Add the Comparison Engine read port, a Core-owned adapter that loads each
timeslot as Work reads it (scoped scanner runs, rolled-up GPO verdicts, backup
creation time only) and GET /projects/:id/timeslot-comparison with validation,
a concurrent-change guard and an off switch. The API client gains an
engine-free getTimeslotComparison and the activity label.
Move compareScanRuns, comparison normalization and local comparison rules
unchanged to src/evaluation/scan-run-comparison so src/comparison can become
the Comparison Engine; only import paths change. Register the engine paths,
document it in AGENTS.md and onboarding, extend the boundary test and add the
permanent on/off invariance test. Record the decision and staging in the plan.
- Engines & relationships: Remediation Engine, its tables and UI in a new
right column; port, published Context, own data and UI relationships.
- DB Structure reads the declared engine schema assessor_remediation and
lists its tables schema-qualified.
Compliance tab layout (user corrections 2026-10-12):
- Overview row is only as tall as its taller tile; Mapping coverage is no longer forced square and both tiles share the row height.
- Annex A and BSI Kernel footprint cells grow with the tile up to 64px (5:3, 14px numbers); each Annex A theme column is as wide as its six cells and the columns spread across the tile. BSI requirements fill an auto-fill grid of the same cell size, also in the Show all window.
- Kernel footprint tile shows only the practice with the most findings.
- BSI Grundschutz++ framework tile shows 4 to 10 practices, as many as its square fits (container queries on the inner width).
Fixes from 2026-10-09:
- Compliance tab signals changed coverage or mapping history so Work reloads its assignments; `Mapping failed` no longer opens Work without its filter value.
- Hybrid ISO retrieval counts hits in rank order against the 8,000 character cap per control and always keeps the best hit, so large mappings no longer fail with regulatory_mapping_prompt_too_large.
Governance (Core-Freigabe: Governance): AGENTS.md BSI tile rule updated to the content-fitted practice count. Engine code only otherwise.
Release 2.23.0 (Phase 13F, BSI Grundschutz++ stage 2).
- Map rules directly to BSI Kernel requirements from a dedicated BSI
corpus in Qdrant; mappings and decisions carry a framework
(migration 029, verified backup before applying).
- Compare directly accepted BSI requirements with those derived from
accepted ISO controls (Consistent / Inconsistent / No ISO reference)
in a new "ISO <-> BSI review" tile in the Client Info Compliance tab,
with Accept/Reject and "Mark reviewed" per fingerprint.
- Hybrid retrieval for ISO and BSI: German restatement of the public
query, dense English + dense German + German BM25 fused by
reciprocal rank; BSI skips BES/DLS/GEB/PERS and sends 16 candidates.
- ISO prompt 1.3.0 gives the model the full ISO/IEC 27002 control text
(never sent to the browser).
- Auto-accept only the first two strong proposals (primary, high or
medium confidence).
- Recognize Purple Knight scoped tool ids and SI indicator ids in the
mapping query.
- Rename Result DB "Map all ISO controls" to "Compliance mapping" with
ISO and BSI batches; "Mapping history" lists ISO and BSI runs.
- Data Hub BSI tile shows the BSI search index state and can build or
rebuild it.
- Update AGENTS.md and the implementation plan with these decisions.
Client Info now loads the Compliance tab overview when a timeslot opens, so the first switch to the tab shows it at once.
AI overview fixes:
- drop references to IDs that are not in the input instead of failing the run
- keep source IDs when chunked analyses are consolidated (prompt version 1.3.0)
- allow CIS rule numbers and other address-like values from the input in the email/IP check
- remove the count limit for questions and assumptions; the panel shows ten with a Show all toggle
- label AI overview requests as Compliance Engine in the activity log and keep one Generating AI overview entry open for the whole run
The product version stays unchanged.
