Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
80 changes: 80 additions & 0 deletions .github/workflows/cbz-auditor-origin.yml
Original file line number Diff line number Diff line change
@@ -0,0 +1,80 @@
name: CBZ auditor assertion-origin controls
'on':
push:
branches: [main]
paths:
- '.github/workflows/cbz-auditor-origin.yml'
- 'tests/cbz_auditor_origin*'
- 'tests/cbz_inventory_origin.py'
- 'tests/test_cbz_auditor_origin*'
- 'tests/test_cbz_inventory_origin.py'
- 'tests/test_cbz_layer_a_scan*'
- 'tests/test_cbz_a3_float_census.py'
- 'tests/test_cbz_auditor_hardening.py'
- 'tools/cbz/layer_a_scan.py'
- 'tools/cbz/a3_float_census.py'
pull_request:
paths:
- '.github/workflows/cbz-auditor-origin.yml'
- 'tests/cbz_auditor_origin*'
- 'tests/cbz_inventory_origin.py'
- 'tests/test_cbz_auditor_origin*'
- 'tests/test_cbz_inventory_origin.py'
- 'tests/test_cbz_layer_a_scan*'
- 'tests/test_cbz_a3_float_census.py'
- 'tests/test_cbz_auditor_hardening.py'
- 'tools/cbz/layer_a_scan.py'
- 'tools/cbz/a3_float_census.py'
workflow_dispatch:
permissions:
contents: read
jobs:
origin-controls:
runs-on: ubuntu-24.04
timeout-minutes: 20
steps:
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1
with:
persist-credentials: false
- uses: actions/setup-python@6d4fd22b20465ccc0f7354476588e8df7c9ca0e0
with:
python-version: '3.11.16'
- name: Run both actual modes and retain each mutation observation
shell: bash
run: |
set -uo pipefail
evidence="$RUNNER_TEMP/cbz-auditor-origin"
mkdir "$evidence" || exit 1
git rev-parse HEAD > "$evidence/commit.txt" || exit 1
git status --porcelain=v1 --untracked-files=all > "$evidence/worktree-before.txt" || exit 1
if [ -s "$evidence/worktree-before.txt" ]; then echo 'source checkout is not clean' >&2; exit 1; fi
python --version > "$evidence/python.txt" 2>&1 || exit 1
sha256sum .github/workflows/cbz-auditor-origin.yml tests/cbz_auditor_origin.py tests/cbz_auditor_origin_mutation_child.py tests/cbz_inventory_origin.py tests/test_cbz_auditor_origin.py tests/test_cbz_auditor_origin_mutations.py tests/test_cbz_auditor_origin_workflow.py tests/test_cbz_inventory_origin.py tests/test_cbz_layer_a_scan.py tests/test_cbz_layer_a_scan_mutants.py tests/test_cbz_auditor_hardening.py tests/test_cbz_a3_float_census.py tools/cbz/layer_a_scan.py tools/cbz/a3_float_census.py > "$evidence/source-sha256.txt" || exit 1
failed=0
for mode in normal optimized; do
mkdir "$evidence/$mode" "$evidence/$mode/census" "$evidence/$mode/adapter" || exit 1
flags=(-B -S)
if [ "$mode" = optimized ]; then flags=(-B -O -S); fi
if CBZ_AUDITOR_ORIGIN_MUTATION_EVIDENCE_DIR="$evidence/$mode/adapter" python "${flags[@]}" -m unittest tests.test_cbz_auditor_origin tests.test_cbz_inventory_origin tests.test_cbz_auditor_origin_mutations tests.test_cbz_auditor_origin_workflow -v > "$evidence/$mode/origin.log" 2>&1; then status=0; else status=$?; failed=1; fi
printf '%s\n' "$status" > "$evidence/$mode/origin.exit" || failed=1
if CBZ_AUDITOR_EVIDENCE_DIR="$evidence/$mode/census" python "${flags[@]}" -m unittest tests.test_cbz_layer_a_scan tests.test_cbz_auditor_hardening tests.test_cbz_layer_a_scan_mutants tests.test_cbz_a3_float_census -v > "$evidence/$mode/census.log" 2>&1; then status=0; else status=$?; failed=1; fi
printf '%s\n' "$status" > "$evidence/$mode/census.exit" || failed=1
grep -Eq '^Ran 22 tests in ' "$evidence/$mode/origin.log" || failed=1
grep -Eq '^Ran 83 tests in ' "$evidence/$mode/census.log" || failed=1
grep -Fxq 'OK' "$evidence/$mode/origin.log" || failed=1
grep -Fxq 'OK' "$evidence/$mode/census.log" || failed=1
if grep -Fq ' ... skipped ' "$evidence/$mode/origin.log" "$evidence/$mode/census.log"; then failed=1; fi
[ "$(find "$evidence/$mode/adapter" -name '*.json' -type f | wc -l)" -eq 10 ] || failed=1
[ "$(find "$evidence/$mode/census" -maxdepth 1 -name '*.json' -type f | wc -l)" -eq 202 ] || failed=1
done
sha256sum -c "$evidence/source-sha256.txt" > "$evidence/source-readback.log" 2>&1 || failed=1
exit "$failed"
- name: Preserve complete observations including failures
if: always()
uses: actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9
with:
name: cbz-auditor-origin-${{ github.run_id }}-${{ github.run_attempt }}
path: ${{ runner.temp }}/cbz-auditor-origin
include-hidden-files: true
if-no-files-found: error
retention-days: 90
15 changes: 15 additions & 0 deletions BASELINE.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,20 @@
# CBZ baseline ledger

