Skip to main content

Adversarial remediation tracker

Baseline: 473c6ec6ffb9c195d49c54caae1267bd069c548f. All 26 remediations implemented and locally validated. One independent Claude/Codex implementation review completed; follow-up fixes integrated. No live tenant delivery proven.

FindingScopeState
AZ-01Concurrent Azure uploads acknowledge the failed event and reject a different eventImplemented; local gates passed
AZ-02Current Sentinel schema still omits target, change and event-specific contextImplemented; local gates passed
AZ-03Identity or DCR configuration failures dead-letter valid recordsImplemented; local gates passed
SEC-01Secrets survive change values and request URLsImplemented; local gates passed
SEC-02Documented management Entra audience mismatches v2 tokensImplemented; local gates passed
SEC-05Producer authentication does not authorize application or tenant identityImplemented; local gates passed
DUR-01SDK reports dropped events as accepted and hides sender failureImplemented; local gates passed
DUR-02Fingerprint merges distinct security eventsImplemented; local gates passed
DUR-03SQL timestamp cursor stalls on a saturated timestampImplemented; local gates passed
COV-01Defaults reject role-change and export SDK eventsImplemented; local gates passed
COV-02Cloud mappings lose permission-change investigation contextImplemented; local gates passed
AZ-04Final Retry-After is lost when SDK-level retries exhaustImplemented; local gates passed
AZ-05Evidence control mappings claim more than canary query establishesImplemented; local gates passed
AZ-06Entra ingestion is implemented but provisioning remains partly manualImplemented; local gates passed
SEC-03Unchecked heartbeat counts persistently crash fleet overviewImplemented; local gates passed
SEC-04Revoked collectors can rewrite protected inventoryImplemented; local gates passed
DUR-04Conflicting event ID returns duplicate successImplemented; local gates passed
DUR-05Stale worker can settle a newer leaseImplemented; local gates passed
DUR-06Legitimate replay after retention reports false ledger corruptionImplemented; local gates passed
DUR-07Unresolved event pins unbounded ledger historyImplemented; local gates passed
DUR-08Delivery health reports healthy when output disabled or stalledImplemented; local gates passed
COV-03Signed pack parser/detection assets never activate in collectorImplemented; local gates passed
COV-04Sigma subset silently changes Boolean and equality semanticsImplemented; local gates passed
COV-05High-value filtering relies on upstream subscription remaining narrowImplemented; local gates passed
OPS-01SQL size checks follow unrestricted row allocationImplemented; local gates passed
OPS-02UI token cache contradicts repository security requirementImplemented; local gates passed

Release gates: byte-level privacy; exact event acknowledgement; outage storage/backpressure; populated database migration and recovery; source/policy/DCR field preservation; supported parser/detection activation; independent destination query. Unsupported evidence dimensions remain not-assessed.

Regression evidence map​

These tests establish local/simulated behavior only. Final gate results must bind to the combined commit, runtime versions and production defaults. Focused child results do not replace integration validation. The Rust tests named below were part of this record; the Rust collector was retired in November 2026.

  • AZ-01/03/04: AzureLogsIngestionOutputTests and both runtimes' delivery-worker suites cover reordered concurrent responses, exact acknowledgement ownership, bounded isolation, persisted outages and final server retry deadlines.
  • SEC-01/05: SemanticPrivacyTests, ProducerAuthorizationTests and Rust redaction/host suites cover canaries across SQLite/WAL/spool/captured requests and producer identity mismatches before persistence.
  • DUR-01: SdkDeliveryTests covers overflow, permanent/transient failures, timeouts, sender faults and bounded shutdown accounting.
  • SEC-02/03/04: ManagementEntraTokenTests and ManagementBoundaryTests cover signed offline token rejection, counter bounds, legacy overflow and revoked inventory writes.
  • DUR-08: DeliveryHealthTests and Rust heartbeat tests cover disabled output, retrying/stalled/backpressure states. Historical receipts cannot prove the currently configured destination worked; an empty backlog reports idle.
  • AZ-02/05: SchemaAssetTests, IntegrityAndDcrTests, PackAndEvidenceTests and Rust parity fixtures pin additive fields and restricted evidence claims.
  • AZ-06: deploy/entra/test_plan_ingestion.py plus Bicep compilation cover deterministic offline identity/federation plans and exact DCR role scope.
  • COV-01/02/05: DefaultSignalCoverageTests, CloudPermissionContextTests, Rust policy tests and the versioned population notes cover shipped SDK events, IAM grant/revoke context and default denial of routine inventory traffic.
  • DUR-02–07/OPS-01: DurabilityAdversarialTests, SqlServerAdversarialTests, Rust adversarial/compatibility tests and bidirectional queue fixtures cover identity conflicts, fenced settlement, retention incarnations, compacted history, DB/WAL admission bounds and populated v4-to-v5 migration rollback. SQL tests inspect generated server projection and keyset behavior locally. SqlServerReplayTests also covers fair bounded overlap across restart, and QueueMigrationPreflightTests verifies read-only collision reports and unchanged DB/WAL bytes before a failed-closed upgrade.
  • AZ-03 recovery: tools/audited-replay/test_replay.py covers exact-ID bounds, durable intent/outcome, audit failure before mutation and ambiguous crash state.
  • COV-03/04: PackActivationTests, DetectionSemanticsTests, Rust host tests and 89 shared semantic vectors cover signed raw intake, component status, Boolean precedence/negation/equality, malformed input, restart and backpressure. Combined replay tests exclude changing local detection evidence from source identity.
  • OPS-02: console auth tests pin in-memory MSAL cache and popup flows.

