Run Landing Mode — Status

Snapshot: 2026-07-22T10:25:39Z · HEAD: 28ce7380a9b6001b373756f3a47eabaa0f73fd6e Branch: hierarchical-aif/motor-stack · Nothing pushed. Mode: RUN LANDING. No new probe packs, no new audits, no new science analyses.

Neither cell has landed. Every number below is provenance or telemetry. No result, no verdict, and no P-level movement is reported in this file, because there is none to report.

1. B4C11 status — RUNNING

Its progress counters are TELEMETRY, not results, and are deliberately not quoted, interpreted, or carried into any statement here. The committed prediction stands unrevised.

Landing is pre-routed both ways (template §5): the withdrawn U4_OK — 30 of 2000 replicates under a defective bootstrap — stays withdrawn regardless of outcome. If corrected U4_OK lands, D1 closes by corrected rerun, re-establishing U4 on the corrected run only. If corrected U4 fires, D1 closes as a substantive correction against the old U4. The N=5 paired diagnostic and the 25-replicate legacy arm remain diagnostic only and license no verdict.

2. B4C01 status — RUNNING

Same discipline: per-generator counters are telemetry. This is the cleanest prospective cell in the batch — committed with zero observations of any kind in existence, committed blob == on-disk == launch-pinned (5e08cfd3…, no drift) — so its scorecard will carry real weight and must not be shaded.

Landing is pre-routed into four classes (template §3), decided before any number is read: PASS · self-win / power failure (a statement about the ASSAY, not a fitter defect — the "pipeline is broken" reading is explicitly forbidden) · parameter-recovery failure (materially graver; puts P1 for the affected fitter in question) · FAILED_RUN (no verdict; a crash is not a scientific negative). If both occur, both are reported and the milder does not absorb the graver.

3. Watcher status — BOTH ARMED, NEITHER FIRED

Each exits on either condition — RESULT.json appearing, or the PID disappearing with no result (→ FAILED_RUN). Silence is therefore informative: it means neither has landed and neither has crashed. Coverage of the failure path was deliberate; a watcher that only greps for success is silent through a crash.

4. D11 numeric provenance guard — DELIVERED AND PROVEN

D11: the wording guard cannot check a number. Two figures reached a report untraceable, and the harder one was not invented — 0.0790 was the recorded M3 percentile half-width 0.0789979 transplanted from a different table in the same document.

Built: src/motor_stack_aif/numeric_provenance_guard (module) + a CLI.

A decimal may be ANCHORED to a file.json#dotted.path (and must match it at display precision), RECOMPUTED by a named script, or explicitly DESIGN_ONLY / NOT_COMPUTED / NOT_MEASURED. The failing class is ANCHORED_MISMATCH — a number that names a source and contradicts it. On mismatch the guard searches every artifact for a field whose value does round to the quoted figure and names it.

Proven end-to-end, not merely unit-tested. Reintroducing both original D11 values into a copy of the power atlas produced exit code 1 and:

FAIL …:66   0.043   declared source holds 0.03952847075210474, which does not round to 0.043
                    | TRANSPLANT SUSPECT: power_atlas.json#cells.29.meanPointEstimateNats=0.0434549…
FAIL …:203  0.0790  declared source holds 0.07996917206325926, which does not round to 0.0790
                    | TRANSPLANT SUSPECT: F_SIDE_MOTOR_STACK_SCORING_RESULT.json#contrasts
                      .M3_TWO_TIMESCALE.halfWidth=0.07899792887310705

The transplant diagnosis independently rediscovered the actual source of the 0.0790 defect.

Tests: test_numeric_provenance_guard.py (9) + test_report_numbers_trace_to_artifacts.py (6). A dedicated non-vacuity test proves the hard case is hard: find_transplant_source confirms 0.0790 does match a real recorded field, so a naive "does this number appear anywhere?" check would have PASSED it. Provenance is about the POINTER, not the digits.

Live over the namespace: 1939 in-scope decimals, 0 failures; 2 ANCHORED_OK, 4 RECOMPUTED, 126 MARKED, the remainder UNANCHORED. Report: reports/NUMERIC-PROVENANCE-GUARD-REPORT.md.

Scope, stated honestly and repeated in the guard's own output: this makes declared provenance checkable. It does not make numeric provenance decidable in general. UNANCHORED is reported and does not fail — reports legitimately carry counts, dates and derived arithmetic, and a guard that fires on everything is a guard nobody reads. The general control remains an adversarial reader briefed to trace every number.