## Auditor-origin candidate — 9 November 2026

The observed published cut has Main `05f2289478a0ee9492adc1eb501466e0f66866df`,
Math `9bdfc54b76173b57912503bfcf901a5f9d43b09c` and Workspace
`bda15fcd29d6d484f9b2e1233663dbd8fdd8b796`. The Workspace pin and contract
runs passed their bounded controls, and the Main normal discovery retained
24 skips. A later Main merge needs a fresh guarded Workspace pin update.

Two Main auditor mutation graders were observed crediting subject assertions
as kills. The source repair and its actual current-head tests are a candidate,
not a final G0/G1 or Layer A clearance. The intentional negative axiom fixture,
historical unparseable Python writer and Workspace gitlink refusal remain
visible in the scoped scan evidence. No CBZ-0 or terminal Partial is declared.


**State: active intermediate baseline; CBZ-0 is not established. No terminal Partial Baseline is declared.**

This ledger binds the currently observed repository snapshot to the twelve
Expand Down
24 changes: 24 additions & 0 deletions DECISIONS.md
Original file line number Diff line number Diff line change
Expand Up @@ -172,3 +172,27 @@ Append-only. Historical source outcomes remain in their original records.
engineering test, not continuum closure or independent scientific review.
Its disposition is reversible through Git history; prior outcomes remain
preserved.


## 2026-11-09 — auditor assertion-origin repair

- Preserve the two real Main grader false kills in each Python mode at
`05f2289478a0ee9492adc1eb501466e0f66866df`, their raw child streams and
the failed local attempts in [the evidence record](cbz/evidence/auditor-origin-20261109/README.md).
A subject assertion is not an authorized oracle outcome.
- Reuse the existing pre-cleanup live-code recorder for each designated test,
require an unchanged-source clean baseline and reject missing provenance,
errors, undeclared skips or source drift. Require actual source mutants and
adverse controls in both modes; keep declared inner recursion skips visible.
Tighten the synthetic formatted-traceback-only expectation to refusal. This
changes no scanner subject, theorem, bound or historical result.
- Record the scoped Main negative-axiom fixture finding, Math's historical
unparseable documentary snapshot and Workspace's gitlink refusal. Do not
suppress or rewrite any of them to obtain a green audit. Exact-head hosted
qualification remains necessary; the action is reversible in Git.

