Extends Method C: for each unknown narrative token that pattern text names, suggest
the keyword_dictionary tag = the RequiredComponentTags shared by the naming
patterns (ranked by frequency, kept only when shared by >=40% of them, top 3).
Surfaces real dictionary gaps like "zwischenkreis" -> stored_energy and
"updates" -> has_software, which close coverage without hand-editing the dict.
Two precision fixes to Method C while here:
- patternsMentioning now matches WHOLE WORDS, not substrings — substring matching
flagged fragments like "stehen" inside "entstehen" and produced nonsensical
tag suggestions.
- a token is only proposed with a tag if one is shared by >=40% of its naming
patterns, so diffuse common verbs (spread across categories) drop out.
Wired into iace-audit propose -> audit-reports/vocab.{md,json}. Residual
common-verb noise is left to the human/LLM filter rather than a hand-grown
stopword list. Type 4 (coverage blind spots) + P3 (pin accepted proposals into a
GT case) remain for slice 6.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Surfaces fired patterns whose zone names terms the machine's narrative never
mentions — foreign framing that leaks through terms not yet in domainGateTerms
(once a term is a gate term, the ghost-pattern invariant already fences it out).
- FindFramingCandidates (proposer_framing.go): per fired pattern, zone terms with
no narrative echo (minus a generic hazard-location stoplist). Echo matching is
bidirectional to survive German compounding (narrative "Steuerung" echoes zone
"Steuerungssystem"). Heuristic verdict foreign (fully orphan) / plausible
(partial). Over-surfaces by design — human/LLM is the precision filter.
- Wired into iace-audit propose -> audit-reports/framing.{md,json}, threshold via
IACE_FRAMING_MIN_ORPHAN (default 0.6).
Honest finding: genuine wrong-MACHINE framing (Walzen, Transportbaender) no longer
fires thanks to the machine-type gate; the residual is mostly cyber/control
patterns with generic-industrial zone vocabulary, candidates for re-framing.
Proposal types 3-4 (vocab->tag, coverage blind spots) remain for slice 5.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Makes the offline proposer runnable end-to-end.
- BuildProposerInput (proposer_input.go): non-test engine->hazards path. The
PatternMatch->Hazard converter is lifted out of the GT test files into
production scope so both the tests and the CLI share one pipeline.
- iace-audit propose <narrative.json> [<ground-truth.json>]: detect candidates ->
GT-screen survivors (when a ground truth is given) -> judge (HeuristicJudge by
default, LLMJudge over ollama when IACE_PROPOSE_LLM=1) -> write the human-review
queue to audit-reports/proposals.{md,json}. Propose-only.
Smoke run on a dishwasher narrative: 32 fired -> 3 candidates -> queue with a
confident duplicate, a confident distinct, and one punted to the LLM judge; GT
wall recall-safe. Live qwen is opt-in via env; the heuristic default keeps the
tool runnable (and CI deterministic) without a model. Proposal types 2-4
(foreign-framing gates, vocab->tag, coverage blind spots) remain for slice 4.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Adds the semantic judgement layer on top of the slice-1 detector + GT wall.
DEV-TIME, propose-only — nothing mutates the library or runtime.
- CandidateJudge interface with two implementations: HeuristicJudge
(deterministic default/fallback, used in tests) and LLMJudge (offline, over the
shared llm.ProviderRegistry via the LLMCompleter adapter). LLMJudge degrades to
"uncertain" on any transport/parse error — it can never break a run.
- BuildJudgePrompt: the ISO 12100 same-vs-distinct prompt, unit-tested
deterministically even though the call is not.
- RenderProposalQueue: markdown human-review queue with a suggested action per
candidate (supersede / keep both / needs review).
On real warewashing output the heuristic punts to "uncertain — needs the LLM
judge" for exactly the two recall-safe near-dupes (HP807/HP033 update,
HP101/HP096 winding-vs-friction), making the LLM's role explicit. All 3 GTs
unaffected (read-only). Live qwen wiring + a CLI/file queue are slice 3.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
First thin slice of the offline library-improvement proposer. DEV-TIME ONLY,
propose-only — it never mutates the pattern library or the runtime.
