Skip to content

Commit df9a433

Browse files
committed
P-037 Phase B / B1: R_B retake at the corrected instrument (T_B = a02ccf1)
Evidence-only. Supersedes 40cb9b4's R_B (measured against the mislabeled, wrong-epoch instrument_identity, before B1-F1 hardening) -- kept exactly as committed in git history, not rewritten. This retake measures against the corrected instrument (formal-kernel.md #10.8b-10.8d): the classifier precedence fix, the production-diff-gate self-authorization fix, and the instrument_identity wrong-epoch-closure fix plus its own mislabeled-value correction. Four governed takes at T_B (a02ccf1, population_commit identical), each independently verified before the next started, none concurrent: MOS repo snapshot (82 files), MOS corpus snapshot (192 files), verdict/python and verdict/rust snapshots (192 corpus files each). All four agree on source_commit, population_commit, epoch=b, environment_id=P037_B_MEASUREMENT_M3, dirty=false, is_evidence=true, and instrument_identity (75764c81ac49185adad6a02657b373a7453fc03ba337bb61b86d4694bd8d5065) -- byte-for-byte, cross-checked directly, and confirmed different from the first R_B's mislabeled value, as expected. Facts and verdicts are unchanged from the first R_B, measured rather than assumed: both MOS snapshots omit cross_engine_mismatches entirely (the tool only writes that key when non-empty; confirmed by reading p037_mos_snapshot.py's own source, not by the key's absence alone), and console output recorded zero for both (82 and 192 documents); the two verdict snapshots were diffed directly, 192/192 files identical (line, code, level) and exit code on both engines, 105 findings each side, 0 mismatches -- identical to the first R_B's own numbers. p037_controls.py --engine both and p037_fact_shapes.py check --engine both, re-run fresh at the new T_B, both report full agreement (all match / 55/55 shapes), also identical to the first R_B. The mos-corpus take's first attempt hit its own 115s shell timeout and was killed; the resulting stale population lock (held by a PID confirmed no longer running) was removed before the take was retried in full -- an operational hiccup in this session's tooling, not a second concurrent take and not a defect in the measurement. Adds nothing new: updates the same five docs/evidence/ paths the first R_B created. No instrument, treatment, spec, tooling, or CI file is touched. named_later.T_B/R_B in docs/evidence/p037-b-epoch.json are not updated by this commit -- that is the naming commit's own act, once this evidence is itself terminal-green, mirroring 40cb9b4 before 2ed4d92. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AN6xHbpovxrmZ4AMjS7WQA
1 parent a02ccf1 commit df9a433

5 files changed

Lines changed: 861 additions & 822 deletions

‎docs/evidence/p037-b-baseline-manifest.md‎

Lines changed: 109 additions & 70 deletions
Original file line numberDiff line numberDiff line change
@@ -2,9 +2,9 @@
22

33
epoch: `b`
44

5-
T_B (terminal-green, the measurement-instrument head): `349c7bcf456bdd33ce3efe7a19da43790c186145`
5+
T_B (terminal-green, the measurement-instrument head): `a02ccf17d22eaf62c046fb95ba5060e6d66ffeae`
66

7-
population_commit (every take): `349c7bcf456bdd33ce3efe7a19da43790c186145`
7+
population_commit (every take): `a02ccf17d22eaf62c046fb95ba5060e6d66ffeae`
88

99
environment: `P037_B_MEASUREMENT_M3`
1010

@@ -16,19 +16,46 @@ predecessors only and are not retaken, compared, or treated as a control
1616
here (`docs/evidence/p037-b-epoch.json`'s own `predecessor`/`environment`
1717
sections state this precisely).
1818

19+
## This is a retake, not the first R_B
20+
21+
The instrument this take measures against is not the one `349c7bc`/`40cb9b4`
22+
measured against. Between that take and this one, B1-F1 instrument
23+
hardening fixed four defects in the instrument itself (formal-kernel.md
24+
§§10.8b-10.8d): two found by owner review (a classifier precedence bug
25+
that could silently resolve an overlap the frozen contract requires
26+
`UNCLASSIFIED`; a production-diff-gate self-authorization hole that read
27+
its own policy boundary from the head it was validating), and two
28+
self-found (`instrument_identity` was computed by calling
29+
`p037_evidence.py`'s own epoch-coupled function directly, so it measured
30+
a2d's eight-path closure instead of Phase B's own nine-path closure; then,
31+
once fixed, the fix commit's own "consequence" prose mislabeled which
32+
commit its "all three fixes" digest actually belonged to, because the fix
33+
commit itself edits an in-scope instrument file and moves the identity a
34+
second time). None of the four touched `mos.rs`/`lower.rs` or any
35+
production semantics; all four are documented in
36+
`docs/notes/p037-formal-kernel.md`. `349c7bc`, `40cb9b4` and `2ed4d92` are
37+
kept exactly as committed -- honest records of the instrument and its
38+
first (flawed) `R_B` -- superseded here, not rewritten: `git log` on this
39+
file's path still shows their content.
40+
1941
## Instrument identity
2042

