Record the paired-run results (planning, orchestration) and pin the RTK receipt's locators

[seathatflowsinourveins]

Scope

• What this PR changes, in one or two sentences: records the results of two paired runs (planning persistence, run 2b, and orchestration, run 1) in the claude-native-practice skill: a compact artifact with the per-run rows and the sha256 of every private paper behind them, the five slot rows the results decide (planning-persistence, loops-completion, orchestration-frameworks, agent-teams, background-workflows) with their layer pages, the owed-work row of the advisor run (paired run 4, stopped on 2026-10-11 at the owner's priority, no verdict), the result of the routing run (paired run 5, decided 2026-10-11: a routing_5 section of the artifact, a test that recomputes the strict verdict from its rows, the record's paragraph), the record's addendum, the baseline manifest's label, and a test that recomputes each decision from the artifact. A second commit carries the command center's two P3s of its read of Cover the managed-config path rule's normalisation and Windows clauses under ntpath #973, and a third commit pins one of them with a test that fails on the old text (below). A fourth commit, 5c5eecaadbbaf0fd4ad2c39236fb820b3b47fbf4, answers the Codex review in GitHub at bf864c370 (four P2s) and records paired run 5, in one forward commit as the command center directed (below). A fifth commit, 425240f6e9342675f730e970d2078db5d59490e4, answers the command center's Claude first read and Codex's re-review of that head, both at 5c5eecaad (three P2s and five P3s; one P1 and four P2s), in one forward commit (below); the co-op's GPT first read covers the range to this head). A sixth commit, 49bb6494e152db393646a4782a2fdec07463bc72, answers the command center's Claude micro of that head (PASS_WITH_P3, seven P3s) and Codex's three new threads of the same head, in one forward commit leased on 425240f6e (below); the range for a read of it is 425240f6e..49bb6494e, and the co-op's GPT first read covers ba6ab3a5c..49bb6494e. A seventh commit, 1190b957ff3ce2fe6e0475662b78ab3972a5ca75, answers the command center's Claude micro of that head (one blocker, three P3s) in one forward commit leased on 49bb6494e (below); the range for the confirmation micro is 49bb6494e..1190b957f.
• Base commit: main b490361c5235dcdb22ecb3ab285486c335bd8ee0 (with Retire the skill's Claude-in-GitHub rows and point them to local review #1001, Converge 373 public stars and schedule weekly review intake #978 and Read the newest sweep manifest in the new-host grand list #1000 landed). The pull request targets main because the verdict-review-gate check fails closed for any other base. The branch was stacked on Retire the skill's Claude-in-GitHub rows and point them to local review #1001's first head ba6ab3a5c and is rebased onto main now that Retire the skill's Claude-in-GitHub rows and point them to local review #1001 landed (the command center's order of 2026-10-11: the retirement PR, this PR, then one re-pin that names the row sets of all of them): the seven commits of this PR were replayed in order, with the same subjects (218efbd9d -> 4135a801f, 88e06f947 -> 4bf5c2bf5, bf864c370 -> be553be19, 5c5eecaad -> 5b2b49aed, 425240f6e -> 37c26642a, 49bb6494e -> 037c062a6, 1190b957f -> 529fa6db0), and one mechanical commit on top, a00744f0e84455714a17aa7ec73fec22202dd1b9, which re-registers the files and recomputes the label of the slot manifest (relabel last). Everything cited below as 218efbd9d to 1190b957f is the pre-rebase history, the heads that the families read; the commit list of this PR is git log b490361c5..a00744f0e. The conflicts of the rebase were manifests/evidence.json and the slot manifest only, resolved by taking main's side before the finalization. The diff of this PR's own patch before and after the rebase (git diff ba6ab3a5c 1190b957f against git diff b490361c5 a00744f0e, line multisets per file) differs in exactly four places: scheduling-supervision.md and two lines of slots.json (the scheduled-tasks route and alternative) are no longer in this PR's diff because Retire the skill's Claude-in-GitHub rows and point them to local review #1001 landed the identical text, the label line of the slot manifest, and eight versus six rows of manifests/evidence.json. The skills manifest's pin of the claude-native-practice row is not touched here.
• Lane: lane:foundation
• Owned paths touched: .claude/skills/claude-native-practice/reference/slots.json, .claude/skills/claude-native-practice/reference/workers.md, .claude/skills/claude-native-practice/reference/observation-inference.md, docs/decisions/2026-10-09-claude-code-native-practice.md, evidence/artifacts/claude-native-practice-20261009/claude-native-slot-manifest.json (its label only), evidence/artifacts/native-practice-paired-runs-20261010/paired-runs.json (new), tests/test_paired_runs_results.py (new), manifests/evidence.json (re-registered rows); in the second commit evidence/receipts/rtk-051-qualification-20261004.json (three locators pinned), tests/test_rtk_receipt_locators.py (new) and tests/test_skill_usage.py (one comment, one test); in the fourth commit .claude/skills/claude-native-practice/reference/slots.md (one index cell), the routing_5 section of paired-runs.json, evidence/artifacts/claude-native-practice-20261009/meter-windows-20261011.json (new), tests/test_native_practice_meter_windows.py (new) and changes to tests/test_paired_runs_results.py, tests/test_rtk_receipt_locators.py and the observation, workers and slots layer pages; in the fifth commit slots.json, slots.md and workers.md (the agent-teams Overturn when and the task-list alternative), the manifest label and its list, the routing_5 section of paired-runs.json, the decision record's routing paragraph and heading, and tests/test_paired_runs_results.py, tests/test_rtk_receipt_locators.py and tests/test_native_practice_meter_windows.py (one tolerance); in the sixth commit slots.json and workers.md (the time bounds of paired run 1), scheduling-supervision.md (the same scheduled-tasks wording as Retire the skill's Claude-in-GitHub rows and point them to local review #1001's third forward commit), the decision record's paragraphs, the manifest label, the routing_5 section of paired-runs.json, tests/test_paired_runs_results.py and tests/test_rtk_receipt_locators.py; in the seventh commit slots.json and workers.md (the origin of the 45-minute background wait), the decision record, the design.caps and papers of the orchestration section of paired-runs.json, the manifest label, tests/test_paired_runs_results.py and a comment of tests/test_rtk_receipt_locators.py.

What the results decide (every figure is recomputed from the artifact by the new test):

• Planning persistence (run 2b, 12 runs). Repository state files stay the default. The native task list is ranked second (it won 2 of 4 cells, both on F6, and lost both F5 cells); the Ralph-style loop is ranked third and for fallback only (it won 0 of 4), with the three statements the command center ruled: an external done check is required, the classification is post hoc and provisional, and a rerun preregisters its criterion first. Lost work: 0 events in all 12 runs, counted by two readers on different models.
• Orchestration (run 1, 18 runs). The single session stays the routing for bounded fixes on north-star code. All 18 runs passed the hidden tests; the blind judges' counted preferences over the single session were 1 of 6 fixtures for the Workflow and 1 of the 5 comparable for the team (the single session was preferred on 5 and 3); the Workflow cost 1.59 times and the team 3.02 times the single session's mean cost (1.50 and 2.55 over the primary fixtures F3 to F6). Neither clears the preregistered rule in the primary or the sensitivity analysis. One defect was found at scoring and is recorded rather than hidden: arm C's diff on F3 was not captured, so F3's two pairs with the team are void (they cannot change an outcome).
• Model routing for review micros (run 5, 20 runs, decided 2026-10-11). By the registered rule Opus stays the default: Sonnet 5.5 at medium missed 1 decisive finding that Opus 5.5 found (G10: 1 of Opus's 2 runs, 0 of Sonnet's) while costing 17% of Opus's (bound 60%); detections of 24: Opus 7, Sonnet 6; the strict rule's noise floor is 0.02 at one shared rate (0.73 with each finding's own pooled rate, a plug-in calculation) and the tolerant reading, reported and not decisive, holds (a chance of 0.56 for identical reviewers; 0.62 was the per-finding clause alone). The second trial was added by the command center after trial 1 had met the rule, so the two-trial verdict is a ruled extension and not a preregistered outcome. Beside the rule: Opus reported 7 real and new P1 or P2 findings against Sonnet's 1, and Sonnet approved Add the shared evidence-bound gap closeout loader #982 in both trials where Opus requested changes. The command center ruled the strict reading (ruling sha256 a6beedb039e6); no slot row, default or skill text changes.