- FindDedupCandidates (proposer_dedup.go): structural near-duplicate detection
over the fired patterns (category + measure/zone/scenario overlap). Bakes in
the P1 lesson: only same-category pairs compare, and pairs with different
operational states are never proposed (normal-operation vs maintenance are
legitimately distinct, e.g. HP011 vs HP077).
- ScreenSupersession (proposer_screen.go): the wall. A proposal is safe only if
(1) dropping the hazard does not reduce GT recall AND (2) keep/drop do not
credit DIFFERENT GT entries. Check 2 catches distinct hazards that merely share
measures (HP2201 hot surface GT 1.3 vs HP2202 hot ware GT 1.4) which recall
alone would wave through.
On real warewashing output: 3 candidates -> 1 BLOCKED (distinct GT), 2
RECALL-SAFE for human/LLM review (the update + winding/friction near-dupes).
Nothing auto-applied. All 3 GTs unaffected (read-only). The LLM judgement and a
CLI/file queue are slice 2.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
HP013 (stored electrical energy) fires for dishwashers via the broad stored_energy
tag but its zone is framed for Batteriefaecher/USV-Anlagen, which a dishwasher does
not have. The precise residual-voltage pattern HP144 (Frequenzumrichter/Zwischenkreis,
Priority 90) already fires and covers the same hazard. Add HP013 to the
warewashing-scoped supersession set so the duplicate is dropped only when
dom_warewashing is present.
Warewashing recall stays 100% (25/25), precision 92.6% -> 96.2%. Kistenhub/Bremse
keep HP013 (no dom_warewashing); 26 Bremse pins + benchmark unaffected.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The generic hot-surface patterns HP016 (high_temperature) and HP018 (actuator
burn) fire for dishwashers via broad tags and duplicate the precise warewashing
pattern HP2201 (Boiler/Tank/Spuelkammer). Suppress HP016/HP018 only when
dom_warewashing is present, so the specific pattern wins and the duplicate is
dropped. Scoped to the domain tag -> Kistenhub/Bremse and every non-warewashing
machine keep the generic patterns unchanged.
Warewashing recall stays 100% (25/25), precision 90% -> 92.6% (2 dupes removed).
Bremse 26 pins and Kistenhub benchmark unaffected.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
ListHazards returned hazards in pattern-firing order, which reads as a jumble.
Sort by EN ISO 12100 hazard group (A. Mechanisch, B. Elektrisch, C. Thermisch,
D. Pneumatik/Hydraulik, E. Laerm, F. Ergonomie, G. Stoffe, H. Software/Steuerung,
I. Cyber, J. KI), stable within a group. Matches the frontend CATEGORY_LABELS.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
HP1717 was gated on the generic stored_energy tag (carried by a frequency
converter's DC link) + pneumatic_pressure (emitted by "Boiler unter Druck"),
so it leaked into the dishwasher despite the absence of any pneumatics. Require
pneumatic_part instead. The Bremse pin is a static pattern->measure check
(unaffected); full suite incl. Bremse coverage and Kistenhub 97.1% unchanged.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
HP753 (lithium thermal runaway), HP754 (battery off-gassing) and HP755 (HV
battery shock) were gated on stored_energy, which a frequency converter (C034,
DC-link capacitors) legitimately carries — so they leaked into any machine with
a VFD (surfaced by the dishwasher after the Frequenzumrichter narrative). Now
require the "battery" tag; add lithium/batteriespeicher synonyms so real
battery-storage machines still emit it.
GT #3 100% recall unchanged, battery themes gone from the dishwasher log;
Kistenhub 97.1% and Bremse pinned mappings unchanged.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Phase 1 complete. GT #3 recall 84% -> 100% (25/25 matched), no regression:
- HP2207 backflow / potable-water contamination (EN 1717) + measure M2209
(Rueckflussverhinderer / Systemtrenner) — the only genuinely new hazard.
- HP2208 cut on sharp edges/screens (new sharp_edge tag from scharfe-Kante/Sieb).
- HP2209 unexpected restart during maintenance (dedicated dom_warewashing pattern;
avoids flooding the log via the broad moving_part tag).
- Spray-arm contact now covered by the enclosure-re-scoped contact patterns.