2143
```
22-
04525002af5383fcdda0c4f0d77c9b647cc2c5f6be759b4ec733067013e8a559
44+
75764c81ac49185adad6a02657b373a7453fc03ba337bb61b86d4694bd8d5065
2345
```
2446

2547
Read from each take's own `instrument_identity` field (computed by
26-
`p037_evidence_b.evidence_fields()` from `p037_evidence.instrument_manifest`/
27-
`_manifest_digest` over Phase B's own `INSTRUMENT_PATHS`/`INSTRUMENT_
28-
CARVE_OUTS`) -- not a separate CLI call: unlike `p037_evidence.py identity`,
29-
`p037_evidence_b.py`'s own `identity` subcommand does not print this digest
30-
directly, so the authoritative value is the one each piece of evidence
31-
actually carries. All four takes below agree on it, byte-for-byte.
48+
`p037_evidence_b.evidence_fields()`, now calling this module's OWN
49+
correctly-scoped `instrument_manifest`/`instrument_identity` over Phase
50+
B's own `INSTRUMENT_PATHS`/`INSTRUMENT_CARVE_OUTS` -- the §10.8c fix) --
51+
not a separate CLI call: unlike `p037_evidence.py identity`,
52+
`p037_evidence_b.py`'s own `identity` subcommand does not print this
53+
digest directly, so the authoritative value is the one each piece of
54+
evidence actually carries. All four takes below agree on it,
55+
byte-for-byte. This differs from the value the first `R_B` recorded
56+
(`04525002af5383fcdda0c4f0d77c9b647cc2c5f6be759b4ec733067013e8a559`) --
57+
expected and confirmed by direct measurement, not assumed: that value was
58+
a2d's instrument identity, mislabeled as Phase B's own (§10.8c).
3259

3360
## Execution profile (M3, identical across all four takes)
3461

@@ -37,8 +64,12 @@ actually carries. All four takes below agree on it, byte-for-byte.
3764
- Rust: rustc 1.94.1 (e408947bf 2026-03-25), cargo 1.94.1 (29ea6fb6a 2026-03-24), host x86_64-unknown-linux-gnu
3865
- Platform: Linux / x86_64
3966

40-
Captured verbatim via `python3 scripts/p037_evidence_b.py profile` immediately
41-
before the first take, and re-captured after the fourth take:
67+
Captured verbatim via `python3 scripts/p037_evidence_b.py profile` after
68+
the fourth take (this retake reuses the same qualified M3 environment the
69+
first `R_B` established a few hours earlier in the same session; only the
70+
instrument changed, not the toolchain, so a fresh before/after pair was
71+
not repeated -- the one capture below, cross-checked against the first
72+
`R_B`'s own recorded profile, is the evidence for that):
4273

4374
```json
4475
{
@@ -65,112 +96,120 @@ before the first take, and re-captured after the fourth take:
6596
}
6697
```
6798

68-
Byte-for-byte identical to `docs/evidence/p037-a2d-baseline-mos-repo.json`'s
69-
own recorded M2 profile -- expected and stated as such in
70-
`docs/evidence/p037-b-epoch.json`'s `environment.qualification`: M3 is a
71-
genuine requalification of M2's exact toolchain, on the same physical
72-
workspace, not an independent machine claim.
99+
Byte-for-byte identical to the first `R_B`'s own recorded profile (and, in
100+
turn, to `docs/evidence/p037-a2d-baseline-mos-repo.json`'s recorded M2
101+
profile) -- the same physical workspace and toolchain, re-verified rather
102+
than assumed unchanged.
73103

74104
## Machine qualification
75105

76-
Measurement environment `P037_B_MEASUREMENT_M3`, requalified rather than
77-
reprovisioned:
106+
Measurement environment `P037_B_MEASUREMENT_M3`, unchanged since B1's own
107+
qualification (this retake re-measures the instrument, not the
108+
environment):
78109

79-
- CPython: pinned to the host interpreter (3.11.15) -- reverified byte-for-
80-
byte against M2's own recorded profile.
81-
- .NET SDK: 8.0.425, installed isolated outside any OS package manager --
82-
same reverification.
83-
- Rust: rustc/cargo 1.94.1, exact commit hashes equal to M2's recorded
84-
profile -- same reverification.
110+
- CPython: pinned to the host interpreter (3.11.15).
111+
- .NET SDK: 8.0.425, installed isolated outside any OS package manager.
112+
- Rust: rustc/cargo 1.94.1.
85113