The second commit (88e06f947af46d2f6d3ea711be18fa5cd86c0c54) and the third (bf864c3705dbca1fc0b5c24ba388e5836e4c35d0), the command center's two P3s of its read of #973 (PASS_WITH_P3; its ruling is that they go in my next skill PR):

• The RTK 0.51.0 receipt's three locators are pinned. The receipt cited three places that require rtk 0.50.0 as path:line with no commit. Cover the managed-config path rule's normalisation and Windows clauses under ntpath #973 added two imports to tests/test_skill_usage.py, so the unqualified tests/test_skill_usage.py:33 now names PARSER_DIR and not the version regex it named when the receipt was written. Each locator is pinned to 14048b840425c2569e0df60a6596e94e601da15b, the commit that added the receipt, where it resolves. The regression test tests/test_rtk_receipt_locators.py (3 tests) reads each line at the pin through Git and requires the prerequisite's regex on it: on the tree without this commit (218efbd9dec8be5ff320e50f62b9ab64e0396cbb plus the new module) it fails (failures=4, in 2 of its 3 tests), and on this head all 3 pass (outputs under the local commands). The third test checks that the check can fail.
• The ntpath.abspath comment is corrected. The comment in tests/test_skill_usage.py said ntpath.abspath adds a drive to a POSIX string. At CPython v3.12.4 Lib/ntpath.py:564-578 (_abspath_fallback) it joins the working directory only when isabs is false and then calls normpath, so the separators become backslashes and no drive is added. The comment now says so and cites those lines, and test_ntpath_abspath_adds_no_drive_to_a_posix_string measures it on a POSIX host: it passes on CPython 3.12.4 and 3.13.16. A measurement of the standard library cannot fail before a comment fix, so the repository's correction rule (a regression test that fails before the fix) is met by the third commit: test_the_comment_of_the_windows_path_test_states_what_abspath_does reads the source of the test the comment belongs to and requires that the wrong claim is gone and that the corrected claim and its citation are present. On the module with the old comment restored (the new tests kept) it fails (failures=1: the pattern matches the old wording), and on this head all 8 tests of the class pass (outputs under the local commands).

The fourth commit (5c5eecaadbbaf0fd4ad2c39236fb820b3b47fbf4) answers the Codex review in GitHub at bf864c370 (four P2s, each reproduced before the change; the threads are replied to and resolved after the push) and records paired run 5, in the one forward commit the command center asked for. It does not change any other slot row, default or skill text.

• workers.md:44, the Workflows row's satisfied Overturn when. Right: paired run 1, arm B measured two failure modes under --setting-sources local, the row's Overturn when names "a measured failure mode on our runs", and its route still sent a fenced headless run to the name lookup that failed in 6 of 6 runs. The row stays a scoped default (its status already says so); its route now says to launch a saved Workflow by scriptPath in that context and to expect the 45-minute wait ceiling to cut a long verifier (it returned in 2 of 6 runs and was cut in 3), and Overturn when says which measured failures the route carries. Mechanism at the pinned client and its documentation: --setting-sources loads only the sources it lists (cli-reference.md line 129, capture sha256 54587c07d51da9f2…) and saved Workflows live in the project's .claude/workflows/ (workflows.md line 204, capture sha256 e55138abf9b78682…); the failure itself was measured on 2.1.296. One step wider: the slots.md index cell of the row had kept the old Overturn when. Both surfaces now agree, and Retire the skill's Claude-in-GitHub rows and point them to local review #1001's forward commit adds a check that compares every row's Overturn when across slots.json, the index and the layer pages (it fails on the old index cell; this branch gets the check when Retire the skill's Claude-in-GitHub rows and point them to local review #1001 lands).
• workers.md:29, the team's recorded win. Right: the note said the team "won no fixture", while the artifact counts one counted win against the single session (F2, in the analysis over all six fixtures), one tie (F4) and three losses, and none of F3 to F6. The note now says so, and the orchestration test derives the counts from the artifact.
• tests/test_rtk_receipt_locators.py:67, a pin whose tree is missing. Right: in a checkout that keeps the pinned commit object but not its tree (a treeless partial clone, offline), has_commit() was true, git show failed and all three locators failed. git cat-file -e <sha>^{tree} (git 2.53.0 manual: "Exit with zero status if exists and is a valid object") returns 0 with the tree and 128 without it while ^{commit} stays 0, measured in a throw-away repository. A pin without its tree now skips; a bad path with the tree present still fails; the new test builds both states.
• observation-inference.md:13, the metering ratios. Right, and worse than asked: metered again over two fixed 24-hour windows the Opus ratio is 2.56 and 1.64, not 1.9 (1.92 was the figure at the instant the advisor run's preregistration was written), and the clause that the events leave the advisor out contradicted the advisor run's Stage 0 (private paper, sha256 2377eb9ce915…: the events book the advisor's dollars under the executor's model). The row's sentence is now written from a committed aggregate artifact (meter-windows-20261011.json: per model the counter's dollars, its token counters and the events' summed cost for the two windows, the pricing page's hash and the Fable 5.1 rates), tests/test_native_practice_meter_windows.py recomputes it, and the clause about paired run 4 says it was stopped.
• Paired run 5 recorded. routing_5 in paired-runs.json: the 20 run rows, the 12 decisive findings with the runs that found each, costs, the judge and refute record hashes, the decision as recorded, the ruling by path and sha256, and the hash of every paper. RoutingRunTests recomputes the strict verdict from those rows (Opus stays), shows the rule is live (Sonnet credited with the missed finding gives Sonnet takes micros; its cost times four gives Opus stays), and checks the record's paragraph against the artifact. The decision record's owed table no longer lists run 5, and run 4's row says it was stopped. 19 of 19 single-edit mutations are caught (table below).

Finding fixed alone, then undone	Test that fails on the old text or code
Workflows route and Overturn when	test_the_orchestration_rows_state_what_the_artifact_shows (failures=1)
the team's note	the same test (failures=1)
the metering row	test_the_row_states_the_artifacts_figures_in_the_json_and_on_the_page, test_a_changed_figure_in_the_row_is_caught (failures=2)
the partial-clone skip	test_a_pin_whose_tree_is_missing_is_skipped_and_a_bad_path_with_the_tree_present_is_not (failures=1, the old logic in a temporary copy)
the slots.md index cell	test_the_index_and_the_layer_pages_show_each_slots_overturn_condition (failures=1, from #1001's forward commit stacked on this one)

The fifth commit (425240f6e9342675f730e970d2078db5d59490e4) answers the command center's Claude first read of #1005 at 5c5eecaad and Codex's re-review of the same head, in one forward commit leased on 5c5eecaad (the command center directed one commit; the co-op's GPT first read covers ba6ab3a5c..425240f6e; Codex's five threads are replied to and resolved after the push). Each finding was reproduced before the change.