Kistenhub 97.1% and Bremse pinned mappings unchanged; 0/28 hazards without a
measure. Completes the commercial-dishwasher (white-goods Phase 1) coverage.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Phase 1 of the commercial white-goods expansion (EN ISO 10472 family). Extend
GT #3 with 8 completeness hazards a Fachmann expects but that were neither in
the GT nor previously questioned: dry-run boiler overheating, residual/stored
electrical energy, sharp-edge cut, tipping, interlock-failure, unexpected
restart, backflow (EN 1717), microbial/legionella. Enrich the UC-M narrative
with the real features so existing library patterns can fire.
Result: 4/8 auto-covered by existing patterns (dry-run, residual voltage,
tipping, interlock-failure) — recall 84% (21/25). Remaining gaps documented:
spray-arm contact (4.3), sharp-edge cut (4.6), backflow (2.3), restart (6.4).
Gate the re-surfaced CNC leak ("spanende Bearbeitung", high_temperature-only)
via dom_cnc. Kistenhub 97.1% and Bremse pinned mappings unchanged.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Class D (generic keyword hygiene, GT-guarded). Two over-broad keyword->component
mappings produced ghost components:
- "kuehl"/"cool" -> Kuehlaggregat (C095) matched product variants
("Cool-Ausfuehrung") and outputs ("kuehle Glaeser"). Narrowed to cooling-UNIT
terms (kuehlaggregat, kuehlanlage, kuehler, kaeltemaschine, chiller, rueckkuehl).
- "filter" -> Absauganlage/Oelnebelabscheider (C124) matched any filter
(Laugen-/Wasser-/Oelfilter). Keep "filteranlage" only.
No pattern or GT test depends on these mappings (Kistenhub/Bremse use hand-crafted
inputs). UC-M now parses 6 plausible components (was 8 incl. the two ghosts).
Warewashing GT recall 82.4% and Kistenhub/Bremse pins unchanged.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Class C (phase-aware, generic EN ISO 14120). A contact/entanglement hazard from
a moving part is removed during NORMAL operation when the part is behind an
interlocked guard; it remains only when the guard is open (maintenance/cleaning).
- New HazardPattern.GuardableByEnclosure flag; set on HP096 (friction at
rotating surfaces) and HP101 (entanglement of hair/clothing).
- Narrative emits interlocked_enclosure for an interlocked door/hood.
- pattern_enclosure.go: suppressedByEnclosure (drop in normal-op-only contexts)
+ guardedLifecycles (re-scope to maintenance/cleaning).
- GT #3 gains the maintenance-phase entanglement/friction rows.
Generic + regression-safe: machines that do not emit interlocked_enclosure are
unaffected. GT #3 recall 80% -> 82.4%, one false positive removed (Aufwickeln).
Kistenhub 97.1% and all 26 Bremse pinned mappings unchanged.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Add ground_truth_warewashing.json + TestWarewashing_GTCoverage. The test runs
the UC-M narrative through the SAME chain as production (ParseNarrative ->
engine -> relevance + cyber filter), so keyword/gating fixes are measured on
the real hazard set, and false positives show up as "extra".
Class A (generic keyword hygiene): spuelarm/spuelfeld no longer map to library
component C004 ("Drehtisch" / rotary table) — that mislabelled the spray arm.
Keep the rotating_part tag. Removes the bogus "Drehtisch" hazard.
GT #3 baseline -> after Class A: recall 80% (unchanged), one false positive
(Drehtisch) removed. Kistenhub 97.1% and Bremse pinned mappings unchanged.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Add commercial-dishwasher hazard patterns (HP2200-HP2206): hot-water/steam
scald on door opening, hot surfaces, hot ware, corrosive detergent/rinse-aid
burn, respiratory irritation, door pinch and wet-floor slip — each gated by
dom_warewashing so they never leak into other machine classes. Add the
matching warewashing protective measures (M2200-M2208).
Tighten capability-domain gating: emit dom_flame/dom_glue and add welding
surface-form gate terms (schweissarbeitsplatz, schweissfunke, lichtbogenzone,
...) so the welding/flame/glue burn patterns stop leaking into thermal-capable
machines such as a dishwasher.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Cumulative resume run added 64 new NASA NTRS docs (query/page pool then
exhausted): 164 total, 73 applicable failures, all public-use licensed.
