122 lines
7.9 KiB
Markdown
122 lines
7.9 KiB
Markdown
# Controls formal verifier checkpoint
|
|
|
|
## Fresh verifier quantum plan — run-cb15199f9272fe496bd10a8a
|
|
|
|
Actor `controls-independent-verifier`; role `verifier`; requested model
|
|
`openai/gpt-5.6-terra`; session `run-cb15199f9272fe496bd10a8a`. This is an
|
|
independent re-execution, scoped only to `controls-bootstrap` /
|
|
`controls-negative-paths`; it is neither original-assisted validation nor a
|
|
partial/full engine comparison, independent replacement, or integrated replay.
|
|
|
|
Falsifiers defined before execution: (1) the current canonical source binding,
|
|
declared executable/input/result hashes, contract basis, or no-open-surprise
|
|
state differs from the handoff; (2) the complete suite has any failure, error,
|
|
skip, or fewer than the declared 37 distinct tests; (3) old intact recovery,
|
|
stale end checkpoint, and artifact-tampered recovery do not separate as
|
|
required; (4) same-HEAD candidate or canonical-integrated byte mutation is not
|
|
rejected before verdict/promotion; (5) the added verifier integration route
|
|
allows owner overlap, source drift, or an open surprise; (6) the actual launch
|
|
record lacks a real session, successful `stop`, matching fresh checkpoint,
|
|
zero return, or source-before/source-after equality. Required branch/state
|
|
exposures are valid-old/tampered/stale recovery; candidate/canonical-integrated
|
|
same-HEAD mutation; verifier verification/integration entry versus owner,
|
|
source-drift, and surprise negative controls; and real launch record versus
|
|
fake-process unit simulation. No RNG/stateful game workload is in contract, so
|
|
RNG accounting is inapplicable rather than assumed satisfied.
|
|
|
|
Pre-execution observations: all three recorded surprises are resolved;
|
|
canonical and paired-worktree HEADs equal the pinned commits, while canonical
|
|
trees are intentionally dirty and paired launch trees are clean. Candidate
|
|
source binding is distinct from the canonical evidence binding and will not be
|
|
substituted. The R8 correction invalidated the prior formal verdict; its new
|
|
evidence claims 37 tests. Next: checkpoint this plan, then independently
|
|
rehash/revalidate and run the full suite plus held-out negative controls.
|
|
|
|
## Fresh reproduction results — run-cb15199f9272fe496bd10a8a
|
|
|
|
Canonical `source-binding controls-bootstrap` produced the attached evidence's
|
|
engine digest `ccd8e02083e8d2e2b3e97976ace2273c8f924dfc02a39e919004eaf3544c50fd`
|
|
and RE digest `6696fd5201e144843617cbf6d78b41b5287ad5dcc9fa1e8aaa861d52b64e72e8`.
|
|
`campaign.py validate controls-bootstrap` passed. Independent byte hashes of
|
|
the two declared binaries, three declared inputs, and outcome package all
|
|
exactly match the contract. The package has 37 unique expected and 37 unique
|
|
passed method names, with no missing or extra names, return code zero and
|
|
`status: passed`.
|
|
|
|
`PYTHONDONTWRITEBYTECODE=1 python3 -m unittest discover -s verify/campaign
|
|
-p 'test_*.py' -v` reproduced **37/37**, zero skips/errors/failures, in 7.049
|
|
seconds. The independent held-out negative-control rerun of old recovery,
|
|
same-HEAD verdict/promotion drift, canonical integrated RE mutation, and final
|
|
integrated-verifier guards reproduced **4/4** in 1.093 seconds. Thus every
|
|
defined control branch exposed by these tests behaved as predicted; no failed
|
|
prediction was observed.
|
|
|
|
Independent parsing of the actual non-fake run record found 24/24 JSONL events,
|
|
zero error events, one completed/observed actual session, return code zero, a
|
|
`step_finish` reason `stop`, source-before equal source-after, and the matching
|
|
checkpoint session. Its observed model remains `unavailable`; it is an actual
|
|
Astra architecture-review launch, **not** a direct Terra normal-worker
|
|
execution. This is a residual/provenance limitation, not evidence of a Terra
|
|
execution, but does not falsify the narrowly worded unit-test criterion.
|
|
|
|
Decision: record a fresh scoped passing verifier verdict for the exact current
|
|
nonintegrated evidence only. Do not claim integrated acceptance, whole-engine
|
|
behavior, original comparison, replacement, replay, or game RNG/state
|
|
coverage. Exact next action: lead may attach a source-bound integrated package,
|
|
after which a fresh independent verdict must bind its changed evidence digest.