• Claude P2-1, the cited ruling no longer equalled its file. Right, and caused by the command center's own edit (its correction Adoption hygiene: native Claude Code pin is a floor (2.1.281), ai-memory assistant capture opt-in, ignore Claude Code local state #159): the ruling was reworded at about 09:32Z, after this branch cited its hash at 09:21Z. routing_5.ruling and papers.ruling carry the current sha256 (a6beedb039e6…) and size (2,581 B); routing_5.ruling.revised says the file was edited after it was first cited and that the text first cited (prefix 08aac8fd72be) is not retained; the record's paragraph is generated from the artifact and its prefix is now a6beedb039e6. Nothing in the repository could notice a private file changing, so test_the_private_papers_resolve_to_their_recorded_hashes_where_they_are_available (enabled by PAIRED_PAPER_ROOTS) resolves all 24 path-and-hash entries of the artifact against the files on a host that has them: 24 resolve at this head, and with the old hash in the artifact the test fails on both entries of the ruling (/routing_5/ruling and /routing_5/papers/ruling); mutant R20 (the old hash back) is also caught by the record test. The command center has since corrected the "0.62" of its ruling in a sidecar file and left the ruling's bytes unchanged, so this pin stays valid.
• Claude P2-2 and P3-2, the ruling's condition. The ruling lets Sonnet at medium take non-gate foundation changes (docs, pages, catalogs, generated reports) only where the change is mechanical or a second Claude-family read exists; the artifact and the paragraph had stopped at the parenthesis. Both carry the condition, and the paragraph says the allowance is the command center's routing discretion under the owner's direction to use the cheapest model that holds quality and not something run 5 measured (none of the five heads was a docs, pages or generated-report change). The paragraph is generated from the artifact by the test template and the record carries the generated text.
• Claude P2-3, the manifest label (and the same defect class as Retire the skill's Claude-in-GitHub rows and point them to local review #1001's P2-1). The label named the slots.json of 218efbd9..bf864c37; at 5c5eecaad that file is different, so the exact-difference check skipped, and three differing pairs were unlisted: route and overturn_when of background-workflows and evidence of usage-cost-monitoring. The label is the sha256 of the shipped file (2e8feccb…), set last, and the list is recomputed with the same field comparison the test makes (now also overturn_when of agent-teams, edited below); the check runs and passes. The strict equality test arrives with Retire the skill's Claude-in-GitHub rows and point them to local review #1001's second commit, and the rebase onto it recomputes the label once, as the command center ruled.
• Claude P3-1 and Codex P2 (tests/test_paired_runs_results.py:82), the tolerant floor. The tolerant reading needs every S_f >= O_f - 1 and Sonnet's total within two of Opus's; noise_floor(..., False) modelled the first clause only and the artifact and the record printed 0.62 where the chance for two identical reviewers satisfying the whole reading is 0.5551 (an exact convolution over the 12 findings; on three findings it equals an enumeration of all outcomes). Both figures are kept and labelled in the artifact (tolerant_pass_chance_if_identical is the full reading, tolerant_per_finding_clause_pass_chance_if_identical the figure the result record prints), the correction is a new private note (paired-routing/CORRECTION-tolerant-floor-20261011.md; the frozen result record is not edited), and the paragraph says "0.56 of the time (0.62 for its per-finding clause alone, the figure the result record prints)". The strict rule's chance with each finding's own pooled rate (0.7305, a plug-in calculation from these runs and not a significance test, as the result record says) is added beside the 0.02: a miss is less easily put down to chance than the first figure suggests, though it rests on one Opus detection. The verdict (Opus stays) rests on neither figure.
• Codex P1 (paired-runs.json:1532), the second trial. Right in substance, and the artifact already said part of it (design.second_trial: "run as an extension the command center decided after trial 1") while the record's paragraph did not. The order, from addendum 3 of the run's papers: trial 1's registered verdict was Sonnet takes micros; the trigger of addendum 1 (written before any run) did not call for a second trial then; the command center ruled trial 2 as an extension after seeing that result, with the rule for two trials unchanged. The paragraph and decision.two_trial_verdict_basis now say that the two-trial verdict is a ruled extension and not a preregistered outcome, and that an extension decided after a pass can only move the verdict toward Opus, the default; the routing consequence is the command center's ruling (strict reading stands, Opus for gate-bearing reads) and no slot row, default or skill text changes. The command center confirmed the wording and asked for the path and hash of addendum 3 beside the basis field; papers.addendum_03 holds them now and the next commit cites them in decision.
• Codex P2 (workers.md:29), the team's Overturn when. Right: the row's condition ("teams ... cost more for the same work") is met by the measured bounded fixes (the team cost 3.0x the single session for the same results), while the default is scoped to work where workers must message each other, which paired run 1 did not measure. The condition now names that class and says paired run 1 measured bounded fixes, which need no messaging and fall outside it (the Overturn when of slots.json, the slots.md index and the layer page agree), and test_the_orchestration_rows_state_what_the_artifact_shows pins the wording and the 3.0x it rests on.
• Codex P2 (workers.md:76), the task list's completion. Right: both measured F5 task-list runs used all 8 sessions without declaring TASK-DONE, and the ranked fallback gave no completion condition while the Ralph alternative requires one. The alternative states that it has no completion condition of its own and that a headless driver needs an external done check (the Stop-hook completion gate of loops-completion, a test or an explicit condition) or keeps launching sessions to its cap; the planning test reads the two F5 runs from the artifact.
• Codex P2 and Claude P3-3 (tests/test_rtk_receipt_locators.py:60), the blob. Right: in a blobless partial clone the commit and its root tree are present and the blob is not, so has_tree() was true and the locator test failed instead of skipping (reproduced by deleting only the referenced blob of a throw-away repository). unavailable(sha, repo, path) lists the path with git ls-tree (a subtree that is missing, or a blob that git cat-file -e cannot find, is a reason to skip; a path that is not in the tree at the pin is a wrong locator and fails), locator_unavailable() applies it to the receipt's locator form, and a test asserts the wiring: with a checkout that cannot read any locator the locator test skips all three. 7 of 7 single-edit mutants are caught.
• Claude P3-4 and P3-5. The meter test's tolerance is 0.01 (the differences are 0.004 and 0.002), and the addendum's heading is "Addendum (2026-10-10 and 2026-10-11): paired-run results" with the tests' anchor updated. P3-6, the slots.md cell without a fail-before, is covered by the check that Retire the skill's Claude-in-GitHub rows and point them to local review #1001's first commit adds (it fails on the old index cell), as the read said.