Core-Release: Platform, Presentation
- Client Info Compliance tab: overview row above the framework tiles with an
ISO/BSI switch. Mapping coverage (square donut, per mapping state) and the
Annex A footprint (all 93 controls, titles from the active corpus) or the
BSI Kernel footprint (top two practices, Show all window with all 901
requirements, unreachable ones marked, pinned detail line).
- Mapping failed and ISO control without BSI requirement open Work through
the new Compliance Engine facet ISO mapping; Client and Assessment context
row in the tab (Core release Presentation for the DashboardPanel drilldown
and context row).
- Fix: an ISO control accepted by an earlier run can be decided again after a
remap of the same pipeline version no longer proposes it.
- Overview response adds coverage, bsiFootprint and annexATitles; Work
assignments add failedFindings. Product version unchanged pending Platform release.
Compliance Engine stage 1 for BSI Grundschutz++ (Kernel requirements only):
- Pinned snapshots of the BSI Kernel catalog and the official ISO 27001
Annex A -> Grundschutz++ mapping with source URL, SHA-256 and CC BY-SA
attribution, generated by server/compliance/update-bsi-assets.mjs.
- Requirements are derived from accepted ISO controls, never stored as
decisions; missing mapping targets are reported.
- Client Info Compliance tab: BSI Grundschutz++ tile and Results window.
- Finding Reader: collapsed BSI requirements section grouped by ISO control.
- Work: BSI requirement and BSI practice facets.
- Data Hub: BSI catalog tile with Update from BSI and Use bundled catalog;
new outbound target documented in SECURITY.md.
The Compliance row moves from the Client Info Overview into its own
Compliance tab (before AI Overview). The framework tile row and its
Results button stay unchanged so later frameworks add a tile.
The tab heading carries a Last mapping button. Its window lists up to 50
mapping batches and single finding mappings of all clients, because
mappings are global per rule. Each entry shows its trigger (Automatic,
Manual, Single finding), where it came from relative to the shown
timeslot (This timeslot, Other timeslot, Other client, Unknown), status
and content-free counts. Mapping runs now record their trigger, and
single mappings their origin as SHA-256 tokens only, since runs are
exported with every client. The tab refreshes while a mapping runs and
shows the last overview at once when returning to it.
Also: history table keeps the 14px table role in every column, and the
flaky sheet-selection-cost test uses 300 instead of 2,537 sheets.
Core changes (released Presentation, Governance, Platform, Core Tests):
DashboardPanel Compliance tab, AGENTS.md decisions, product version
2.20.0 -> 2.21.0, sheet-selection-cost test.
The Compliance Engine maps newly uploaded scans and GPO evaluations to
ISO controls while the new Data Hub switch "Map new scans automatically"
is on (default off). An engine-side watcher reads only through
ComplianceReadPort: active scan runs and, for GPOs, the deviation rows of
each timeslot as filtered Result DB pages. It starts the existing ISO
batch per new scan run or per timeslot with new GPO rows; already mapped
rules are reused, so only new rules reach the model. Existing data forms
the baseline and is not mapped retroactively.
The switch lives in the engine's own settings file, not a Core table.
Switching the Compliance Engine off in the Engines popup also switches
automatic mapping off server-wide.
Core changes (released Platform, Core Tests, Presentation, Governance):
PUT /reference-data/iso-corpus/auto-mapping route, setIsoAutoMapping API
client method, subscribeEnginePreferences export, AGENTS.md decision and
product version 2.19.1 -> 2.20.0.
- Move AI overview and privacy-safe JSON export into Compliance Engine
- Add ISO control and theme facets to Work, including GPO settings
- Enrich ISO queries with baseline context, Windows code meanings and technical explanations
- Count GPO settings once and preserve editable decisions across mapping runs
- Fix GPO mapping version conflicts and backup domain identity resolution
- Add regression coverage and update architecture and reimport documentation
Extract the regulatory ISO control mapping from the Core into the
Compliance Engine (src/compliance, server/compliance, tests/compliance).
- Fix ISO mapping after the Result DB privacy merge: look up a finding's
Result DB row scoped to its project instead of comparing a stripped
project id.