NASA stays a generic component failure-mode library; not scaled further.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
100 NASA NTRS docs processed (Claude Haiku 4.5), 55 applicable failures
extracted with verbatim source quotes; all licences public-use-permitted
(NTRS GOV_PUBLIC_USE_PERMITTED / PUBLIC_USE_PERMITTED), each passes the
failure-knowledge allowlist. Register now serves the real corpus in the
FMEA "Quelldokumente" panel.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Quote-verifiable failure extraction via Claude (Haiku 4.5): PDF sent
directly, tool-schema forces verbatim source quotes + applicable flag +
confidence — replaces the unreliable local llama run. Only applicable=true
tuples ingest into bp_iace_failure_kb; every processed doc lands in the
source manifest.
Frontend: FMEA tab now shows a "Quelldokumente" register (every document we
use, with source + licence + link + what was extracted) served from the
embedded manifest via GET /iace/failure-knowledge/sources. Manifest is
placeholder until the 100-doc Haiku run is folded in.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Stage 1 of the FailureKnowledge bulk loader: harvest NASA NTRS
lessons-learned with a strict public-reuse gate (NTRSUsable: public
release, not export-controlled/EAR/ITAR, not CUI, PUBLIC_USE_PERMITTED,
no third-party copyright). NTRSPDFURL prefers the PDF download for
downstream text/OCR extraction. GET /iace/failure-knowledge/ntrs runs
the live harvest and returns only the licence-clean records.
Pure parse/gate helpers are fixture-tested (usable vs ITAR / third-party
/ restricted / video-only); accepted licences also pass the FK allowlist.
Next: tuple extraction (abstract -> FailureKnowledge) + Playwright/OCR for
scanned PDFs -> bp_iace_failure_kb.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
P1 of the auto-FMEA build plan: establish the public-domain methodology
foundation (no AIAG-VDA/SAE/IEC tables reproduced).
- fmea_data_sources.go: MIL-STD-882E severity (Cat I-IV→1-10) + probability
(A-F→1-10 with per-hour λ bands), OccurrenceFromRate(λp·α), SeverityForCategory,
MIL-STD-1629A CriticalityCm = λp·α·β·t. Own 1-10 projection, government-anchored.
- 4 versioned source docs (MIL-STD-1629A, MIL-STD-882E, NASA RCM, FMD-91/NPRD-91)
ingested into the new RAG collection bp_iace_fmea_kb (whitelisted).
- Tests for all scales/mappings/criticality (green).
Next (P1 step 2): fetch FMD-91/NPRD-91 bulk λ/α tables from DTIC.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- architecture.go: DataSources now reflect the real ingested set (ESAW 2023,
BLS CFOI, OSHA OTM, PRISM, cobot CC-BY, HSE) with their RAG collections;
risk stage cites BLS + the searchable RAG layer; matrix stage now mentions
the distance-benchmark dimension.
- Architektur & Datenfluss tab: new DataFlowDiagram — 4 lanes (input →
knowledge/RAG-evidence → deterministic engine → outputs) with live counts.
- scripts/ingest_iace_kb.sh: idempotent E1 ingest — creates the 2 collections
and uploads the 6 datasources docs against a configurable RAG_URL (for prod
Qdrant), with retry.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Cross-checked cobot_biomech_limits.md against both source papers:
- Behrens et al. 2022 (Frontiers): 10 body regions spot-checked, force
values match the paper EXACTLY in both columns (pinching + impact).
- Park et al. 2019 (PLOS ONE): lowest/highest/range pressure values exact.
Fix: 28 -> 29 body locations; add a verification stamp. Threshold VALUES
were already correct (no data change), so no RAG re-ingest needed.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Versioned, license-tagged source docs for the multi-layer GT knowledge base,
ingested into the new core RAG collection bp_iace_safety_kb (whitelisted in
the RAG search handler):
- prism_risk_methodology.md — OPSS PRISM v2 (OGL v3): full severity(4)×
probability(8) → risk-level matrix (Serious/High/Medium/Low), RAPEX-aligned.
- cobot_biomech_limits.md — CC BY 4.0 papers (Behrens 2022 / Park 2019):
force (N) & pressure (N/cm²) pain thresholds by body region (the data behind
ISO/TS 15066, cited from the open papers — standard tables NOT reproduced).
- hse_example_risk_assessments.md — HSE (OGL v3): qualitative hazard→control.
- osha_robot_safety.md — OSHA OTM (public domain): 250 mm/s teach anchor,
robot hazard taxonomy, safeguarding hierarchy.