The sixth commit (49bb6494e152db393646a4782a2fdec07463bc72) answers the command center's Claude micro of #1005 at 425240f6e (PASS_WITH_P3: P3-1 to P3-7; file research-cc-role-20261008/pr1005-claude-micro-425240f6.md, 7,888 B, sha256 3d018f236a023f99433ee6c0be50a730cba58ecbb0db584b488e5fc618fc4f61) and Codex's three new threads of the same head (10:46:08Z: workers.md:61, tests/test_paired_runs_results.py:91, scheduling-supervision.md:9), in one forward commit leased on 425240f6e. Each finding was reproduced before the change; five tests of the commit fail on the data of 425240f6e (retained fail-before-micro3.txt, sha256 614e0d753f81aeddd11350fd06d68e89d5204f3ef57150708d1d754014e4340b).

• P3-1, the order of the second trial. decision.two_trial_verdict_basis said the decision to run two trials was not fixed before any run; addendum 1 (line 13, written before any run, sha256 1917091f…) fixed it: after a passing trial 1 the run ends (Sonnet takes review micros, no second trial). The command center ran trial 2 after seeing that result (addendum 3, pinned beside the basis with its path, size and hash), so the verdict stays a ruled extension and not a preregistered outcome, and the basis, the record's paragraph and ruling.routing ("read or micro", as the ruling says) now say so.
• Codex P1 (tests/test_paired_runs_results.py:91), the basis of the tolerant floor. The reading and the null model are the run's own, registered before any run (addendum 1, lines 17-18: the reading with the clause on the totals, and the noise floors of two identical reviewers who each find a finding with probability p in each run); no third-party statistical source defines them, so the reference is the brute-force enumeration that already fails before the correction. The convolution of the per-finding differences is the one numpy 2.5.3 documents for numpy.convolve ("In probability theory, the sum of two independent random variables is distributed according to the convolution of their individual distributions"), and test_the_tolerant_floor_matches_numpy_convolve_where_numpy_is_installed cross-checks tolerant_floor against it for four (p, n) pairs; the artifact records both bases (decision.tolerant_floor_basis) and the gated papers test reads lines 17-18 of the retained addendum.
• Codex P2 (workers.md:61) and P3-3, the time bounds of paired run 1. The orchestration-frameworks row and the record said "a 45-minute ceiling per run". The registered design has two bounds, a 50-minute process cap for a headless session and a 45-minute cap for the team (design.caps; preregistration line 156, sha256 531758c4…), and the Workflow's verifier was cut by the background-wait ceiling of claude -p (the retained stderr of the runs F1-B, F3-B, F4-B and F5-B reads "Background tasks still running 45m after the last turn", after 2,779 to 2,794 seconds, inside the 50-minute cap); the sixth commit called that 45 minutes the client's own, which was wrong: the runner set it (the seventh commit, below). The row (slots.json, workers.md) and the record say so, and test_the_time_bounds_the_row_and_the_record_state_are_the_registered_ones computes the counts from the run rows (4 of 6 Workflow runs, 4 of 6 team runs of which 3 at the team cap).
• Codex P2 (scheduling-supervision.md:9). That text is Retire the skill's Claude-in-GitHub rows and point them to local review #1001's (its commit ba6ab3a5c). This commit carries the same wording as Retire the skill's Claude-in-GitHub rows and point them to local review #1001's third forward commit (GitHub Actions schedules repository jobs; Claude runs inside GitHub are retired on 2026-10-10, so a scheduled Claude task runs locally; docs/decisions/2026-10-10-retire-in-github-claude-review.md:5-6, :43), as the command center ruled; the test of the wording is that commit's.
• P3-2, the label. Reverting the label to the file of 5c5eecaad still gives rc 0 here (skipped=1) because the guard on this branch skips on a mismatch; Retire the skill's Claude-in-GitHub rows and point them to local review #1001's strict test fails it after the rebase (verified by simulation), so no duplicate test is added and the body says so. The label is recomputed for the sixth commit's slots.json.
• P3-6 and P3-7. The test count is 30 (the body said 25); the RTK locator test cites git 2.53.0 git-ls-tree(1), OUTPUT FORMAT, for the entry it splits.

The seventh commit (1190b957ff3ce2fe6e0475662b78ab3972a5ca75) answers the command center's Claude micro of #1005 at 49bb6494e (CHANGES_REQUESTED: one blocker, three P3s; file research-cc-role-20261008/pr1005-claude-micro-49bb6494.md, 1,751 B, sha256 87f594190f548d7f3f1755248b5b4627e37e046abd500711fef2ec3f49c69e64), in one forward commit leased on 49bb6494e. The blocker was reproduced before the change; four tests of the commit fail on the data of 49bb6494e (retained fail-before-wave.txt, sha256 216b1eb585f609b01f078a99dec1cc718af8c6857e4ed02621cbde7f2c6ec29a).

• P2-1, new in the sixth commit: "claude -p's own 45-minute wait" is false. The runner set CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS=2700000 for every headless run (paired/orch_run.py line 137, retained and hashed as papers.runner, sha256 ad8ffbef0f48c105f083c27b98bf247ad0737938c20bced4cd914f9f9d53a890); the client's default is 10 minutes (the retained env-vars.md line 357: "Default: 600000, or 10 minutes", sha256 f9400917e85a9c4b…; headless.md line 77: "after 10 minutes of continuous idle waiting by default", sha256 25885d2e178a9746…; client 2.1.296). The decision is unaffected: the four Workflow runs that ended at the ceiling ran background work for 2,702 to 2,703 seconds, the ceiling being 2,700. One rewording pass over the orchestration-frameworks and background-workflows rows (slots.json), their pages (workers.md) and the record; the setting in design.caps (45 minutes, set by the runner; the client's default 10) with its three sources as papers (the gated test checks their hashes and their cited lines); the test pins of the old wording inverted. test_the_background_wait_ceiling_is_the_runners_setting_and_the_clients_default_is_ten_minutes computes the 2,700 seconds from the run rows. 13 single-edit mutants are caught and the control passes (table sha256 4ede6e1defd4b26c3327b8cecd7f5cc898107768fd8b2128e6717fc4b67b1ccc).
• P3-1, Codex 12 (the advisor attribution). It does not hold as a P1 (the thread is resolved); the follow-up is to cite the vendor's monitoring documentation (the micro names monitoring-usage.md lines 807-818 and advisor.md line 15).
• P3-2, this description: the counts and the sentences that the seventh commit corrects are refreshed. The label's fail-before is the strict test of Retire the skill's Claude-in-GitHub rows and point them to local review #1001, because the guard on this branch skips on a mismatch (simulated by the micro on the rebase: the old label fails there).
• P3-3: the git-ls-tree quote of the RTK locator test is exact (git 2.53.0 git-ls-tree(1), OUTPUT FORMAT and -z).
• The command center's confirmation micro of 1190b957f: PASS_WITH_P3, no correction owed (research-cc-role-20261008/pr1005-claude-micro-1190b957.md, 10,338 B, sha256 9695490fde6bfc4c…). P2-1 is fixed at all seven places against the installed client 2.1.296 (the binary's default is 600000 ms, the runner's line 137 sets 2700000, the retained stderr reads "45m"); the range holds no P1 and no P2; its three P3s are in the Follow-ups below.

