feat(iace): ESAW accident-stats RAG pipeline + real 2023 risk anchors

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>

ai-compliance-sdk/internal/api/handlers/rag_handlers.go

@@ -24,15 +24,16 @@ func NewRAGHandlers(corpusVersionStore *ucca.CorpusVersionStore) *RAGHandlers {
 
 // AllowedCollections is the whitelist of Qdrant collections that can be queried.
 var AllowedCollections = map[string]bool{
-	"bp_compliance_ce":           true,
-	"bp_compliance_gesetze":      true,
-	"bp_compliance_datenschutz":  true,
-	"bp_compliance_gdpr":         true,
-	"bp_dsfa_corpus":             true,
-	"bp_dsfa_templates":          true,
-	"bp_dsfa_risks":              true,
-	"bp_legal_templates":         true,
-	"bp_iace_libraries":          true,
+	"bp_compliance_ce":          true,
+	"bp_compliance_gesetze":     true,
+	"bp_compliance_datenschutz": true,
+	"bp_compliance_gdpr":        true,
+	"bp_dsfa_corpus":            true,
+	"bp_dsfa_templates":         true,
+	"bp_dsfa_risks":             true,
+	"bp_legal_templates":        true,
+	"bp_iace_libraries":         true,
+	"bp_iace_accident_stats":    true,
 }
 
 // SearchRequest represents a RAG search request.

ai-compliance-sdk/internal/iace/DATA_SOURCES.md

@@ -27,9 +27,16 @@ do **not** reproduce any standard's risk-graph table, decision tree or matrix.
   any generated risk-assessment export that shows engine risk numbers.
 - **URL:** https://ec.europa.eu/eurostat/statistics-explained/index.php/Accidents_at_work_-_statistics_on_causes_and_circumstances
 - **Aggregate facts used (anchor only):** contact-mode shares of accidents at
-  work, e.g. impact with stationary object ~24%, struck by moving object ~13%
-  (non-fatal) / ~24% (fatal), trapped/crushed ~14% (fatal), contact with sharp
-  agent ~15%. Retrieved 2026-06.
+  work. **Dataset `hsw_ph3_08`, reference year 2023** (Figure 7, "contact —
+  mode of injury"), EU shares:
+  - Physical/mental stress: 24.7% (non-fatal)
+  - Impact with stationary object (victim in motion): 24.0% (non-fatal) / 21.4% (fatal)
+  - Contact with sharp/pointed/rough agent: 14.5% (non-fatal)
+  - Struck by object in motion / collision: 13.0% (non-fatal) / 23.8% (fatal)
+  - Trapped / crushed: 13.8% (fatal)
+
+  Retrieved 2026-06. The source document is also ingested into the core RAG
+  collection `bp_iace_accident_stats` for searchable evidence at seeding time.
 
 ## Acceptable supplements
 

ai-compliance-sdk/internal/iace/datasources/esaw_accident_stats_2023.md

@@ -0,0 +1,57 @@
+# Accidents at work — contact mode of injury (EU, 2023)
+
+Canonical, citable source document for the IACE risk-frequency/severity anchors.
+This file is the versioned artifact that is ingested into the core RAG
+collection `bp_iace_accident_stats` so seeding can full-text search the evidence
+and surface the figure with its citation.
+
+## Primary source — Eurostat ESAW
+
+- **Source:** Eurostat — European Statistics on Accidents at Work (ESAW)
+- **Dataset:** `hsw_ph3_08` — accidents at work by contact / mode of injury
+- **Reference year:** 2023 (Statistics Explained, Figure 7)
+- **License:** CC BY 4.0 (reuse permitted, source acknowledgement required)
+- **Attribution:** `Quelle: Eurostat (ESAW) hsw_ph3_08, Bezugsjahr 2023, CC BY 4.0`
+- **Retrieved:** 2026-06
+- **URL:** https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Accidents_at_work_-_statistics_on_causes_and_circumstances
+
+### Contact mode of injury — EU shares, 2023
+
+| Contact mode | Non-fatal | Fatal |
+|---|---|---|
+| Physical or mental stress | 24.7 % | — |
+| Impact with a stationary object (victim in motion) | 24.0 % | 21.4 % |
+| Contact with a sharp / pointed / rough-coarse agent | 14.5 % | — |
+| Being struck by an object in motion / collision | 13.0 % | 23.8 % |
+| Being trapped or crushed | — | 13.8 % |
+| No contact / no information | 9.6 % | 15.1 % |
+
+Reading: the non-fatal column anchors the **frequency / probability tier (W)**
+of a contact mode; the fatal column (and the fatal-vs-non-fatal gap) anchors its
+typical **severity (S)**. Struck-by and trapped/crushed are comparatively rare
+among non-fatal but over-represented among fatal accidents — i.e. lower
+frequency, higher severity.
+
+## Supplementary context — US OSHA (public domain)
+
+- **Source:** OSHA — Commonly Used Statistics (U.S. Government work, public domain)
+- **Retrieved:** 2026-06 · **URL:** https://www.osha.gov/data/commonstats
+- 2023: **5,283** fatal work injuries in the US (3.5 per 100,000 FTE workers).
+- Most frequently violated standards: Fall Protection, Hazard Communication,
+  Control of Hazardous Energy (Lockout/Tagout).
+
+US BLS CFOI/SOII event-level tables (public domain) are an intended further
+supplement; the BLS site blocks automated retrieval, so those figures are to be
+added from a manually downloaded release.
+
+## How these numbers are used
+
+1. **Anchor (ordering):** the relative frequency/severity ordering of contact
+   modes above sets the *direction* of the W and S tiers in `risk_estimation.go`
+   (`contactModeTable`).
+2. **Calibrate (values):** tier *values* are adjusted to BreakPilot ground truth;
+   well-sampled modes use the GT mean, sparse modes use conservative defaults —
+   no overfitting to a small GT sample.
+
+No standard's risk-graph table, decision tree or SIL/PL matrix is reproduced.
+Excluded by license: DGUV statistics, DIN/Beuth/ISO/IEC tables.