No DIN/EN/ISO/IEC/DGUV content reproduced; each doc states its license + attribution.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
US severity anchor complementing ESAW: BLS Census of Fatal Occupational
Injuries (public domain), event/exposure distribution 2023-24 + the
machine-relevant "Contact incidents" breakdown (struck/caught/compressed
by running powered equipment: 226/213). Key finding: in MANUFACTURING,
contact is the leading fatal event (104/353 = 29.5%) — independent support
for the model's mechanical-contact emphasis. Ingested into the core RAG
collection bp_iace_accident_stats.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
C1: drop the misleading OSHA §1910.212(a)(5) fan-guard citation from M602
(overhead lift clearance) — EN 349 + EN ISO 13854 already cover it.
C2: frame M237's 25/500 mm as Richtwerte to be determined per EN ISO 13854
(single factual values in prose are facts, not table reproduction — but
keep the conservative caveat).
C3: keep ergonomic W=2 deliberately and document why — ESAW ranks it the most
frequent non-fatal mode (24.7%) but that population doesn't transfer to an
acute machine point-hazard; the machine GT governs.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Adding M605 (drive-limited general speed) and M606 (limited descent on
energy loss) to the library wasn't enough — measures only get suggested
if a pattern's SuggestedMeasureIDs references them. Add M605 to the three
lift crush patterns and M606 to the floor-stop descent pattern (HP2100),
so a re-seed actually attaches them and the distance benchmark closes the
≤150 mm/s gap.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CompareBenchmark now also compares the engine's numeric dimensions
(mm gaps, mm/s speeds) against the professional's GT measures: parses
distance tokens from both sides (German thousands/decimal aware),
reports matched / gt_only (gaps) / engine_only + an agreement %.
Surfaces as result.distances on the existing benchmark endpoint.
Deterministic, no LLM. On the GT-derived seed sessions it mainly guards
DRIFT; its real value is new sessions. Real-GT test pins that the engine
covers the Bremse (250 mm/s, 250/850 mm) and Kistenhub (25/120 mm,
150/75 mm/s) headline dimensions.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The GT distance benchmark surfaced three Fachmann lift values the engine
carried no measure for: general lift/lower speed (≤150 mm/s), the low-zone
inching regime (<200 mm floor clearance, ≤75 mm/s), and limited descent on
power loss (≤100 mm). Extend M603 (inching) and add M605 (drive-limited
general speed) + M606 (load-holding on energy loss). Values framed as
generic hoist recommendations with EN 1570-1 reference, not GT-memorised.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Makes the OSHA minimum-distance anchor visible per measure in a project
without a DB schema change or re-seed: persisted mitigations store the
measure NAME verbatim (not the catalog ID), and measure names are unique
across the 578-entry library (pinned by test), so a name→ID resolver
bridges the gap.
Backend: MeasureIDByName + MinimumDistancesForMeasureName/LinksForMeasureName;
/iace/minimum-distances now accepts ?measure_name=; link table enriched with
measure_name for one-request UI matching.
Frontend: useMinimumDistances loads the link table once and keys it by name;
OshaDistanceNote renders the anchor (value/CFR/license/EU-hint/relation) on the
matching measure group in the Maßnahmen tab.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Executes the accident-statistics pipeline for the risk anchors:
- Refresh contactModeEvidence with real Eurostat ESAW figures
(dataset hsw_ph3_08, reference year 2023): impact 24.0%/21.4%,
struck-by 13.0%/23.8%, sharp 14.5%, trapped/crushed 13.8% (fatal),
+ new physical/mental-stress mode 24.7% → ergonomic. GT-calibrated
tier VALUES unchanged; the real data confirms the ordering.
- Add the versioned source document (datasources/esaw_accident_stats_2023.md,
ESAW CC BY 4.0 + OSHA public-domain context) that is ingested into the
core RAG collection bp_iace_accident_stats for searchable evidence.
- Whitelist bp_iace_accident_stats in the RAG search handler so seeding
can full-text search the statistics with citation at seed time.
Two-layer design: the small license-tagged code table stays the deterministic
tier/citation lookup; the RAG holds the searchable source evidence.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The May-built OSHA distance library (minimum_distances.go, 29 CFR 1910,
US public domain) was dead code — zero callers, no route, no test, while
the mm values that actually appear in measures are independent hand-prose
(some carrying ISO 13854/13857 values, not OSHA).