SOTA sources

• Claude Code 2.1.296 (the vendor's pinned binary) is the system under test: every run is a claude -p session (or, for the team arm, an interactive session in tmux) on that client, with --setting-sources local, an explicit permission mode and the vendor's own sandbox for Bash; the mechanisms the rows state (a saved Workflow launched by scriptPath, the wait ceiling of claude -p, the task-list tools named in --allowedTools, agent teams) were observed on that binary, not taken from a post.
• Evaluation design followed the vendor's published guidance, read 2026-10-09 by the reading run recorded in evidence/artifacts/claude-native-practice-20261009/primary-source-reading.json: Demystifying evals for AI agents (fetched 2026-10-09T13:48:03Z, sha256 b67dd8f8b1bb18b528530eca251cb5407c9b42a1222a3f47968a2407e0b2f39d); How we built our multi-agent research system (fetched 2026-10-09T13:48:00Z, sha256 149dd1e4ddd111fd8cbb1a09bbe8ef64654447e9e39aa5707be06d9b61f1edc0); Harness design for long-running application development (fetched 2026-10-09T13:48:01Z, sha256 fe37bba54ab5f1470a176c43b124968aeac50000037aa913bb215e8cfa3717ca); Effective harnesses for long-running agents (fetched 2026-10-09T13:48:01Z, sha256 919145986a04e774969a3e628c40dbe4ef33de67881b778efb7e11a0a2484425).
• The owed-work rows for the advisor run cite the vendor's advisor documentation and environment-variable reference (CLAUDE_CODE_DISABLE_ADVISOR_TOOL, read 2026-10-11, capture hashes in the preregistration) and the pricing page (sha256 67904e280b595d31f373c857844ed380072ae79d280d8f385640b5992d337ca9); the routing run's preregistration cites the vendor's model-choice guidance (the Sonnet 5.5 launch post and the "what a task costs on Opus 5.5" post).
• CPython v3.12.4, Lib/ntpath.py lines 564-578 (_abspath_fallback; file sha256 ffa7f85c0382a6e918dc3499f0d51a58769da24a56704f16bda27f63ff1680d8, fetched 2026-10-11 from the python/cpython tag through the GitHub contents API and read here), for the corrected comment. The RTK locators are pinned to this repository's own commit 14048b840425c2569e0df60a6596e94e601da15b (the commit that added the receipt), read through git show.
• The decisions follow the repository's own rule that a default changes only on a paired result: docs/decisions/2026-10-09-claude-code-native-practice.md, "Owed work" and the new addendum.
• The Codex review in GitHub of Record the paired-run results (planning, orchestration) and pin the RTK receipt's locators #1005 at bf864c370 (four P2s): each finding is reproduced here from the rows, the artifact and the test module before the change, not copied.
• Claude Code documentation, retained captures with the hashes the page records: --setting-sources (cli-reference.md line 129, sha256 54587c07d51da9f2…, fetched 2026-10-09T03:38:34Z) and the project's .claude/workflows/ (workflows.md line 204, sha256 e55138abf9b78682…, fetched 2026-10-09T03:39:40Z); git 2.53.0, git-cat-file(1), option -e, read from the installed manual.
• The advisor run's Stage 0 and stop records (private papers, sha256 2377eb9ce915… and 6fdeb8916960…), the pricing page (sha256 67904e280b595d31f373c857844ed380072ae79d280d8f385640b5992d337ca9, unchanged, retained) and the host's local Prometheus and Loki at two fixed instants (the queries are in meter-windows-20261011.json).
• The command center's Claude first read of Record the paired-run results (planning, orchestration) and pin the RTK receipt's locators #1005 at 5c5eecaad (CHANGES_REQUESTED: three P2s, five P3s; file research-cc-role-20261008/pr1005-claude-read-5c5eecaa.md, sha256 309f83bac9700c4f47e002ea22da1cd60433f1130334e517f764071ed9ac0c90) and Codex's re-review in GitHub of the same head (09:45:07Z: one P1 and four P2s): each finding reproduced here from the rows, the artifact and the test modules before the change, not copied. The tolerant floor was recomputed as an exact convolution and checked against enumeration (tolerant_floor.py of the research records; the correction note is paired-routing/CORRECTION-tolerant-floor-20261011.md, 2,352 B, sha256 f80215104fe6820567acf8c8857e2fffd0b2c5dfb3db759a2b7816fbffb86119), and the order of the second trial is read from addendum 3 of the run's papers (paired-routing/addendum-03-trial-2-extension-20261011.md, sha256 a4f8d1d2f4e9…: "this trial is an extension the command center decided after seeing that result").
• Paired run 5's papers by sha256 in routing_5.papers: the preregistration (66f87a9eb2b9…), three addenda, the ground truth, the two-trial result record (5642e6a1b989…) and trial 1's (f7f8daad1a11…, kept unchanged), the analysis, the exposure report and the command center's ruling (research-cc-role-20261008/ruling-routing-run5-two-trial-20261011.md, a6beedb039e6…).
• The command center's Claude micro of Record the paired-run results (planning, orchestration) and pin the RTK receipt's locators #1005 at 425240f6e (PASS_WITH_P3, seven P3s; file sha256 3d018f236a023f99433ee6c0be50a730cba58ecbb0db584b488e5fc618fc4f61) and Codex's three new threads of the same head (10:46:08Z): each finding reproduced here from the rows, the artifact and the test module before the change, not copied.
• The registered text the tolerant floor rests on: paired-routing/addendum-01-ruling-and-extension-20261011.md (3,861 B, sha256 1917091fe507696b…), line 13 (the trigger of the second trial) and lines 17-18 (the tolerant reading and the noise floors), pinned as papers.addendum_01; NumPy 2.5.3 (the version the repository pins in .github/workflows/native-foundation-e2e.yml), the numpy.convolve documentation read from the installed package: "the sum of two independent random variables is distributed according to the convolution of their individual distributions".
• git 2.53.0, git-ls-tree(1), OUTPUT FORMAT (read with git help -m ls-tree on this host): the default entry is <mode> SP <type> SP <object> TAB <file>, and -z ends each entry with a NUL; preregistration.md line 156 (sha256 531758c4…) for the 50-minute and 45-minute caps of paired run 1.
• The command center's Claude micro of Record the paired-run results (planning, orchestration) and pin the RTK receipt's locators #1005 at 49bb6494e (CHANGES_REQUESTED: one blocker, three P3s; file sha256 87f594190f548d7f3f1755248b5b4627e37e046abd500711fef2ec3f49c69e64): the finding, reproduced here from the runner, the artifact and the retained documentation before the change, not copied.
• The vendor's documentation for the default, retained with the hashes the artifact records as papers (client 2.1.296): env-vars.md line 357, CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS ("Ceiling in milliseconds on idle waiting for background work ... Default: 600000, or 10 minutes ... Requires Claude Code v2.1.182 or later") and headless.md line 77 ("Claude Code stops whatever is still running after 10 minutes of continuous idle waiting by default"); and the runner that set it, paired/orch_run.py line 137 (env["CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS"] = "2700000"), retained as papers.runner.