- A later synthetic-producer probe of the actual workflow shell found that an
exit-status evidence write could fail while the shell still returned 0.
Preserve the red output in the evidence record and fail the job on either
metadata-write error; test both guard removals in each Python mode. This
controls evidence completeness, not mathematical truth or scientific status.
13 changes: 13 additions & 0 deletions GOVERNANCE.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,18 @@
# CBZ-0 governance

## Auditor assertion-origin boundary — 9 November 2026

A mutation kill requires a clean designated baseline and an actual failure
from the selected test body or an explicitly authorized standard unittest
assertion helper. Capture live traceback code provenance before unittest
cleanup and reject missing provenance, subject/helper failures, unexpected
skips, errors and source drift. The recorder's supported domain is trusted
synchronous unittest, not hostile code. Historical filename-based false kills
remain preserved and cannot qualify a future gate. Candidate commands and
limits are registered in [TOOLING.md](TOOLING.md#auditor-origin-controls);
current-head tests and the final federation G0–G11 pass remain required.


The later Math mutation-grader identity defect and its bounded draft correction
are recorded in [STATUS.md](STATUS.md#later-node-inventory-and-pin-observations--9-november-2026).
Their historical kernel evidence and scientific limits are unchanged.
Expand Down
11 changes: 11 additions & 0 deletions PROOF_INDEX.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,16 @@
# CBZ proof and quarantine index

## Auditor-origin boundary — 9 November 2026

The Main and Workspace engineering changes and the auditor-grader repair do
not modify any theorem, proof disposition, authorship or scientific claim.
The unapproved axiom in `cbz/contributions/math514/kernel_replay/tests/fixtures/UnapprovedAxiom.lean`
is a deliberate negative test fixture, not a closed mathematical module; its
scanner finding remains recorded. The bounded lexical census does not prove
its transitive isolation or establish a complete axiom audit. Historical Math
conditional kernel results and continuum limits remain as recorded below.


**State: incomplete inventory; CBZ-0 is not established.**

This is the Main cold-start map for proof status. It records exact boundaries;
Expand Down
29 changes: 29 additions & 0 deletions STATUS.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,34 @@
# CBZ-0 execution status

## Auditor-origin repair and published pin cut — 9 November 2026

Main #791 merged as `05f2289478a0ee9492adc1eb501466e0f66866df`; its hosted
navigation run executed 2,609 tests with 24 skips, and the selected historical
#525 binding test passed in each mode. Workspace #62 merged as
`bda15fcd29d6d484f9b2e1233663dbd8fdd8b796`, pinning that Main commit and Math
`9bdfc54b76173b57912503bfcf901a5f9d43b09c`. The Workspace tested and
postmerge contract runs each executed 254 tests per mode without top-level
skips; pin checks covered eight current pins and four canonical inputs. These
sequential observations do not establish final G5 or G7.

At Main `05f22894`, actual normal and optimized subject-origin probes showed
that both Lean and Python-census mutation graders could count a subject's
`AssertionError` as a test-oracle kill when the subject was under `tests/`.
The original failures and source identities are preserved in
[the evidence record](cbz/evidence/auditor-origin-20261109/README.md). This
candidate uses live pre-cleanup code provenance and requires a clean baseline
for each source mutant; current-head qualification remains required. Earlier
mutation counts are historical observations, not sound final G0 credit.

The scoped Lean lexical scans found a deliberately unapproved Main test axiom,
an `unsafe` review token and no listed tokens in 1,974 Math Lean blobs; the
Workspace scan refused a gitlink. The Math Python census failed on one
intentionally preserved historical writer parse error. These are scoped
observations, not whole Layer A clearance or proof that float signals certify
bounds. Exact commands and limits are in [TOOLING.md](TOOLING.md#auditor-origin-controls).
No CBZ-0, terminal Partial or scientific acceptance follows.


