Skip to content

status.md

Latest commit

 

History

History
58 lines (51 loc) · 7.01 KB

status.md

File metadata and controls

58 lines (51 loc) · 7.01 KB
Raw

Implementation status

Updated: 2026-09-01

Working now

  • Capability probe with AVAILABLE, UNAVAILABLE, UNVERIFIED, and BLOCKED states.
  • Case IR v1 validation, canonical JSON, and stable SHA-256 case IDs.
  • Typed campaign lifecycle backed by an append-only SQLite journal.
  • Resume from RUNNING without a second dispatch.
  • ILabBackend plus eight deterministic, virtual-time simulator scenarios.
  • The ordinary-campaign Hyper-V backend always fails closed; destructive Track B execution is available only through the explicit gated G1 command.
  • Versioned synthetic triage parser, stack normalization, signature v1, and exact grouping primitive.
  • Sequence delta debugging demonstrated from 14 to 3 operations.
  • Hierarchical minimization reduces the same fixture from 552 to 188 canonical bytes by shrinking schedule, optional fields, strings, and integers while preserving the signature.
  • Deterministic 3/3 replay policy with infrastructure errors excluded from the vote.
  • Replay voting can be used as a probabilistic minimization oracle for flaky synthetic targets.
  • Content-addressed artifact copy, SHA-256 manifest, semantic bundle verifier, and static report.
  • Deterministic boundary scalar, payload block, and sequence insert/delete/swap mutation operators with lineage and seed-keyed hash-ranked candidate caps that do not privilege early enumeration positions.
  • Energy-ranked novelty corpus driven by synthetic semantic feedback.
  • Exact-signature finding deduplication and deterministic fuzz campaign runner.
  • Verifiable fuzz artifacts containing corpus, per-execution metrics, coverage, findings, manifest, and static report.
  • Paired E1 2×2 policy ablation crossing keep-all/novelty-only admission with uniform/energy-ranked parent selection; operator and candidate selection remain identical across arms.
  • Censoring-aware E1 Kaplan–Meier discovery curves plus paired factorial contrasts, raw CSV, claims limits, and manifest verification.
  • Paired E2 stateful-versus-single-call experiment with sequence length as the only changed variable.
  • Versioned experiment provenance: distinct source-commit/artifact time semantics, deterministic timestamp policy, clean HEAD == git_commit binding for canonical records, GIT_TREE_SHA256_V1 object-database source digest, definition digest, Case IR version, and engine version. M1 additionally records and verifies canonical simulator-environment, scenario-fixture, minimizer, and replay-policy digests.
  • Deterministic evidence text uses LF, UTF-8 without BOM, invariant culture, and ordinal ordering; generated JSON, CSV, HTML, reports, and manifests are contract-tested against platform newline drift.
  • Fuzz termination telemetry distinguishes budget completion from scheduler-limit termination and records duplicate/empty scheduling work.
  • Draft 2020-12 structural Case Schema/runtime parity tests execute a real schema validator over checked-in cases, randomized property order, and generated valid/invalid boundaries for depth, integers, Unicode rune lengths, lineage, fields, mutation metadata, operations, and schedules. Cross-field schedule-length equality and the byte-size envelope are tested separately as runtime semantic and transport invariants. Raw byte parsing rejects malformed UTF-8 before canonicalization.
  • Exact .NET SDK pinning, locked NuGet dependency graphs, immutable Action revisions, and Windows 2022 plus Ubuntu 24.04 simulation CI lanes.
  • Fault injection for ambiguous journal commits, duplicate/colliding events, CAS interruption/corruption, manifest loss/truncation, and invalid virtual-clock jumps.
  • Fail-closed clean reviewer packaging using git archive, ZIP inspection, and separate recorded-results packaging.
  • Safe CI that builds and tests simulation only; dependency audit treats NuGet advisories as errors.
  • Track B source boundary: a safe-default repository-owned KMDF target, explicit lab-fault build property, deterministic Case IR compiler, fixed device path, driver hash pinning, write-through guest dispatch journal, strict real-lab profile validation, and dump-stability state tracking. This source has not yet been WDK/Hyper-V validated.
  • One-command Track B G1 host automation with exact VM/checkpoint checks, private-switch enforcement, clean restore, heartbeat/reboot observation, one discovery plus three cold replays, stable dump acquisition, WinDbg batch triage, exact-signature voting, and finally recovery.
  • Real WinDbg parser and sanitized REAL_LAB evidence builder/verifier; public bundles require 3/3 matching signatures and record—but do not publish—the raw dump hash.

Last checked-in simulation baseline

  • Evidence source commit: 55c31b25902157cddc9014bb9fdaa598bede40a2; GIT_TREE_SHA256_V1 source-tree digest: bc0d06a2759e2429b0fec5683f3102e4644628f30b3fe6032a21a233d1333338. Recorded evidence is committed in its child evidence/release revision to avoid circular self-reference; exact reruns must check out the source commit.
  • Release build: zero warnings and zero errors.
  • Latest protected-main CI baseline: 66 tests passed on Windows 2022, and the cross-platform simulation lane passed on Ubuntu 24.04. The final evidence-freeze source passes 73 tests locally with exact .NET 8.0.100 on Windows; both canonical evidence runs produced byte-identical manifests. This does not constitute WDK/Hyper-V Track B runtime validation.
  • All eight controller scenarios reach the expected classification.
  • Evidence corruption and untracked files are rejected.
  • Fuzz campaign corruption and untracked files are rejected; identical campaigns produce identical manifests.
  • The checked-in G3 current-policy record discovers the known synthetic signature at execution 211 in its 256-execution fixture, with 99 semantic feedback elements and 92 retained corpus cases.
  • Recorded E1 v2 four-arm outcomes are 14/20 keep-all energy-ranked, 7/20 keep-all uniform, 18/20 novelty-only energy-ranked, and 12/20 novelty-only uniform discoveries. These are descriptive paired-seed observations, not a general superiority claim.
  • Recorded E2 expressiveness validation: stateful 12/20 discoveries versus single-call 0/20 under paired synthetic trials; by construction, only the stateful arm can represent the required three-operation chain.
  • The recorded minimization/replay bundle reduces the known synthetic trigger from 14 to 3 operations and 552 to 188 canonical bytes while preserving the exact signature in 3/3 simulated replays.
  • Duplicate JSON keys, unknown contract members, invalid state transitions, unsafe paths, and real-backend acquisition fail closed.
  • Generated case, scenario, capability, environment, finding, run, fuzz campaign, and E1 experiment documents pass their JSON Schemas.

Not implemented yet

  • Similarity-assisted clustering and manual split/merge audit records.
  • SQLite metadata for CAS references and retention garbage collection.
  • Controller resume from every intermediate evidence-production state.
  • A genuine sanitized G1 evidence bundle from the owner's pinned Windows/WDK/Hyper-V lab. Source completeness and CI syntax/build validation do not substitute for this runtime evidence.