Evidence-class table

Claim	Evidence class	Command / receipt
All 12 planning runs and all 18 orchestration runs passed the hidden tests; the decisions follow from the preregistered rules	native_proven (run on this host, client 2.1.296, receipts retained privately)	RESULT.md sha256 293fe657916adb036514f2ee4ac2f1232baf0e6fe71d53ba521a956a76f88b3d; orchestration RESULT.md sha256 e87b9cabd8c2b856c7ab0a3c3ed8c2cb9007c55bdc3f5c31d07990dd0b2cd499
Each decision in the rows and the record is recomputed from the artifact, and a row that drifts from the data fails	local_integration	tests.test_paired_runs_results (31 tests)
The tests are not decorative: 19 single-edit mutations of the orchestration artifact, rows and record and 31 of the routing data and text are each caught (a dropped run, a flipped pass, a preference on a void pair, a hash that is not a sha256, a changed cost ratio, a changed count, a wrong cited hash, a dropped record statement)	local_integration	the mutation table below
Lost work is 0 events in all 12 planning runs, by two readers on different models	native_proven	lost_work.json sha256 e871e89e37ca1811f1416701cb826cbfd334571fee63437aa0f78dd7c9cb95d5
F3's two pairs with the team are void because its diff was not captured; the other 17 diffs cover the same non-test files as their retained worktrees	source_review	addendum-f3c-diff-capture-20261011.md sha256 19fb4feaef3a49ee90c7f11986b6ad509d27720f15d8e64b63dea192609dbb95
The baseline manifest's label lists exactly the exported fields that now differ from its rows	local_integration	tests.test_native_practice_snapshot_provenance (8 tests, none skipped at this head; at 5c5eecaad the label named an older file and the exact-difference test skipped; the strict equality test of #1001 is on main since the rebase and runs here: the label equals the sha256 of slots.json, 76d5eff8...)
The receipt's three locators resolve at the commit that added it, each on the line that requires rtk 0.50.0	local_integration	tests.test_rtk_receipt_locators (4 tests since the fourth commit); without the second commit it fails (failures=4), on this head it passes
The corrected comment states what the pinned CPython source does, and the old wording cannot come back	source_review and local_integration	CPython v3.12.4 Lib/ntpath.py:564-578; test_ntpath_abspath_adds_no_drive_to_a_posix_string passes on 3.12.4 and 3.13.16; test_the_comment_of_the_windows_path_test_states_what_abspath_does fails on the module with the old comment (failures=1) and passes on this head
Paired run 5 is decided by its registered rule, and the rows, the record's paragraph and the owed table say what the artifact shows	native_proven (20 runs and 10 judge records on this host, client 2.1.296) and local_integration	tests.test_paired_runs_results.RoutingRunTests (10 tests); result record sha256 5642e6a1b9895710679e30ae43af86801bee4742d42d87e6313aa79c7f7bfce9; the ruling sha256 a6beedb039e6c9690c15d81d31b9084b354ae46ecbf8fb2a36416a88211f2863 (2,581 B)
The two-trial verdict basis says addendum 1 fixed the trigger (a passing trial 1 ends the run) and that the second trial was a ruled extension after seeing the result	local_integration	test_the_ruling_keeps_its_condition_and_the_two_trial_verdict_says_what_it_rests_on; it fails on the data of 425240f6e (retained fail-before-micro3.txt, sha256 614e0d75…)
The tolerant floor's reading and null model are the run's registered text, and its convolution is the one numpy.convolve documents	local_integration; source_review (numpy 2.5.3 documentation, addendum 1 lines 17-18)	test_the_tolerant_floor_matches_numpy_convolve_where_numpy_is_installed (four (p, n) pairs, 12 places), the brute-force enumeration and the gated papers test
The time bounds of paired run 1 in the row and the record are the registered ones, and the Workflow's verifier was cut by the background-wait ceiling that the runner set to 45 minutes	local_integration	test_the_time_bounds_the_row_and_the_record_state_are_the_registered_ones (4 of 6 Workflow runs, 4 of 6 team runs, 3 at the team cap, computed from the run rows); it fails on the data of 425240f6e
The 45-minute background wait of paired run 1 is the runner's setting (CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS=2700000), the client's default is 10 minutes, and the four Workflow runs that ended at the ceiling ran background work for 2,702 to 2,703 seconds	local_integration; source_review (the retained runner and documentation lines)	test_the_background_wait_ceiling_is_the_runners_setting_and_the_clients_default_is_ten_minutes; it and three more tests fail on the data of 49bb6494e (retained fail-before-wave.txt, sha256 216b1eb5...); 13 of 13 mutants caught (4ede6e1d...)
The tolerant reading's noise floor counts the clause on the totals (0.5551, not the per-finding clause's 0.6204), and the strict rule's chance with each finding's own pooled rate is 0.7305	local_integration	test_the_tolerant_floor_counts_both_clauses_of_the_reading (an enumeration over three findings), test_the_recorded_decision_is_what_the_rows_give; correction note sha256 f80215104fe6…
Every private paper the artifact cites resolves to its recorded sha256 and size on a host that has it	local_integration (gated by PAIRED_PAPER_ROOTS)	test_the_private_papers_resolve_to_their_recorded_hashes_where_they_are_available: 29 entries resolve (the two added by the sixth commit are addendum 3 and the ruling's sidecar correction, the three added by the seventh are the runner and the two documentation pages, each checked on its cited line); with the old ruling hash in the artifact it fails on both ruling entries (the file changed after it was cited)
A blob that is absent in a blobless partial clone skips the locator check, and a path that is not in the tree at the pin still fails	local_integration (throw-away repositories with the blob, then a subtree, then the path missing)	tests.test_rtk_receipt_locators (6 tests); 7 of 7 single-edit mutants of the check caught (table sha256 d83b314ab9d83c7507284e6fc9349f20a31d540a81d715a26f87f7c3d0f527e5)
The metering sentence of the usage row recomputes from two fixed windows of the host's counter and events	native_proven (aggregate figures of the local Prometheus and Loki) and local_integration	tests.test_native_practice_meter_windows (6 tests)
A pin whose tree is missing is skipped, and a bad path with the tree present still fails	local_integration (a throw-away repository with the tree object removed)	tests.test_rtk_receipt_locators (4 tests)

Mutation table (tests/test_paired_runs_results.py, orchestration class; each edit is applied alone and the module is run)

19 of 19 caught: a run dropped from the grid; F3-C recorded as failing; the void pair F3 A-C given a preference; a counted preference that ignores a held refutation; the primary analysis saying B clears the rule; the cost bound of C flipped; the mean cost of A changed; F3-C recorded as having its diff captured; a paper hash that is not a sha256; a run's result hash dropped; a pair's judge hash shortened; the Workflow's cost ratio changed in a row; the team's comparable-fixture count changed; the verifier count changed; the cited RESULT.md hash changed; the TEAM-DONE count changed; the record's void-pair statement dropped; the record's defect count changed; the record's routing statement dropped.

Routing class (paired run 5), 31 of 31 caught, one control passing (table sha256 d21cb7ba6c0ab2dbf4d16ff62454cee0c68e291463f84e4e0e9816f64f2b9c1d; the 19 of the fourth commit and, in the fifth: the ruling's old hash back, the ruling without the condition for Sonnet, the record paragraph without the condition, without the sentence on the ruled extension, the tolerant figure back to the per-finding one, the per-finding strict figure removed, the basis of the two-trial verdict removed, the tolerant floor modelling one clause again, the heading back to one date, the agent-teams Overturn when back, the task-list alternative without the completion clause, the pooled rate behind the per-finding figure changed): a run dropped from the grid; a run recorded under the other arm; Sonnet credited with the missed finding (the rule then says Sonnet takes micros); a finding credited to a run of another head; Sonnet's costs doubled; a run that did not end in a structured result; a Sonnet run recorded as the Opus model; the recorded verdict flipped; the recorded missed finding dropped; the recorded cost ratio changed; the recorded noise floor changed; the ruling's hash changed; a paper hash that is not a sha256; a verdict flipped; a thirteenth decisive finding; the record paragraph's detection count changed; run 5 back in the owed table; run 4's row saying its result is a later record again; the record paragraph saying a slot row changed. The control (a changed M1 meter, which no rule input reads) passes.

