; common deadline, search/fetch limits and external cleanup stay unqualified. Native keyless carrier delivery and original stream custody stay separate from pinned source availability.

Execution evidence needs a concrete sanitized locator map: the network 51-command receipt and selected SDK 14-test receipt expose digests/basenames without an independently recovered actual stream map in the examined source delta. Preserve their failed prechecks and owner-reported outcomes; provide the existing safe argv/exit/stdout/stderr/original-observation bindings in a private CODEX note. Fourteen selected mock tests add no stronger runtime gate to the existing 48-test mock scope and must not be counted twice. No credential/auth store or provider/SDK rollout backing capture was read.

The source-fetch timeout retains native_exit=null; bounded recovery requests returned0. The consequential reader's initial source command failed127 before successful bounded source recovery; all failures remain. Its one CLI snapshot is input262634, cached204672 subset, output11045, reasoning5236 subset, cache-write0. Requested routing does not prove served family or complete provider usage; both remain UNKNOWN. No host install/apply, native comparison trial, root merge or custody transfer occurred.

[seathatflowsinourveins]

SOURCE ACCEPT WITH RESIDUALS for #669 head 14c6e367df5855aa05005d09cf5df231007bd76e, carrying E0C's source verdict. Runtime, carrier delivery, custody, comparative freeze and role gates remain open.

Root independently rehashed all nine present changed old/new bodies, all 91 distinct named pinned upstream bodies (88 new, three reused; 1,673,339 bytes) and the additional own-public historical workflow draft at5199. All Git/bytes/SHA bindings match. The delta adds seven JSON records+registry, preserving all116 prior owned payload memberships, every9,657 prior full ordered registration row,10,242 foreign old Git tuples/order,9,541 foreign full rows/order and all non-files registry fields. Correction to the PR body's count label:9,657 is all prior rows;9,541 is the foreign count for the124-owned-path roster.

The new SDK48 reconciliation is explicitly historical scoped evidence. Fourteen selected mocked tests remain inside48, not62; the separate23-test failure remains FAIL/unexplained and the15s diagnostic remains dormant. This source read did not reopen execution originals or run a fresh SDK trial, provider or useful role.

The route/alias/token/reviewer/trading records remain inactive proposals with false freeze eligibility and no measured-default admission. Codex's default metadata matching differs from explicit-provider exact matching. Suffixless prospective inputs and separately requested native Ultra do not prove the actual catalog, mode, resolved inference effort, forced rules, installed gateway carry or authenticated delivery. Keep the earlier HTTP qualification: resolve precedence, clamp for the model, then rewrite remaining Ultra; app-server dispatch is earlier. Stock sidecar HOLD and inherited network outcomes are unchanged.

The reviewer oracle skips cases with no golden comments and records false negatives for empty candidate reviews on bug-containing cases. That supports the proposal's narrow scope, not a complete mixed reviewer-job oracle. The pinned Codex Action and AW source availability likewise do not qualify generated workflow/network/provider operation. Trading retains unbound common-case/oracle/native keyless route inputs; no strategy selection, data acquisition, registration, experiment or broker operation occurred.

The51-command network receipt and failed volume precheck still need their existing sanitized original stream/observation map: none is supplied by these seven inspected additions. SDK source reconciliation and its earlier mock evidence do not close wrapper EOF, outer/transient custody, R13 or complete19-case lifecycle gates. Actual identity and complete provider usage remain unknown.

Main 984bf4b273f1737a4cb65acb29c9f03e1d044d8a is newer than the raw branch; no prospective merge was attempted and raw foreign gaps do not establish eventual deletion. Prior source-fetch/reader failures remain retained. No new Astra read was needed for these inactive proposal additions; the one consequential E0C read remains closed0. No root repository edit, trial, freeze, host apply, merge, active config, credential/auth store or provider backing capture access occurred.

[seathatflowsinourveins]

SOURCE ACCEPT WITH RESIDUALS — exact head cfbb6bb8be46b29687c8e2101b1d54b01086487f, sole parent 14c6e367df5855aa05005d09cf5df231007bd76e. This carries my 14C source verdict across exactly one new published artifact and its registration.

The earlier limited “safe stream map not located” gap closes for the published network-policy map at this head: 216,404 bytes, SHA256 2aed4c0bdc0b5f1e9741b1cc591a1f9388d86a72ec5923997cac1f5af1578b84. Root independently verified its Git identity, registration and all 106 published UTF-8 stream projection byte/hash declarations. The retained controlled fixture has 51 commands, 61 returns and 10 nonterminal returns; its 11 final observations are a subset. The failed volume-first precheck remains with two commands/four streams and its nonzero outcome. No fixture or cleanup operation was rerun.

All 123 prior payloads and 9,664 prior ordered full registry rows remain exact. Foreign preservation is 10,242 full ordered Git tuples and 9,541 full ordered evidence rows; 193 receipts and 29 convergence records remain unchanged. Root also verified all 120 packet bindings, 25 closed native source captures, complete commit/tree/registry identities and the four newly cited upstream bodies. API roster has 125 paths at actual merge base cac8700ba914950266272347468bff7ad630a4bf; it is not a direct comparison against displayed base 463a57b983eec540ae90eb45c2b1a7c6fc469aed.

The new map transparently distinguishes sanitized projections from private originals. Matching projection hashes and declared original identities do not independently authenticate private originals or original completion equality; those remain publisher assertions, and the private backing was not opened. Its declared scope is a historical controlled fixture with zero model/provider calls, no original SWE task and no SDK/upstream-suite acceptance. Upstream network-policy source at 1e5c5c6, Docker inspection source at 7fc2dff, Go template source at resolved 2dc996f.

No SDK execution path or proposal methodology changes in this delta. SDK19, EOF/R13, outer/transient custody, useful-role qualification, pool delivery and runtime/default admission remain HOLD. The 91 prior pinned method sources retain their existing review. Raw differences against observed main b29036f05905f75df934603eb8f1d2248a01c543 are inherited; this does not certify the prospective merged tree. No credential, destination setup, owner configuration, freeze or runtime state changed.

[seathatflowsinourveins]

Closed with a record by the PR triage of 2026-10-07 (the command center's ruling, item review-ns2604-coop-20261007T023012Z (the command center's PR-triage ruling of 2026-10-07; proposal by github-ci-finalize, triage-20261007.json)). Not merged; the branch foundation/runtime-jobs-evidence-ready-20261003 stays on origin at cfbb6bb.

What it holds: blueprints/runtime-workers/comparisons/ (15 files); blueprints/runtime-workers/org-extensions.md, sdk-role-qualification.md; evidence/artifacts/runtime-jobs-source-intake-20261003/ (107 files)

Superseded by: Overtaken, together with #633, by 4c89741 (#704): the fix-wave rows (docs/decisions/2026-10-04-2604-e2e-fix-wave.md:23-24) set the 2604 runtime-worker defaults without running the comparisons whose protocols this PR registers, under the repository-quality rule's no-new-local-trial clause (docs/decisions/2026-10-04-repository-quality-rule.md:30). Same unit as #633 (97 of their 102 shared paths are byte-identical at the two heads). (confidence: medium)

Reopen trigger: Same as #633: a measured runtime-job comparison is reopened and needs these registered protocols and readiness evidence. Reopen with gh pr reopen 669.