This surfaces it without touching the measures response contract:
- GET /iace/minimum-distances (+ ?measure_id=) returns the distances, the
curated measure→distance link table and the licensing note.
- AllMeasureDistanceLinks/MinimumDistancesForMeasure resolve only the
defensible links (M600 value_source; M254/M065 public-domain crossref to
ISO), with the relation made explicit so the join stays honest.
- architecture.go lists the OSHA library so it shows in the audit explainer.
- Tests: inch→mm conversion + license completeness, link integrity, and a
consistency test pinning that a value_source measure's prose still
matches the OSHA source (codifies the audit finding as a regression gate).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Adds an auditor-facing view of the IACE engine: a clickable 10-stage
pipeline flow (Grenzen-Formular → ParseNarrative → Pattern-Gates →
Relevanz → Caps → Gefährdungen → Maßnahmen → Risiko → Normen → Matrix),
plus live library counts, the data-source/license register (incl. the
DIN/Beuth + DGUV exclusions), and the norm-matching logic that reconciles
DIN/ISO/OSHA machine-type vocabulary via canonicalMachineType folding.
Backend: BuildArchitecture() with LIVE counts so the diagram can never
drift; GET /iace/architecture; collectAllNorms() extracted from
SuggestNorms as the single source of truth for the norm-library count.
Frontend: useArchitecture hook + page + new IACE nav tab.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Surfaces the public-statistics provenance for the contact-mode probability
tiers so generated risk numbers are auditable and attributed (not RAG —
~a dozen stable aggregate facts are better as a license-tagged code table).
- risk_data_sources.go: RiskEvidence register (Eurostat ESAW figures + CC BY
4.0 attribution) for the documented contact modes; RiskDataSourcesNote.
- risk_suggestion.go: the W justification now cites the actual ESAW share +
license where documented; RiskSuggestion gains a data_source field.
- GET /iace/risk-data-sources returns the evidence register + attribution.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds GET /projects/:id/risk-matrix — a confidence-aware risk view computed
on read from each hazard's category/scenario/lifecycle using the SAME model
as the GT benchmark (no persistence, so it never goes stale against the
model; the hand-defaulted iace_hazards risk columns stay untouched).
- risk_matrix.go: EstimateHazardRisk (single source of truth for S/F/W/P +
range + level + confidence) and BuildRiskMatrix (per-hazard list + a 5×5
Severity×Probability aggregation grid with dominant level per cell).
- Frontend: RiskMatrix grid in the Risikobewertung tab (muted colours per
the confidence-aware tonality), level counts + tool-confidence summary,
fed by useRiskMatrix. Shows risk for EVERY project, not only GT ones.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Report the tool's risk number as a plausible range with a confidence
label instead of a false-precision point value (confidence-aware
tonality — the assessment is confirmed by the DSB / safety expert).
- risk_estimation.go: EstimateConfidence (hoch/mittel/niedrig from how the
contact mode resolved), EstimateRiskRange (S±1 and aggregate L=F+W+P ±1,
the empirically validated per-parameter accuracy), RiskLevelRange; share
the riskBandLabel thresholds with EstimateRiskLevel.
- risk_benchmark.go: RiskComparisonPair gains eng_risk_point/low/high +
level + level_range + confidence; RiskAgreement gains high_confidence_pct.
- RiskComparison.tsx: per-hazard range "low–high (level range)" + point,
confidence chip, and an aggregate confidence line; types in useBenchmark.ts.
- Unit tests for the range/confidence helpers.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Domain-gate ~15 foreign machine classes (pool, amusement, paint booth,
tank farm, reactor, lathe/chips, saw, film/carton, robot, mobile cab,
asbestos, playground swing) in pattern_domain_gates.go so ungated hazard
patterns stop leaking into unrelated machines; matching emit keywords
added in keyword_dictionary.go (gate+emit share one vocabulary).
- Extend the cross-domain precision guard to 6 machine classes (press,
cobot, motor, welding + the 2 GTs) with per-case homeDomains, so a
machine's own domain terms are never flagged. GT coverage stays 100%.
- Reconcile the fine-grained norm machine-type vocabulary (455 keys) with
the 68 canonical dropdown keys via canonicalMachineType() family folding
in matchNorm — welding 0->17, robotics_cobot 0->6, press 8->13,
circular_saw 1->35 machine-specific C-norms. Pattern gating left strict.