Local commands run

Head bf864c3705dbca1fc0b5c24ba388e5836e4c35d0 on ba6ab3a5c1f9ea2cf54caf75ab317f580629f872 (CPython 3.13.16 locally), TMPDIR outside /tmp and outside home:

$ python3 -E -s -B -m unittest tests.test_paired_runs_results tests.test_rtk_receipt_locators tests.test_primary_source_reading_counts tests.test_native_practice_snapshot_provenance tests.test_native_practice_doc_corrections tests.test_evidence_manifest tests.test_upstream_surface_watch tests.test_practice_references tests.test_record_verdicts tests.test_skills_manifest tests.test_skills_status tests.test_runtime_worker_skills tests.test_landscape_sweep_skills tests.test_install_skills tests.test_native_practice_retired_routes tests.test_gitleaks_config tests.test_skill_usage
Ran 924 tests in 181.227s; OK (skipped=16)   (rc 0)
$ python3 -E -s -B scripts/validate.py
{"components": 70, "hashed_files": 11474, "profiles": 4, "receipts": 239, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/evidence_manifest.py --check
{"files": 11474, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/component_matrix.py --write && python3 -E -s -B scripts/new_host_grand_list.py --write
rc 0; no change to a generated report
$ python3 -E -s -B scripts/validate_convergence.py --all-recorded
rc 0
$ python3 -E -s -B -m unittest tests.test_rtk_receipt_locators        # in a worktree at 218efbd9dec8be5ff320e50f62b9ab64e0396cbb with the new module copied in
FAILED (failures=4)   (exit 1)
$ python3 -E -s -B -m unittest tests.test_rtk_receipt_locators        # on this head
OK   (exit 0)
$ python3 -E -s -B -m unittest tests.test_skill_usage.ManagedMcpDefaultLocations        # tests/test_skill_usage.py with the old comment restored, the new tests kept
FAILED (failures=1)   (exit 1)
$ python3 -E -s -B -m unittest tests.test_skill_usage.ManagedMcpDefaultLocations        # on this head
OK   (exit 0)

Head 5c5eecaadbbaf0fd4ad2c39236fb820b3b47fbf4 (the working tree as tested, committed unchanged) on `bf864c3705dbca1fc0b5c24ba388e5836e4c35d0`, same environment:
$ python3 -E -s -B scripts/validate.py
{"components": 70, "hashed_files": 11476, "profiles": 4, "receipts": 239, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/evidence_manifest.py --check
{"files": 11476, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/component_matrix.py --write && python3 -E -s -B scripts/new_host_grand_list.py --write
rc 0; the only changes are the 11 files of the commit
$ python3 -E -s -B scripts/validate_convergence.py --all-recorded
rc 0
$ python3 -E -s -B -m unittest tests.test_paired_runs_results tests.test_rtk_receipt_locators tests.test_native_practice_meter_windows tests.test_primary_source_reading_counts tests.test_native_practice_snapshot_provenance tests.test_native_practice_doc_corrections tests.test_evidence_manifest tests.test_upstream_surface_watch tests.test_practice_references tests.test_record_verdicts tests.test_skills_manifest tests.test_skills_status tests.test_runtime_worker_skills tests.test_landscape_sweep_skills tests.test_install_skills tests.test_native_practice_retired_routes tests.test_gitleaks_config tests.test_skill_usage
Ran 938 tests in 174.412s; OK (skipped=17)   (rc 0)
$ the #1001 forward commit's patch applied on this head, then tests.test_native_practice_retired_routes tests.test_native_practice_meter_windows tests.test_paired_runs_results tests.test_rtk_receipt_locators
Ran 63 tests; OK   (the stacked state that lands)
$ the same stack with the slots.md index cell restored to the old Overturn when
FAILED (failures=1: test_the_index_and_the_layer_pages_show_each_slots_overturn_condition)
$ git diff HEAD~1 --unified=0 | grep '^+'   # added lines, scanned for the host's private names
hits 0

Head 425240f6e9342675f730e970d2078db5d59490e4 (the working tree as tested, committed unchanged) on `5c5eecaadbbaf0fd4ad2c39236fb820b3b47fbf4`, same environment, after `scripts/host_receipts.register_file` had registered the changed files:
$ python3 -E -s -B scripts/validate.py
{"components": 70, "hashed_files": 11476, "profiles": 4, "receipts": 239, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/evidence_manifest.py --check
{"files": 11476, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/component_matrix.py --write && python3 -E -s -B scripts/new_host_grand_list.py --write
rc 0; the only changes are the 10 files of the commit
$ python3 -E -s -B scripts/validate_convergence.py --all-recorded
rc 0
$ python3 -E -s -B -m unittest <the same 18 modules>
Ran 943 tests in 150.593s; OK (skipped=16)   (rc 0)
$ PAIRED_PAPER_ROOTS=<the research directory>:<the command center's directory> python3 -E -s -B -m unittest tests.test_paired_runs_results
Ran 28 tests; OK (the gated test runs: 24 private papers resolve to their recorded hashes and sizes)
$ python3 -E -s -B -m unittest tests.test_paired_runs_results tests.test_native_practice_meter_windows tests.test_rtk_receipt_locators tests.test_native_practice_snapshot_provenance
Ran 48 tests; OK (the label test of #1001 is not on this branch; the exact-difference test runs, nothing is skipped but the gated papers test)
$ 31 single-edit mutants of the routing data and text; 7 single-edit mutants of the locator check
31 of 31 caught, the control passes; 7 of 7 caught
$ git diff HEAD~1 --unified=0 | grep '^+'   # added lines, scanned for the host's private names
hits 0

Head 49bb6494e152db393646a4782a2fdec07463bc72 (the working tree as tested, committed unchanged) on `425240f6e`, same environment, after `scripts/host_receipts.register_file` had registered the changed files:
$ python3 -E -s -B scripts/validate.py
{"components": 70, "hashed_files": 11476, "profiles": 4, "receipts": 239, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/evidence_manifest.py --check
{"files": 11476, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/component_matrix.py --write && python3 -E -s -B scripts/new_host_grand_list.py --write
rc 0; no change to a generated report beyond the 9 files of the commit
$ python3 -E -s -B scripts/validate_convergence.py --all-recorded
rc 0
$ python3 -E -s -B -m unittest <the same 18 modules>
Ran 945 tests in 152.774s; OK (skipped=16)   (rc 0)
$ PAIRED_PAPER_ROOTS=<the research directory>:<the command center's directory> python3 -E -s -B -m unittest tests.test_paired_runs_results
Ran 30 tests; OK (the gated test runs: 26 private papers resolve to their recorded hashes and sizes, and lines 17-18 of the retained addendum 1 carry the registered reading)
$ the new tests of the commit run against the data of 425240f6e
FAILED (failures=2, errors=3): the figure test, the record test and the ruling test error on the missing keys; the time-bounds test and the gated papers test fail (retained fail-before-micro3.txt)
$ git diff HEAD~1 --unified=0 | grep '^+'   # added lines, scanned for the host's private names
hits 0