- Engine reads the Core only through ComplianceReadPort and writes only
its own ComplianceStore (existing migration 026 tables, own pool);
engine writes no longer advance the Core revision.
- Engines popup gains Compliance Engine; switched off it shows nothing
and sends no requests. Permanent on/off invariance test.
- Map ISO controls opens a window like Accepted Risk; accepted controls
appear as an ISO controls section in the Finding Reader and as tiles
below the Context heatmap in Client Info (drill-down into Work).
- First two proposals of a finished mapping are accepted automatically
(provisional user decision); existing decisions are never overridden.
- Activity log names Compliance Engine work, keeps running mappings,
batches and corpus embedding until the backend ends them, and no
longer logs repeated status checks.
- Data Flow diagram shows the Compliance Engine.
- Core guard treats a bare engine folder name as engine path; new guard
area test.
Adapted pick of fdd46e8: recognition, source extractor, domain model and
tool profile for Defender for IoT alert exports, the common-description
first read with its AGENTS.md rule, and the shared markdown, recognition,
status-outcome and scan-sequence changes.
Also brings forward two fixes Defender needs on this base: the title
fallback in recordTitleFor (from 210bbc0) and the Source Rule ID
exception for ID columns that are the finding title (from 21a6266).
Product version 2.12.0 for the Defender series.
Adapted pick of d9ff6a8: device-finding and device-snapshot targets with
immutable snapshots, migration 028, repositories, API, Defender panel,
Finding Reader and the Accepted Risks register.
Unlike the original, the register keeps the base rule for its current
count, and a scoped decision covers only the occurrences it captured.
After a reimport it stays with its full history, shows Review required
and covers nothing, like a whole-finding acceptance.
Adapted pick of fe71246. Adds the vulnerability_finding result type,
missingStatusOutcome for sources whose rows are the finding, and Defender
profile 1.1.0. The Work deviation filter uses the canonical predicate
through isDeviationRecord; Michael's GPO status logic and status cache
stay in front of it.
The constraint migration is renumbered from 022 to 027. The CVE
enrichment commit e048295 is skipped for now.
Partial pick of 95ec322: a failed or queued regeneration keeps the previous successful AI overview visible, and a stored ISO mapping stays readable while the retrieval index is unavailable. The Defender device risk-acceptance fix of that commit is not taken here.
policy-evaluation rows map like scanner findings in the Finding Reader and the filtered Result DB batch. The query is built after the Result DB disclosure projection, so the customer value never reaches retrieval or the mapper. The global rule key is the technical setting ID; Work GPO rows carry their current comparison ID. Raw GPO and environment observations stay excluded. Retrieval version unchanged.
- Result DB is checked and stored on creation (migration 024, assessor_result_db_rows)
with a content-free per-row trace, per-client identity guard and Open original in Work
- Disclosure policy 3.4.0: review state no longer revealed via withheld-field trace,
per-finding scanner scores withheld, CIS numbers keep all seven levels
- Less over-redaction: "Limited" is no legal form, attached versions (Tls1.2, v1.0),
Windows 95, 2012-R2, step/option/page references, documented Microsoft GUIDs and
documented vendor values in the recommendation field stay
- Maester first read shows the native result statement verbatim
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Reference recognition (Phase 13D, 2.8.0-2.8.5):
- ATT&CK 19.2 / D3FEND 1.6.0 catalogs, versioned recognizer and
append-only reference runs (migration 023), catalog references in the
reader, Data Hub status and reprocessing, MITRE online update.
Scanner report and Work (2.8.6):
- Resolve scope-named Purple Knight runs to their profile via runNames.
- Maester "Remediation action" section as next step; PowerShell scripts
as command blocks with display-only line layout.
- Purple Knight likelihood without the appended references section.
- CIS section titles from the run's own report in the CIS Category
filter, Work headings and badge tooltips.
- Keep all record metadata when GPO results rebuild All Records
(withRebuiltLayout) with a guard test for future metadata.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Binh
Michael
Tronth