86-
recipe sha256 (identical to M2's, independently re-hashed at B1 time, not
87-
copied from the old manifest):
114+
recipe sha256 (re-hashed at this retake, not copied from the earlier
115+
manifest -- confirmed unchanged):
88116

89117
```
90118
5fe4ece4c24b720f659a49bbc68b8438080c9e4b8ba554ee40f6eea94c66eb26
91119
```
92120

93121
Recipe: `provision.sh`, retained outside this repository at
94-
`/root/p037-a2d-m2-recipe/provision.sh` on the qualified machine. M3 reuses
95-
M2's exact recipe because it independently requalified byte-for-byte, per
96-
`docs/evidence/p037-b-epoch.json`'s own `environment.m2_rule`: M2 takes
97-
nothing for this epoch and is not a control of M3; this is a new epoch's own
98-
environment record, never a silent inheritance. The environment/toolchain
99-
was not changed after the first governed take began.
122+
`/root/p037-a2d-m2-recipe/provision.sh` on the qualified machine. The
123+
environment/toolchain was not changed between the first `R_B` and this
124+
retake.
100125

101126
## Workspace
102127

103-
The same preserved measured workspace checkout root M2 used, reused here
104-
because the toolchain on it was independently reverified, not assumed:
128+
The same preserved measured workspace checkout root used throughout this
129+
epoch:
105130

106131
```
107132
/home/user/Own.NET
108133
```
109134

110135
## Takes
111136

112-
Four sequential governed takes, each independently `verify`d against T_B
113-
before the next was started. None ran concurrently.
137+
Four sequential governed takes, each independently `verify`d against the
138+
new T_B before the next was started. None ran concurrently. (The MOS
139+
corpus take's first attempt hit its own timeout and was killed; the
140+
resulting stale population lock, held by a PID confirmed no longer
141+
running, was removed before the take was retried in full -- an
142+
operational hiccup in this session's tooling, not a defect in the
143+
measurement, and not a second concurrent take.)
114144

115145
1. MOS repo snapshot (`--epoch b --source repo`)
116146
2. MOS corpus snapshot (`--epoch b --source corpus`)
117147
3. verdict/python snapshot (`--epoch b --engine python`)
118148
4. verdict/rust snapshot (`--epoch b --engine rust`)
119149

120150
All four takes and verifies exited 0. The checkout was clean at
121-
`349c7bcf456bdd33ce3efe7a19da43790c186145` before the first take and remained
122-
at that exact commit, clean, after the fourth verify.
151+
`a02ccf17d22eaf62c046fb95ba5060e6d66ffeae` before the first take and
152+
remained at that exact commit, clean, after the fourth verify.
123153

124-
| p037-b-baseline-mos-repo.json | `3abb0a25ca0c60efc2882d5d8695517e65ce53398370f7d97d6366dea5ac0be7` |
125-
| p037-b-baseline-mos-corpus.json | `ebbb84d9e45a761eec5260dad7143f8b625b4f68b832f97d295d5608ef06f2f0` |
126-
| p037-b-baseline-verdict-python.json | `5668d59f9df041841c7d40a26b5fe6c56b7e0bc385297d82b84e517657143a43` |
127-
| p037-b-baseline-verdict-rust.json | `60ea9b11b060045f632e172971c1ffae511052e614f216db7db0a013cdd84d6a` |
154+
| p037-b-baseline-mos-repo.json | `8d33712811f32e75c7d26dceab03a5a7829cf1f39c0c8139b6c647f46781313a` |
155+
| p037-b-baseline-mos-corpus.json | `a26f8801869831afe4e06e15d449b5c4090c56312d4ceda47014b40a09532342` |
156+
| p037-b-baseline-verdict-python.json | `5d7f9490abdbb2a7ca9faf855a9c35b93f3779351a963b6a0128feff4658d95b` |
157+
| p037-b-baseline-verdict-rust.json | `f653d812679363aebea6b7015f1126c7647f75b608c2cf2ad1e57d34f129aa4a` |
128158

129159
Every record above carries, and all four agree on:
130160

131-
- `source_commit`: `349c7bcf456bdd33ce3efe7a19da43790c186145`
132-
- `population_commit`: `349c7bcf456bdd33ce3efe7a19da43790c186145`
161+
- `source_commit`: `a02ccf17d22eaf62c046fb95ba5060e6d66ffeae`
162+
- `population_commit`: `a02ccf17d22eaf62c046fb95ba5060e6d66ffeae`
133163
- `epoch`: `b`
134164
- `environment_id`: `P037_B_MEASUREMENT_M3`
135165
- `dirty`: `false`
136166
- `is_evidence`: `true`
137-
- `instrument_identity`: `04525002af5383fcdda0c4f0d77c9b647cc2c5f6be759b4ec733067013e8a559`
167+
- `instrument_identity`: `75764c81ac49185adad6a02657b373a7453fc03ba337bb61b86d4694bd8d5065`
138168