ai-compliance-sdk/internal/iace/risk_data_sources.go

@@ -9,9 +9,17 @@ import "sort"
 // this table only carries the provenance so generated risk numbers are
 // auditable and correctly attributed. No raw dataset is vendored; only these
 // aggregate facts. Excluded by license: DGUV, DIN/Beuth/ISO/IEC. See
-// DATA_SOURCES.md. RAG/Qdrant ingestion is deliberately NOT used here: ~a dozen
-// stable aggregate facts are better served by a license-tagged code table than
-// by vector retrieval.
+// DATA_SOURCES.md.
+//
+// Two-layer design: this small license-tagged CODE table is the deterministic
+// tier/citation lookup (fast, stable, no nondeterminism). The underlying SOURCE
+// documents are additionally ingested into the core RAG collection
+// `bp_iace_accident_stats` so the seeding UI / an auditor can full-text search
+// the evidence and pull the original figure — the RAG is the evidence/search
+// layer, not the tier lookup.
+//
+// Figures below are the EU aggregate shares from Eurostat ESAW dataset
+// hsw_ph3_08, reference year 2023 (Figure 7, "contact - mode of injury").
 
 // RiskEvidence is the public-statistics provenance for one contact mode.
 type RiskEvidence struct {
@@ -25,9 +33,9 @@ type RiskEvidence struct {
 }
 
 const (
-	esawSource      = "Eurostat (ESAW)"
+	esawSource      = "Eurostat (ESAW, hsw_ph3_08, 2023)"
 	esawLicense     = "CC BY 4.0"
-	esawAttribution = "Quelle: Eurostat (ESAW), CC BY 4.0"
+	esawAttribution = "Quelle: Eurostat (ESAW) hsw_ph3_08, Bezugsjahr 2023, CC BY 4.0"
 	esawRetrieved   = "2026-06"
 )
 
@@ -40,10 +48,11 @@ func esawEvidence(mode, label, stat string) RiskEvidence {
 // figure is documented; other modes are anchored by the ESAW ordering and
 // GT-calibrated without a single citable share, so they carry no fabricated stat.
 var contactModeEvidence = map[string]RiskEvidence{
-	"impact_stationary": esawEvidence("impact_stationary", "Anstoßen an ruhendem Objekt", "~24 % der Arbeitsunfälle"),
-	"struck_by":         esawEvidence("struck_by", "Getroffen von bewegtem Objekt", "~13 % (nicht-tödlich) / ~24 % (tödlich)"),
-	"crushing":          esawEvidence("crushing", "Quetschen / Einklemmen", "~14 % der tödlichen Arbeitsunfälle"),
-	"cutting":           esawEvidence("cutting", "Kontakt mit scharfem Gegenstand", "~15 % der Arbeitsunfälle"),
+	"impact_stationary": esawEvidence("impact_stationary", "Anstoßen an ruhendem Objekt", "24,0 % (nicht-tödlich) / 21,4 % (tödlich)"),
+	"struck_by":         esawEvidence("struck_by", "Getroffen von bewegtem Objekt", "13,0 % (nicht-tödlich) / 23,8 % (tödlich)"),
+	"crushing":          esawEvidence("crushing", "Eingeklemmt / zerquetscht", "13,8 % (tödlich)"),
+	"cutting":           esawEvidence("cutting", "Kontakt mit scharfem/spitzem Agens", "14,5 % (nicht-tödlich)"),
+	"ergonomic":         esawEvidence("ergonomic", "Physische/psychische Belastung", "24,7 % (nicht-tödlich)"),
 }
 
 // RiskEvidenceFor returns the documented public statistic for a contact mode.