52 lines
3.1 KiB
Markdown
52 lines
3.1 KiB
Markdown
# Independent integrated verdict — run-d94d4516b9d898793318805e
|
|
|
|
## Decision
|
|
|
|
**Scoped PASS** for the current two-record integrated evidence package. This establishes independent
|
|
static reproduction and archived-state inspection only. It is not original-assisted runtime,
|
|
partial/full live comparison, independent replacement, or integrated replay, and makes no live
|
|
allocator-safety claim.
|
|
|
|
## Reproduction and challenge
|
|
|
|
From canonical RE:
|
|
|
|
```sh
|
|
python3 -m py_compile verify/results/research-completion-abi-independent/reproduce_run_d94d4516b9d898793318805e.py
|
|
python3 verify/results/research-completion-abi-independent/reproduce_run_d94d4516b9d898793318805e.py
|
|
python3 -m py_compile verify/results/research-completion-abi-independent/heldout_run_d94d4516b9d898793318805e.py
|
|
python3 verify/results/research-completion-abi-independent/heldout_run_d94d4516b9d898793318805e.py
|
|
python3 -m unittest discover -s verify/campaign -p 'test_*.py' -v
|
|
python3 tools/campaign.py --state-root /home/alex/sots-re validate
|
|
```
|
|
|
|
All commands exited zero. The reproduction executed 22 nonempty static windows and two strict state
|
|
tools, with no stderr or skips. Its manifest equals the integrated manifest except for session, and
|
|
its independent-state artifact is byte-identical. Seven narrow stops expose truncation while seven
|
|
widened stops expose complete `ret 4`/`ret 8` instructions. Direct PE reads agree.
|
|
|
|
The held-out ablation independently read each return at exact, minus-one and plus-one starts. All
|
|
seven exact encodings matched and all fourteen shifted controls differed. Static branch exposures
|
|
confirm separate `0x2c`/`0x74` strides, both deep-copy calls, three guarded string-delete calls,
|
|
description inequality participation, three float unordered-or-unequal rejection sequences, and
|
|
event iteration. This challenges boundary selection rather than repeating reported counters.
|
|
|
|
Strict actual-state inspection found EvNxID 4, turn-3 bucket count 2, event ID 3 text/action/location/
|
|
chain and `0x7f7fffff` position words. The checksum rebuilt all 609080 inflated bytes with no first
|
|
difference; the actual `/Sim/RNG` leaf is 2503 bytes with digest
|
|
`0978fdf34ff7962f76c2de810dc93e0a`.
|
|
|
|
## Identity, failures, and residuals
|
|
|
|
Assigned HEADs match the baseline and their Git common directories are canonical. Canonical
|
|
integrated source binding remains engine `ccd8e020...50fd`, RE `e74ef6cf...bd4`, exactly matching
|
|
both evidence records. The assigned paired RE digest `6696fd52...72e8` remains distinct as expected
|
|
from excluded campaign/result material and was not substituted for integrated identity. Binary,
|
|
save, objdump, evidence, input and outcome identities match; campaign validation passes.
|
|
|
|
No new prediction failed. Historical truncated-capture provenance and description-omission
|
|
predictions remain recorded failures resolved by Astra; matching repaired evidence does not erase
|
|
them. Unexecuted residual states remain empty/full capacity, short/long strings, duplicate versus
|
|
description-different events, NaN runtime behavior, allocation failure/unwind, allocator-family
|
|
compatibility, live original differential and replay. Whole-state equality is scoped to the single
|
|
archived turn-3 save.
|