Implemented engineering
Local training, checkpoint resume, memory, rights, permissions and preregistered research records.
Inspect model evidenceEvidence ceiling E3Proof boundary
Verification research
Inspect the verifier. Challenge the conclusion.
V128 is an unpublished assurance candidate. Seller-controlled lab evidence is available for scrutiny; hardware execution and external validation remain unestablished.
APEX research
Engineering, research and external validation are distinct.
Active research
General, industry and solutions intelligence; causal reasoning, transfer and governed learning remain unqualified.
Inspect capability statusExternal validation
No qualifying outside witness, buyer-authored result or independent capability evidence.
Inspect the evidence boundaryContract Challenge
Bounded declared-predicate omission and challenge checks.
Active Evidence Acquisition
Rights-bound measurement requests; authoritative evidence remains external.
Proof-Bound Permissions
Host-enforced authority, trust, scope and resource boundaries.
Causal Outcome Modeling
A program objective with no qualified learned result or independent reproduction established.
Verified Invention
A program objective with no qualified learned result or independent reproduction established.
Federated Intelligence
A program objective with no qualified learned result or independent reproduction established.
Cross-Domain Transfer
A program objective with no qualified learned result or independent reproduction established.
Corrected-Outcome Learning
A program objective with no qualified learned result or independent reproduction established.
Adaptive Research Allocation
A program objective with no qualified learned result or independent reproduction established.
Capability Creation
A program objective with no qualified learned result or independent reproduction established.
Governed Self-Improvement
A program objective with no qualified learned result or independent reproduction established.
Long-horizon intelligence objectives remain research goals, with no achieved superintelligence claim.
Run and verify
One audit command, explicit trust boundaries.
Authenticate the package digest from a separately trusted source before execution. The supplied signature identifies a seller-controlled candidate record.
make -f research/audit/Makefile audit DIGEST=PINNED_MANIFEST_SHA256Node 24, Python 3.11+, OpenSSL 3 and the pinned audit dependencies are required.
Research access
Verification tools and limitations.
Each path exposes its evidence class rather than awarding external assurance.
Verify release and runtime chain
Exact published bytes, provenance, checkpoint and software-profile packet checks. Actual measured runtime remains blocked.
Explore →Run hostile corpus
Permanent certificate, source-quorum, rollback, provider and parser cases.
Explore →Run references
Legacy Python/OpenSSL and WebCrypto references plus an internal primary-aware Azure control-plane checker. Shared code and provider limits are explicit.
Explore →Verify transparency and run witness
Persistent local log, Merkle proofs, own keys and retained checkpoints; no Finality credentials required.
Explore →Clean-room implementation kit
Source-free public specifications, legacy and high-assurance vectors. No hidden expected outputs.
Explore →Buyer test kit
Buyer-held truth, sealed cases, signed execution and later reveal. Required sandbox execution remains blocked here.
Explore →Formal assurance
SMT quorum cardinality lemma and bounded state models; production refinement and whole-system temporal proof remain unverified.
Explore →Known limitations and hardware blocker
Exact implemented, tested, untested and resource-required boundaries.
Explore →Release history
V126 remains production; V127 and V128 candidate work is not production evidence.
Explore →