Competitive execution-selection r3 candidate (2026-08)¶
This document describes an additive handoff candidate. It does not replace the active profile or the historical selection receipt.
Status¶
The checked-in candidate is NOT_READY, benchmark_eligible: false, and
claim_eligible: false. Its current structural validation is:
The candidate binds the current checkout bytes for every runner (ours, GLIM,
and FAST-LIVO2), the GT-blind driver, six scorer/result producers, and the
three process-RSS producer/gate files. These are source identities only; they
are not execution receipts. It also binds the exact seven-field CPU thread
policy (cpu_affinity, max_threads, OMP_NUM_THREADS,
OPENBLAS_NUM_THREADS, MKL_NUM_THREADS, TBB_NUM_THREADS, and
accelerator_policy) with canonical identity
262d656a7c382cd27696b3150215ae563a014796de69718754858bcb814ba993.
The selection lineage is explicit:
- current recipe selection:
competitive_execution_selection_2026-08-r2.yaml, selection idcompetitive-execution-selection-2026-08-r2, closure idcompetitive-rival-source-closure-2026-08-r2, closure identitye876339f843b28c1be5d835e8f6a9a062c52715498c85d93afbdb83c6dbae4e8; - historical
competitive_execution_selection_2026-08.yamlremains in place and is recorded as rejected/superseded under the r2 closure revision; - the active profile still points to the historical path. No profile switch or hash update is performed by this candidate.
The candidate reports, rather than infers away, the remaining blockers:
- rival-source legal provenance is
NOT_READYand image publication is blocked; - dataset source closure remains
NOT_READYwith no reviewed pin manifest; - fresh-holdout authorization lacks the independent custodian attestation;
- machine identity, image inspection, and container toolchain evidence are external execution artifacts and are unresolved here; and
- the shared checkout has pending changes, so source bytes require external review before promotion.
The dataset and holdout section identities are recorded from the current profile, but no dataset, bag, ground-truth, or scorer content is opened. Machine fields are intentionally null; no machine id, container digest, or license identity is invented.
Unsigned handoff¶
An unsigned request can be prepared without a signer:
PYTHONDONTWRITEBYTECODE=1 python3 scripts/prepare_competitive_execution_selection_r3_handoff.py \
prepare --output /tmp/competitive-execution-selection-r3-handoff.json
PYTHONDONTWRITEBYTECODE=1 python3 scripts/prepare_competitive_execution_selection_r3_handoff.py \
validate --handoff /tmp/competitive-execution-selection-r3-handoff.json
The request is an atomic JSON/sidecar pair. The two output files are created
exclusively, fsynced, and sealed read-only (0444), with single-link regular
file and device/inode checks on every reopen. Existing output/sidecar paths,
hard links, symlinked parents, path traversal, writable files, partial pairs,
and changes during validation are fail-closed; a failed second write removes
only the inode owned by that write. With no external manifest it lists all seven
required artifact classes as missing. A separately reviewed external manifest
may provide exact regular-file SHA-256 bindings for the rival closure, dataset
closure, holdout authorization, machine, image inspection, toolchain, and
execution receipt. The manifest and each referenced artifact are likewise
bounded, read-only (0444), single-link files. The handoff binds the manifest
root and file device/inode/mode/size as well as its bytes, and reopens those
identities before accepting coverage. The manifest is defined by
competitive_execution_selection_r3_external_artifacts_v1.schema.json.
Coverage is exact; duplicate, extra, symlinked,
hard-linked, stale, or cross-selection artifacts are rejected. The handoff
contains no private key, signer, signature, or promotion result.
The checked-in candidate sidecar is a source-control binding and may retain
normal repository permissions. The stricter 0444 requirement applies to
fresh handoff and external evidence artifacts, not to that checked-in source
sidecar.
Promotion is a separate custodian workflow requiring a reviewed READY r3
execution receipt and a canonical profile reseal. The workflow is atomic at
the policy boundary: historical receipts remain immutable, and this candidate
cannot switch the active profile. A signed or self-rehashed JSON file is not
treated as a receipt merely because its status says READY.
The additive r3 external capture
utility can observe the three local image identities and, only with explicit
opt-in, probe their fixed toolchains. It emits immutable but
NOT_REVIEWED_EXTERNAL image/toolchain artifacts; a complete observation still
cannot become a reviewed receipt or alter this candidate.
No Docker, network, SSD mount, dataset/GT access, scorer, benchmark replay, real signing, or profile promotion was performed for this candidate.