139169
## Cross-engine parity at baseline
140170

141-
The MOS snapshots (repo and corpus) each record their own `cross_engine_
142-
mismatches` field internally: empty on both, over 82 and 192 documents
143-
respectively -- Python and Rust MOS agree completely at T_B. The verdict
144-
snapshots do not have a built-in cross-engine comparator by design
145-
(`p037_verdict_snapshot.py compare` explicitly refuses a cross-engine pair:
146-
"measures the engine, not the change"), so parity was checked directly: all
147-
192 corpus files carry identical `(line, code, level)` finding sets and
148-
identical exit codes on both engines -- 105 findings each side, 0
149-
mismatches. This is the empirical form of the frozen `facts_expectation`
171+
`scripts/p037_mos_snapshot.py` only writes a `cross_engine_mismatches`
172+
key at all when that list is non-empty (`if parity_moved:` -- confirmed
173+
by reading the tool's own source, not assumed from its absence); neither
174+
MOS snapshot below carries the key, and both takes' own console output
175+
recorded `cross-engine mismatches=0` explicitly (82 and 192 documents
176+
respectively) -- reconfirmed directly by comparing the `python-ownlang`/
177+
`rust-own-bridge` recorded documents for equality, not merely by the
178+
key's absence. Python and Rust MOS agree completely at the new T_B. The
179+
verdict snapshots do not have a built-in cross-engine comparator by
180+
design (`p037_verdict_snapshot.py compare` explicitly refuses a
181+
cross-engine pair: "measures the engine, not the change"), so parity was
182+
checked directly: all 192 corpus files carry identical `(line, code,
183+
level)` finding sets and identical exit codes on both engines -- 105
184+
findings each side, 0 mismatches, exactly matching the first `R_B`'s own
185+
numbers. This is the empirical form of the frozen `facts_expectation`
150186
claim (`docs/evidence/p037-b-epoch.json`'s `measurement_policy`): nothing
151-
about the pre-treatment facts or verdict surface moved.
187+
about the pre-treatment facts or verdict surface moved, on the corrected
188+
instrument any more than on the mislabeled one.
152189

153190
## P-037 conformance baseline at T_B
154191

155-
Both re-run fresh at `349c7bcf456bdd33ce3efe7a19da43790c186145`, `OWEN_RUST_
156-
CORE` pointed at the qualified `own-cli` release build:
192+
Both re-run fresh at `a02ccf17d22eaf62c046fb95ba5060e6d66ffeae`, `OWEN_
193+
RUST_CORE` pointed at the qualified `own-cli` release build:
157194

158195
- `python scripts/p037_controls.py --engine both`: `RESULT: all match` (the
159196
three G-V4 known-false-positive controls and the one legacy-honesty
160197
control all agree between engines and against the recorded record).
161198
- `python scripts/p037_fact_shapes.py check --engine both`: `RESULT: all 55
162199
shape(s) match their recorded facts and verdicts`.
163200

164-
Recorded here as the P-037 conformance state at T_B, per the B1 brief's own
165-
instruction to measure and record it, not to chase or repair it as part of
166-
this baseline.
167-
168-
Run time/date: 2026-09-23, ~19:50-20:35 UTC, operator run (the mos-repo and
169-
mos-corpus takes were first measured at the interim head `d5f9134`, then
170-
discarded and re-measured in full at `349c7bc` once that head's own
171-
allowlist-completeness fix was confirmed terminal-green; the times above are
172-
for the accepted run).
173-
174-
This commit is evidence-only: it adds exactly five files under
175-
`docs/evidence/` (these four JSON artifacts plus this manifest) and nothing
176-
else. No instrument, treatment, spec, tooling, or CI file is touched.
201+
Identical results to the first `R_B`'s own recorded conformance state.
202+
Recorded here as the P-037 conformance state at the new T_B, per the B1
203+
brief's own instruction to measure and record it, not to chase or repair
204+
it as part of this baseline.
205+
206+
Run time/date: 2026-09-24, ~00:59-01:35 UTC, operator run.
207+
208+
This commit is evidence-only: it updates exactly five files under
209+
`docs/evidence/` (these four JSON artifacts plus this manifest, all
210+
already tracked from the first `R_B`) and nothing else. No instrument,
211+
treatment, spec, tooling, or CI file is touched. `docs/evidence/p037-b-
212+
epoch.json`'s `named_later.T_B`/`R_B` are not updated by this commit --
213+
that is the naming commit's own, separate act, after this evidence is
214+
itself terminal-green, mirroring exactly how `40cb9b4` preceded `2ed4d92`
215+
the first time.

0 commit comments

Comments
 (0)