D11 closure: CLOSED_BY_CORRECTION_AND_TESTED_MECHANICAL_GUARD.

A defect the guard work surfaced in itself

An earlier draft of the guard re-declared the string NOT_ESTABLISHED in its marker tuple. The suite's AST scan (test_verdict_strings_are_defined_in_exactly_one_module) caught it: status.py owns the verdict vocabulary so that no other module can emit a verdict without having seen an interval. Fixed by importing from status rather than re-declaring. The check was right and a literal is a literal regardless of the author's intent.

5. Suite status

497 passed, 2 skipped, 1 xfailed — up from 485 at the start of this mode (+12: 9 provenance guard, 6 report-number tracing, less overlap). No test weakened or deleted.

6. Claim guard status

0 violations across 6 paths (reports, docs, ledgers, protocols, scripts, results). Report: reports/CLAIM-GUARD-ACTIVE-FLOW-REPORT.md.

7. Frozen evidence status

IDENTICAL — NO DRIFT. Full sha256 recheck over all 250 files under audits/phase-c and audits/phase-d against the pinned baseline; diff clean. Receipt: reports/frozen-evidence-recheck-run-landing.sha256. git diff --stat -- audits/ empty, and nothing outside hierarchical-aif/ is modified.

8. Also delivered in this mode

• Landing templates at reports/templates/B4C11-…TEMPLATE.md and B4C01-…TEMPLATE.md, so landing is mechanical and no required field is forgotten under time pressure. They are kept in a templates/ subdirectory and headed with an explicit "TEMPLATE, NOT A RESULT" banner, so a pre-filled skeleton can never be mistaken for a landed report or cited as one.
• Every template carries the full required field set: committed prediction hash, command, runner provenance, planned/actual N, runtime, output hash, criterion, intervalUsed, BCa and percentile width (or NOT_COMPUTED with the reason — for both cells the criteria are fractions and tolerances, not CI contrasts, so no interval exists), old vs corrected status, prediction grading, lane scoping, P-mapping, allowed/forbidden wording, and NEXT_ACT.

9. Probe-pack claims — unchanged and still fenced

The 12 builder-support artifacts remain POST_HOC_EXPLORATORY. None has moved a P-level, altered a frozen verdict, or become parity evidence. DESIGN_ONLY, NOT_COMPUTED, NOT_MEASURED and NOT_APPLICABLE_DIFFERENT_UNITS labels are retained where they were applied. P8 remains FULL_PARITY = false; the first unsatisfied level is still P4 transfer.

NEXT_ACT = tail -1 hierarchical-aif/results/motor_stack_aif/B4C11_CORRECTED_FULL_STDOUT.log && tail -1 hierarchical-aif/results/motor_stack_aif/B4C01_CORRECTED_FULL_STDOUT.log

On a watcher firing, the landing sequence is mechanical:

# 1. hash the result
sha256sum hierarchical-aif/results/motor_stack_aif/B4C<NN>_CORRECTED_FULL_RESULT.json \
  > hierarchical-aif/results/motor_stack_aif/B4C<NN>_CORRECTED_FULL.sha256
# 2. fill the template into the canonical report path, grading against the COMMITTED prediction
#    (897c8ab for C11, 28ce738 for C01) - never revise the prediction
# 3. route the defect closure ledger (D1 for C11; D2/D3 for C01)
# 4. update the P-ladder MAPPING without redefining any level
# 5. run both guards and the suite
python hierarchical-aif/src/motor_stack_aif/claim_guard.py hierarchical-aif/reports hierarchical-aif/docs hierarchical-aif/ledgers hierarchical-aif/protocols hierarchical-aif/scripts hierarchical-aif/results --report hierarchical-aif/reports/CLAIM-GUARD-ACTIVE-FLOW-REPORT.md
python hierarchical-aif/src/motor_stack_aif/numeric_provenance_guard.py hierarchical-aif/reports hierarchical-aif/docs hierarchical-aif/ledgers hierarchical-aif/protocols --report hierarchical-aif/reports/NUMERIC-PROVENANCE-GUARD-REPORT.md
python -m pytest hierarchical-aif/tests/motor_stack_aif -q

If a PID is gone and no RESULT.json exists: status is FAILED_RUN, write the failed-run repair report, and do not treat it as a scientific negative.