Acceptance preflight¶
The checked-in competitive_execution_selection_r3_acceptance_manifest.json
is a deliberately non-promoting preflight manifest. The exact checker is:
It reopens the current candidate/profile bytes, verifies their fixed SHA and
canonical identities, and requires the exact seven closure names. All seven
slots are currently PENDING with no path or hash, so the checker returns
NOT_READY (exit 3). Any self-rehashed READY edit, candidate/profile drift,
missing closure, or invented extra closure is INVALID (exit 2); this slice
does not accept an in-place promotion. Signed legal/machine/image/toolchain,
dataset, holdout, and source-review artifacts are not opened or fabricated;
the future aggregator must receive them through a separately reviewed input
contract.
Source-worktree observation¶
The additive source_worktree_review closure now has a non-promoting local
observation path, scripts/capture_competitive_execution_r3_source_worktree.py.
It uses only fixed read-only Git probes and records the current HEAD, the
existing staged+unstaged binary-diff hash contract, sorted untracked regular
file paths/content hashes, and recursive submodule revisions. Raw untracked
content is never copied into the artifact. Symlinks, hardlinks, non-regular
files, traversal paths, races, unreadable files, and bounded-size violations
fail closed. JSON and its 0444 sidecar are sealed with O_EXCL,
O_NOFOLLOW, descriptor identity checks, fsync, and a second immutable read.
The result is NOT_REVIEWED_EXTERNAL, with benchmark_eligible=false,
claim_eligible=false, and no active-profile switch. Validation reopens the
candidate/profile/source-manifest bindings and recaptures the worktree; it is
not an external custodian decision and cannot satisfy the acceptance manifest
without a separate signed review. No capture was promoted to the checked-in
acceptance manifest in this change.
The producer binding intentionally distinguishes the compatibility wrapper
competitive_execution_r3_source_worktree_v1.schema.json from the strict
target competitive_execution_source_worktree_snapshot_v1.schema.json; both
paths and file SHA-256 values are sealed. The signed review verifier reopens
both bindings (plus the sidecar schema), so a wrapper-only or strict-schema
substitution is rejected. scripts/validate_competitive_execution_r3_source_worktree_review.py
accepts only a separately supplied READY custodian policy; the checked-in
zero-key policy remains NOT_READY, and even an accepted response is
offline, candidate-review-only evidence requiring a later acceptance
aggregator and profile reseal.
Signed acceptance aggregation boundary¶
The additive scripts/validate_competitive_execution_r3_acceptance.py is a
verifier only; it contains no signer or key-generation path. It reopens all
seven closure slots through fixed authoritative adapters, then binds each
closure's immutable path/file SHA/canonical identity, validator source, trust
policy, candidate, profile, campaign, signature, and one-shot nonce. The
selection handoff and dataset-source closure now have dedicated signed
custodian packet validation in
scripts/validate_competitive_execution_r3_unsigned_closure_review.py rather
than being accepted from status fields. Their checked-in trust policies have
zero keys and remain NOT_READY.
An externally signed aggregate, even when structurally valid, is always
READY_FOR_SEPARATE_PUBLICATION_REVIEW with benchmark, claim, active-profile,
and promotion flags false. It is offline review evidence only and cannot
replace the checked-in seven-slot PENDING preflight, alter the active
profile, or publish a README/SOTA claim. Missing, partial, duplicate,
cross-kind, expired, replayed, self-rehashed, symlinked, hard-linked, or
validator/policy-drifted inputs fail closed. No external key, signature,
review packet, Docker run, dataset, GT, scorer, or benchmark was created in
this change.
The aggregate's execution_authorized=true is deliberately narrower than
publication or competitive-claim authorization. publication_authorized,
sota_claim_authorized, and active-profile/promotion remain false, and
requires_separate_publication_review remains true. The machine closure is
accepted only through the dedicated runtime signed-review adapter after the
authoritative PoP validator performs a live-host check; an offline machine
artifact cannot be promoted by projection. Fresh-holdout reviews likewise
pass through the dedicated adapter, which reopens the profile and requires
the authoritative fresh-holdout verifier's external-attestation signature and
one-shot replay result. The checked-in policies have zero keys and the
checked-in holdout profile has no valid external attestation, so both runtime
slots remain NOT_READY until an independently reviewed policy and subject
are supplied. Neither wrapper infers signature_valid or
replay_protected from status fields.