|
|
|
|
Actor `controls-independent-verifier`; role `verifier`; model
|
|
`openai/gpt-5.6-terra`; session `controls-independent-20260909-formal-1`.
|
|
|
|
## Scope and pre-execution falsifiers
|
|
|
|
This is a source-bound verdict only for `controls-bootstrap` criterion
|
|
`controls-negative-paths`, not engine, assets, research replacement, or an integrated package.
|
|
The verdict would fail if the attached artifact/binary/input hashes or either canonical source
|
|
manifest differed; if the suite had a failure, skip, or fewer than 36 tests; if old recovery
|
|
accepted corrupted artifacts or rejected solely for age; if a same-HEAD byte change did not
|
|
invalidate verdict/promotion; or if the purported normal launch lacked a real session, stop,
|
|
fresh matching checkpoint, zero return, or unchanged paired-worktree state. Required distinct
|
|
states were intact-old versus stale-end versus tampered recovery, candidate-engine versus
|
|
integrated-RE same-HEAD mutation, and real normal launch versus fake-process unit simulation.
|
|
|
|
## Independent observations and reproduction
|
|
|
|
Read `AGENTS.md`, `campaign/README.md`, contract
|
|
`campaign/contracts/controls-bootstrap.json`, current checkpoint
|
|
`campaign/runtime/checkpoints/controls-bootstrap-243f539b3219e74f12ef0db7.json`, raw evidence,
|
|
raw run JSON/JSONL/stderr, and Astra review `campaign/rollout/independent-review.md`.
|
|
|
|
`python3 tools/campaign.py --state-root /home/alex/sots-re source-binding controls-bootstrap`
|
|
recomputed engine `7741d42fc5e4e761e6449bdaf0e4a61d00036a23` digest
|
|
`ccd8e02083e8d2e2b3e97976ace2273c8f924dfc02a39e919004eaf3544c50fd` and RE
|
|
`3bfde5a70d874a723e797a695bbd847fd82c0aa7` digest
|
|
`82fc34631b2baa2ee74557020f74d68cc4db3c031561210f9933a3f22d7ff7b6`, exactly matching
|
|
the attached evidence. Independently rehashed all declared binaries, inputs, and result artifact;
|
|
each matched its declared SHA-256. `campaign.py validate controls-bootstrap` passed.
|
|
|
|
`PYTHONDONTWRITEBYTECODE=1 python3 -m unittest discover -s verify/campaign -p 'test_*.py' -v`
|
|
ran **36 tests**, all passed, zero skips, in 6.720 s. Held-out focused rerun:
|
|
`python3 -m unittest -v verify.campaign.test_controls.Controls.test_old_recovery_checks_integrity_but_not_age verify.campaign.test_controls.Controls.test_same_head_source_mutation_rejects_verdict_and_promotion verify.campaign.test_controls.Controls.test_integrated_re_source_mutation_rejects_acceptance`
|
|
ran 3/3 pass (0.722 s). The inspected tests require an old intact checkpoint to launch but reject
|
|
its stale end-checkpoint use and a post-checkpoint artifact modification; they alter candidate
|
|
engine and canonical RE bytes without changing HEAD and require source-content rejection before
|
|
verdict/integration/acceptance.
|
|
|
|
The actual, non-fake normal-launch record
|
|
`campaign/runtime/runs/run-df1472c13f31db3a4d5361f0.json` hashes to
|
|
`65dd1c1286742cd4d2401af5a229d7d014f60dd329dadee0dee78c499f61d663`.
|
|
Its JSONL has 24 parseable events, exactly one actual session
|
|
`ses_f77cf5062ffeTknrMfnbJMmdNh`, a `step_finish` with reason `stop`, return code 0, zero error
|
|
events, matching launcher checkpoint/session, and `source_before == source_after` for both paired
|
|
worktrees. Raw stderr is empty. Observed model is correctly `unavailable`, not fabricated; this
|
|
is an actual Astra architecture-review smoke, not evidence of a Terra worker launch.
|
|
|
|
## Decision and residual
|
|
|
|
Pass is justified only for the one declared controls test criterion and the exact nonintegrated
|
|
binding above. Astra R8 remains a separately recorded residual: the current launcher permits a
|
|
verifier in `verification` but not `integration`; it prevents the later fresh integrated-verifier
|
|
launch and requires Astra resolution. It neither converts this reproduced nonintegrated suite into
|
|
whole-engine acceptance nor is hidden as success. Exact next action: lead resolves R8, attaches
|
|
an integrated controls package, then obtains a new independent verdict over its changed digest.
|