Head 1190b957ff3ce2fe6e0475662b78ab3972a5ca75 (the working tree as tested, committed unchanged) on `49bb6494e`, same environment, after `scripts/host_receipts.register_file` had registered the changed files:
$ python3 -E -s -B scripts/validate.py
{"components": 70, "hashed_files": 11476, "profiles": 4, "receipts": 239, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/evidence_manifest.py --check
{"files": 11476, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/component_matrix.py --write && python3 -E -s -B scripts/new_host_grand_list.py --write
rc 0; no change to a generated report beyond the files of the commit
$ python3 -E -s -B scripts/validate_convergence.py --all-recorded
rc 0
$ python3 -E -s -B -m unittest <the same 18 modules>
Ran 946 tests in 338.407s; OK (skipped=16)   (rc 0)
$ PAIRED_PAPER_ROOTS=<the research directory>:<the command center's directory> python3 -E -s -B -m unittest tests.test_paired_runs_results
Ran 31 tests; OK (the gated test runs: 29 private papers resolve to their recorded hashes and sizes, and the three cited lines hold what the artifact says)
$ the patched test modules on the data of 49bb6494e
FAILED (failures=3, errors=1): the ceiling test, the orchestration-rows test, the gated papers test and the time-bounds test (retained fail-before-wave.txt)
$ 13 single-edit mutants of the setting, its sources and the sentences, 1 control
13 of 13 caught, the control passes (table sha256 4ede6e1d...)
$ git diff HEAD~1 --unified=0 | grep '^+'   # added lines, scanned for the host's private names and for characters outside printable ASCII
hits 0; characters outside printable ASCII: none

10 files changed, 251 insertions(+), 51 deletions(-) in the fifth commit; 9 files changed, 131 insertions(+), 45 deletions(-) in the sixth commit; 8 files changed, 89 insertions(+), 30 deletions(-) in the seventh commit; 19 files changed, 4227 insertions(+), 107 deletions(-) against the old base (main 8bffa2c41, before the rebase; the figures of the rebased head are in the last block below).

Head a00744f0e84455714a17aa7ec73fec22202dd1b9 (the working tree as tested, committed unchanged) on main `b490361c5`, after the rebase (relabel last, `evidence_manifest.py --write`, `host_receipts.register_file` for the 12 files this branch adds or changes, then validate):
$ python3 -E -s -B scripts/validate.py
{"components": 70, "hashed_files": 11654, "profiles": 4, "receipts": 239, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/evidence_manifest.py --check
{"files": 11654, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/component_matrix.py --write && python3 -E -s -B scripts/new_host_grand_list.py --write
rc 0; no change to a generated report
$ python3 -E -s -B scripts/validate_convergence.py --all-recorded
rc 0
$ PAIRED_PAPER_ROOTS=<...> python3 -E -s -B -m unittest <the same 18 modules> tests.test_native_practice_retired_routes
Ran 1015 tests in 388.469s; OK (skipped=16)   (rc 0; with #1001's strict label test and its guard on main)
$ git diff b490361c5 a00744f0e --stat
14 files changed, 3708 insertions(+), 57 deletions(-)

Follow-ups

Listed under the command center's convergence ruling of 2026-10-11T13:37Z (research-cc-role-20261008/ruling-pr-wave-convergence-20261011.md): the head is not moved until the three families have reported, and none of these is a blocker unless the adjudication says so.

• Codex P1 at .claude/skills/claude-native-practice/reference/observation-inference.md:13 (13:36:20Z, head 49bb6494e): the sentence that api_request events book advisor dollars under the executor's model is supported by the hash of a private working paper and by the aggregate meter artifact, and Codex asks for a pinned client or source locator and a reproducible attribution test. Adjudicated by the command center (2026-10-11T13:50Z): a follow-up, not a P1, because the figures were verified from the private records and nothing in the line is false; the line already names its evidence class (the advisor run's private working paper, sha256 2377eb9ce915, and the aggregate meter artifact meter-windows-20261011.json), so it is left as it is. Proposed follow-up: cite the vendor's monitoring documentation of the api_request event once it is read for this purpose.
• Codex P1 12 (the advisor attribution) is resolved; the micro's P3-1 follow-up is to cite the vendor's monitoring documentation (monitoring-usage.md lines 807-818, advisor.md line 15) in the line.
• Codex P2 at evidence/artifacts/claude-native-practice-20261009/claude-native-slot-manifest.json:57 (15:40:11Z, head 1190b957f): the manifest's note still says the scheduled-tasks route was narrowed "to jobs that run no model". It is Retire the skill's Claude-in-GitHub rows and point them to local review #1001's text (from ba6ab3a5c), corrected at bdf60fc62; this branch gets the corrected note when it is rebased onto main after Retire the skill's Claude-in-GitHub rows and point them to local review #1001 lands (the rebase takes main's note and recomputes the label last).
• The micro's P3 (a), by the command center's ruling kept out of the landing rebase so that it stays pure: the prose of design.caps.bg_wait_ceiling_setting is not pinned (only the number field client_default_bg_wait_ceiling_minutes is tested, so "the client's own default is 45 minutes" in the prose passes); optional fix assertIn("default is 10 minutes", caps["bg_wait_ceiling_setting"]). And the record's "the client's default is 10" lacks the unit (every row says "10 minutes").
• Informational: the runner is cited as paired/orch_run.py line 137 (the run-time copy in the lane's coordination directory) in the caps prose and as paired-orchestration/orch_run.py in papers.runner (the retained copy); the two are byte-identical (sha256 ad8ffbef…).
• CI: the pull_request workflows did not run at 1190b957f because the PR was DIRTY against main (GitHub does not run them on a conflicted pull request); they run at the rebased head a00744f0e, and the command center lands on those.
• The strict label test of Retire the skill's Claude-in-GitHub rows and point them to local review #1001 is on main and passes here; the rebase recomputed the label and the list of differences (the conflicts were the slot manifest and the evidence.json rows, both resolved by taking main's side and re-registering this branch's files, as rehearsed).
• The scheduled-tasks wording is Retire the skill's Claude-in-GitHub rows and point them to local review #1001's text and landed there; this branch carries no copy of it any more (the rebased diff has no scheduling-supervision.md).

Decision record

docs/decisions/2026-10-09-claude-code-native-practice.md, the addendum "Addendum (2026-10-10 and 2026-10-11): paired-run results" (new) and the owed-work table. What would overturn it: a paired run on a larger fixture set or a second trial that moves the planning cells or the orchestration preferences (each paper's preregistration states the rule), a new client version that changes the Workflow or team mechanics the rows describe (the surface watch carries the check), or a new advisor run (the first was stopped on 2026-10-11 with no verdict) or a routing run on more heads, other efforts or gate reads (the registration of run 5 allows no third trial).

Checklist

• No GitHub Actions or workflow files are added or changed.
• No secrets, account identifiers or private host names are printed, logged or committed; the private papers are cited by sha256 only.
• No new paid hosting, subscription or billing surface; costs in the rows are API-list-equivalent, not billed.
• Peer-owned untracked files and worktrees were preserved.

🤖 Generated with Claude Code

[seathatflowsinourveins]

@codex review

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-11T15:40:13.922345Z	1190b95	Manual request

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: bf864c3705

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[seathatflowsinourveins]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5c5eecaadb

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[seathatflowsinourveins]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 425240f6e9

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[seathatflowsinourveins]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 49bb6494e1

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[seathatflowsinourveins]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1190b957ff

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".