- Fix initialize?force=true summary index-shift that mislabeled counts
(reported matched-patterns as "hazards"); now uses named step vars.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Engine precision (stop foreign-machine patterns leaking into a project):
- Wire project.MachineType into the engine machine-type gate (empty input no
longer fires every machine class — press/cnc/excavator/crane/medical...).
- Capability-domain gating extended by 7 domains (outdoor, ventilation,
machining, bulk, palletizer, playground, fitness) so domain-specific hazards
only fire when the narrative names that domain; emitted via keyword_dictionary.
- Relevance backstop moved into iace (single gating contract, testable), and its
dominant false-anchor class removed (a long pattern word no longer matches a
short common token; prepositions/leitung added to the generic stoplist).
- New guard tests: TestCrossDomainPrecision (full pipeline, 0 foreign per GT) and
TestPatternReachability now asserts 0 dead patterns. Both GTs keep coverage 1.0.
Reachability fix: the 51 dead patterns required electrical/pneumatic/hydraulic
tags nothing produced — renamed to the canonical electrical_energy/
pneumatic_pressure/hydraulic_pressure/hydraulic_part.
Component review (negation is best-effort + expert-correctable):
- Parser surfaces negated components (ComponentMatch.Negated) instead of dropping
them; negated contribute no tags/energy → no phantom hazards.
- presence_status (vorhanden|nicht_vorhanden|geloescht) + ce_marked on components;
only `vorhanden` feed matching. CE+safety-relevant flags the PL/SIL obligation.
- Force re-seed preserves the expert's component decisions instead of wiping them.
- Tag-based component→hazard assignment (was: all on the first component).
- Negation-aware narrative parsing ("keine Pneumatik" no longer extracts it).
Local-dev DB: ai-sdk sets search_path=compliance,core,public; reconcile migrations
152-156 bring the consolidated local iace tables to the current schema + add the
presence_status/ce_marked columns. Machine-type vocabulary endpoint for the form.
[migration-approved]
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Diagnosis: engine F mean 3.56 vs professional 2.56; the dominant disagreement was
normal-operation hazards getting F=4 where the professional assigned 2. Lowered
the lifecycle→F mapping (normal operation 4→3, occasional phases 3→2). New
TestGT_RiskComparison_CrossGT runs the exact production comparison on BOTH GTs:
F within±1 rose to 95% (robot cell) and 94% (lift) — generic, not lift-tuned.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
#1 Risk-number comparison in the benchmark: ComputeRiskComparison derives the
tool's S/F/W/P + Fine-Kinney per matched hazard and compares to the GT values;
exposed on the benchmark response and rendered in a new RiskComparison table
with GREEN/YELLOW/RED traffic lights on the risk number R (like the Excel),
plus per-axis within-1 agreement cards.
#2 Generic misuse pattern HP2103 "Personenbefoerderung auf Hebezeug" — gated to
lift-family machine types, fires for ANY lifting device (not machine-specific).
#3 Benchmark matcher is now 1:n — one broad engine hazard may cover several
fine-grained GT sub-scenarios (foot/hand/leg crush), so coverage reflects real
risk coverage rather than 1:1 wording matches.
Validated on BOTH ground truths (robot cell + lift): leakage 0, ghosts 0,
coverage held.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
A "Splitterflug bei Werkzeugbruch" pattern leaked into a lift re-seed because
its press hint ("Pressraum") lives in ZoneDE, which applyDomainGates did not
scan. Add ZoneDE to the gated text. Leakage stays 0, ghosts 0, coverage held.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Observed on a real Kistenhubgeraet (lift) project: generic mechanical patterns
(e.g. HP1000 "Quetschen Arm zwischen Pressenteilen") carry NO machine type and
only generic tags (crush_point, rotating_part), so they fired for a lift; the
narrow domain-gate terms missed their press/welding/glass wording.