**State: active intermediate baseline; CBZ-0 is not established. No terminal Partial Baseline is declared.**

This entry adopts the owner's expanded [Master Directive](governance/MASTER_DIRECTIVE_20261029_SWARM.txt)
Expand Down
35 changes: 35 additions & 0 deletions TOOLING.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,40 @@
# CBZ mission-tool registry

## Auditor origin controls

**Purpose and scope.** `tests/cbz_auditor_origin.py` adapts the existing
`tests/cbz_inventory_origin.py` pre-cleanup recorder for the Lean lexical and
Python A3 census mutation suites. It admits only selected synchronous test
bodies and the recorder's finite standard-unittest oracle set; it is not a
hostile-child sandbox, a mathematical proof or a complete federation auditor.
For each selected mutant it records unchanged subject-byte hashes and requires
a clean selected baseline. The two historical full-suite baselines explicitly
check and retain their two and four inner recursion-guard skips, respectively.

**Commands.** From a clean Main checkout, with new evidence directories outside
the checkout (omit either variable only when retained child evidence is not
needed), run each group in normal and optimized mode:

```sh
mkdir /absolute/new-census-evidence /absolute/new-adapter-evidence
CBZ_AUDITOR_ORIGIN_MUTATION_EVIDENCE_DIR=/absolute/new-adapter-evidence python3 -B -S -m unittest tests.test_cbz_auditor_origin tests.test_cbz_inventory_origin tests.test_cbz_auditor_origin_mutations tests.test_cbz_auditor_origin_workflow -v
CBZ_AUDITOR_EVIDENCE_DIR=/absolute/new-census-evidence python3 -B -S -m unittest tests.test_cbz_layer_a_scan tests.test_cbz_auditor_hardening tests.test_cbz_layer_a_scan_mutants tests.test_cbz_a3_float_census -v
# Repeat both commands with -O and separate new evidence directories.
```

**Verification.** The preserved pre-repair real subject-origin failures are
[here](cbz/evidence/auditor-origin-20261109/README.md). The added controls
exercise clean subjects, actual selected-body failures, direct and nested
subject failures, changed source, missing provenance, forged text and
failed evidence writes. Five adapter source mutations have designated
clean-baseline and actual-failure controls; the two existing source campaigns
contain 74 Lean and 27 A3 mutations. The dedicated workflow preserves original
child reports and source snapshots in both modes. Its actual shell guards are
also exercised with synthetic child outcomes and seven guard-removal controls. Actual exact-head hosted
execution and independent readback are required before qualification; historical
pilot observations do not establish final G0, G1 or scientific acceptance.


**State: component registry; the final federation toolchain is incomplete.**

This file lists runnable mission-relevant tools present in Main. The latest
Expand Down
67 changes: 67 additions & 0 deletions cbz/evidence/auditor-origin-20261109/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,67 @@
# Auditor assertion-origin repair: preserved observations

The base Main commit is `05f2289478a0ee9492adc1eb501466e0f66866df`. At that
commit the Lean grader was blob `faab4ff7e4b515320a95238168ed2d641c226677`
and the A3 grader was blob `b7b56615fa6b11b88d74800478e156be883bedc5`.
Both treated the innermost *formatted traceback filename* as origin authority.
In actual subprocess probes on CPython 3.12.14, an unchanged subject passed,
while an `AssertionError` raised by the subject itself under a directory named
`tests` was incorrectly accepted as a kill by each grader in each Python mode.
`red-results-*.json`, `*-subject-*.stderr` and `initial-probe.py.txt` preserve
those observations and source hashes. `raw-failure-logs.zip` preserves the exact `red-controls-*.log` red
adverse tests and hashes. These failures affect the graders' mutation-credit soundness;
they do not refute a mathematical statement or erase earlier executions.

The repair uses the existing `tests/cbz_inventory_origin.py` recorder, which
captures actual traceback code identities before unittest cleans the exception.
A selected synchronous test-body failure is a kill only if all later frames
are explicitly authorized standard unittest assertion helpers. A clean
selected-test baseline is required for each actual source mutant. Subject
bytes are hashed before and after the run. Missing or inconsistent provenance,
subject/helper/nested callback exceptions, unexpected skips, errors and source
drift are non-kills. The recorder is for trusted synchronous unittest tests; it
is not a hostile-code sandbox, a Lean kernel, a proof checker or a federation
certificate. Existing historical mutation outcomes are not rebound.