External acceptance gates​

The following remain unexecuted until an authorised test target is supplied:

  1. Actual Entra token acquisition for each supported machine identity, management login/heartbeat, correct-DCR success and unrelated-DCR denial.
  2. Destination query of unique application fixture IDs in the intended LAW/table, with exact sanitised target/change/export fields. See deployment acceptance checks.
  3. Live SQL Server tied-timestamp/restart/late-arrival exercise. Predicate and local queue tests do not certify a production database or its permissions.
  4. Any historical reconciliation or replay on a deployed queue. Production loss is conditional, not established by the simulated misattribution probe.

No live delivery, historical completeness, regulatory control or universal software-support claim follows from local green checks. Generic raw-log shipping remains outside scope; unsupported events remain denied.

Implementation council follow-up​

One independent Claude/Codex round reviewed combined implementation 2999b65. Both returned SHIP WITH CHANGES. Their review was source-based; final tests below remain the lead's responsibility. No second approval round is implied.

ObjectionDisposition and local regression
Unbounded server retry deadline pauses every leaseFixed at client and durable-worker boundaries. Long headers return without SDK sleep; deadlines cap at one hour with fixed metadata. Huge date/millisecond and restart tests in both runtimes.
Isolated HTTP 400 continually pauses healthy backlogSelectively rejected singleton retained in dead letters when another record succeeds in the same delivery call; healthy backlog continues. All-failing/configuration-wide 400 remains recoverable destination fault. Immediate audited recovery is safer than reclassifying every singleton as poison.
v4 identity collisions abort upgrade without diagnosticsAdded bounded read-only canonical-identity preflight, collision pairs and backup/reconciliation instructions. Populated v4, malformed/future schema, unchanged DB/WAL and transactional upgrade tests.
Producer metadata suppresses detection without failureAmbiguous or excessive fields now cause explicit evaluation failure and quarantine before source persistence. Matching/repaired/boundary host tests in C# and Rust.
Forward SQL pages skip/starve overlapCursor v3 pins replay bounds and alternates bounded forward/replay pages. Backlog, late rows, timestamp ties and restart regressions; v1/v2 migration documented.
Global event IDs can collide across producersImplemented HTTP 409 with repair guidance, bounded authenticated producer counters, and forwarding tests proving no source acknowledgement/discard/cursor advance. On-disk primary key remains unchanged.

The 400 decision is deliberately narrower than classifying every failed singleton as poison: a wrong DCR may reject every valid record. Without a successful sibling, Bower preserves those records pending operator repair and applies bounded storage backpressure. Dead letters are retained, never counted as delivered.

Final combined validation​

Validated source revision: ebd19267ce956359dd06dcd51b7dc18e7e520332 (2026-10-24 local date). The final reporting commit changes only this tracker; source/build assets remain identical to the validated revision. Every gate ran against the same committed, clean source tree, with production defaults covered by the targeted adversarial fixtures. Fixtures use synthetic data and simulated transports.

  • .NET SDK 10.0.402: locked restore, Release build, 839 tests, Debug reference build and full dotnet format --verify-no-changes passed.
  • Rust 1.99.0 / Cargo 1.99.0: fmt, workspace 487 tests passed; 4 existing manual tests ignored, Clippy --all-targets -- -D warnings, Release build passed. Pinned DOTNET_ROOT/PATH enabled actual C#/Rust interoperability tests.
  • Node 22.23.2: clean install, lint, 11 UI tests, production build passed. Existing bundle-size warning remains informational.
  • Python 3.11.6: 13 audited replay tests, 3 Entra planner tests passed.
  • Bicep compilation and Docker Compose Native AOT build on explicit local orbstack context passed. No containers deployed or cloud resources changed.

All 17 gates passed. Exact commands, runtimes, timing and logs are preserved in artifacts/remediation/verified-gates/manifest.json in the remediation worktree. Earlier runs remain separately retained: final-gates records the initial Rust fixture compile failure; rust-followup-gates records its successful correction; council-final-gates records the combined whitespace-only formatting failure. The final run repeats all gates after that formatting correction. A focused intermediate MSBuild run also exited prematurely; the bounded-parallelism build and both full combined Release builds subsequently passed without warnings.

Council reviewed implementation 2999b65, not the post-review fixes. Both verdicts were SHIP WITH CHANGES. The lead implemented and regression-tested the objections; no second reviewer approval is claimed. Full replies and disposition transcript: ~/.council/20261023-221225-bower-implementation/transcript.md.

This is a local remediation result, not a production release certification. Live Entra acquisition, correct/wrong-DCR authorization, LAW field-preservation query, SQL Server exercise and conditional historical reconciliation remain external acceptance gates above. Generic file/REST/Windows Event Log adapters, additional language SDKs and collector SQL polling orchestration remain roadmap work; no universal software support is asserted.