Broadens domainGateTerms (pressenteil, pressraum, blechbearbeitung,
punktschweiss, schweisselektrod, elektrodenspalt) and adds a dom_glass domain
(glasschneid/glasbearbeitung/...) with its emit keywords. New test pins that the
four observed leakers now require a dom_* tag. Ghost=0, Leakage=0, coverage held
on both GTs.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
GET /projects/:id/hazards/:hid/risk-suggestion returns BreakPilot's justified
starting values for BOTH risk models per hazard:
- EN-62061-style F/W/P/S (the Excel format the professional knows)
- Fine-Kinney P/E/C (US-recognized)
each with a plain-language justification + the visible formula. Read-only and
computed from public-data anchors (ESAW/NIOSH/OSHA via the engine estimators) —
the professional adjusts the values; no norm table is stored or reproduced.
Adds EstimateFrequency (lifecycle -> 1-5) and BuildRiskSuggestion. Go SDK has no
OpenAPI baseline, so the only contract surface is the frontend consumer (the new
Risikobewertung tab, next).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Fine-Kinney (Fine 1971 / Kinney-Wiruth 1976): Risk = Probability x Exposure x
Consequence — a PUBLISHED, freely-usable method (not a DIN/Beuth/ISO standard),
widely used incl. CE-marking. Gives the professional a second, US-recognized
model alongside the EN-62061-style one; German exporters get both for free and
adjust with their own licensed norm data.
risk_fine_kinney.go: SuggestFineKinney derives justified P/E/C from public
anchors (ESAW frequency -> P, lifecycle -> E, de-biased severity -> C on the
Fine-Kinney consequence scale) + ComputeFineKinney(p,e,c) so the professional
can override with his own values. No norm table stored.
GT benchmark (rank concordance vs the professional): Fine-Kinney 75.4% — beats
the EN-62061-style model (69.3%) and the raw engine (57%).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The engine's hand-set DefaultSeverity systematically over-estimates severity
(GT shows crushing 3.3 vs 2.2, struck_by 3.1 vs 2.5; electrical was already
close). EstimateSeverity blends the pattern default 50/50 with the contact
mode's GT-calibrated typical severity (baseS) — keeps pattern-specific signal,
removes the bias. Our own model, no norm table.
Effect across both GTs: severity within +-1 78%->88%; risk RANK concordance
57%->69% (Kistenhub 45%->70%). Wired into iace_handler_init.go so the
BreakPilot risk line uses the de-biased severity.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
IP/copyright fix: ComputePLr reproduced the EN ISO 13849-1 Anhang A risk-graph
decision table (S/F/P -> PLr a..e) and SeverityToS/ExposureToF its parameter
binning, emitted into every hazard description. Removed — we may not reproduce
DIN/Beuth norm logic.
Replaced with BreakPilot's OWN risk model:
- risk_estimation.go: probability (W) + avoidance (P) estimated from public,
permissively-licensed accident statistics (Eurostat ESAW, CC BY 4.0) by
contact mode, calibrated to our ground-truth corpus; own risk index + bands.
- iace_handler_init.go now emits "Risikoeinschaetzung (BreakPilot-Modell):
S F W P -> Risiko: <level>" instead of the norm PLr string.
- DATA_SOURCES.md: data provenance + license register (ESAW CC BY 4.0; BLS/OSHA
public domain; HSE OGL; DGUV + DIN/Beuth explicitly excluded).
- gt_risk_benchmark_test.go: first GT validation of risk numbers — W within +-1
99%, P 93% vs the professional across both ground truths.
Removed risk_graph_test.go (pinned the reproduced norm table).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Three follow-ups to the 671-norm cross-reference matrix:
1. Tech-file renderer (Go): standards_applied section now gets a deterministic
Markdown appendix with the DIN/ANSI/GB/JIS mappings for the project's
suggested norms. Built from registry, never hallucinated by LLM. Applied
both to LLM and fallback content paths.
2. Frontend NormCrossRefPanel (Next.js): expandable row in the IACE library
norms tab now has a "Internationale Aequivalenzen anzeigen" button that
lazy-loads /iace/norms-library/:id/crossref and renders a colour-coded
table (relation + confidence). Region labels humanised (US — ANSI,
China (GB), Japan (JIS), etc.).
3. Contract tests (Go): 4 new handler tests pinning the response shape of
GetNormCrossRef and ListNormCrossRefs. Equivalent to an OpenAPI snapshot
for these specific endpoints — ai-compliance-sdk has no full OpenAPI
baseline yet (separate ticket).
Tests: 6 renderer tests + 4 handler contract tests, all green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Benjamin Admin