The initial adapter self-mutation attempt failed in both modes because the
copied adapter inferred the temporary directory as its test checkout; full
failed logs are retained inside `raw-failure-logs.zip`; the explicit
limitations note is in `attempts.log`. The child then explicitly binds the copied helper to the
checkout owning the designated tests. `full-evidence-red.log` preserves the
subsequent red control for missing full-suite evidence. One partial-state log
was accidentally overwritten; its known exit and that loss are recorded in
`attempts.log`, not reconstructed.

The separately recorded pre-repair lexical scan returned exit 1 on Main's
intentional, unapproved kernel-replay axiom fixture, and reported the replay
adapter's `unsafe` token for review; `main-lexical.json` is unchanged. The
Math scan at `9bdfc54b76173b57912503bfcf901a5f9d43b09c` returned exit 0
for 1,974 Lean blobs with no listed tokens; `math-lexical.json` is the raw
report. These scans alone cannot establish axiom closure or scientific
alignment. Workspace candidate `e9b6ef598ae249120be0d7498c6d5adcc95a1cab`
returned exit 2 on its gitlink; it was not silently skipped. The Math Python
census returned exit 1 on the preserved historical documentary writer
`frontiers/source_derivative_gaussian_20261025/actual-source-finite-superlevel-components-evidence/raw/producer/history/freeze_documents.001.py`
(blob `f588617889ed04eb2c76bb1ab428fcfa5217ac4b`), whose original parser
failure and later correction are preserved in that packet. Python AST float
signals are a census, not a verdict that a bound is or is not certified.

The first synthetic workflow-guard control failed because its stub expected
21 tests while the candidate shell still expected 20; that red output remains
in `raw-failure-logs.zip` as `workflow-guards-red-normal.log`, and the workflow now includes the guard
control in its group (now 22 tests). Its seven guard removals use synthetic child
outcomes and do not claim real scanner or mathematical execution.

A later actual shell-guard probe showed that an unwritable exit-status file
could still produce exit 0 even though the write failed. The two raw
synthetic-producer failures are retained unchanged in `exit-metadata-red.zip`.
The candidate now fails if either exit-status evidence write fails.

The candidate workflow runs the explicit controls and mutation suites in
both Python modes, retains selected child evidence and full-suite logs even
on failure, and checks source hashes. Its hosted results must be observed on
the exact head before this repair can receive current source credit. The
legacy full-suite baseline intentionally records two Lean and four A3 inner
recursion-guard skips; the exact expected skip identities are checked and
retained, while all 101 selected mutation baselines per mode must have no skips.
This is not the final G0/G1 or a full Layer A audit.
20 changes: 20 additions & 0 deletions cbz/evidence/auditor-origin-20261109/a3-subject-normal.stderr
Original file line number Diff line number Diff line change
@@ -0,0 +1,20 @@
test_float_and_complex_calls (tests.test_cbz_a3_float_census.Signals.test_float_and_complex_calls) ... FAIL

======================================================================
FAIL: test_float_and_complex_calls (tests.test_cbz_a3_float_census.Signals.test_float_and_complex_calls)
----------------------------------------------------------------------
Traceback (most recent call last):
File "/workspace/scratch/7abea556a9be/federation-vector-post-integration-20261109/main/tests/test_cbz_a3_float_census.py", line 48, in test_float_and_complex_calls
self.assertEqual(kinds("a = float(x)\nb = complex(1, 2)\n"),
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/scratch/7abea556a9be/federation-vector-post-integration-20261109/main/tests/test_cbz_a3_float_census.py", line 34, in kinds
r = M.census_text(src, "x.py")
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/scratch/7abea556a9be/auditor-origin-probe-20261109/tests/a3-subject_assertion-normal.py", line 311, in census_text
raise AssertionError("ACTUAL_SUBJECT_ORIGIN_PROBE")
AssertionError: ACTUAL_SUBJECT_ORIGIN_PROBE

----------------------------------------------------------------------
Ran 1 test in 0.000s

FAILED (failures=1)
Loading
Loading