1.5.5

docs/ENVIRONMENT.md

@@ -75,6 +75,14 @@ NEUROFORGE_KB_STAGING_MIN_SOURCES=2
 NEUROFORGE_KB_STAGING_MIN_CORROBORATIONS=0
 NEUROFORGE_KB_STAGING_MAX_EVIDENCE=12
 NEUROFORGE_KB_STAGING_SYNTHESIS_MODE=llm
+NEUROFORGE_KB_STAGING_REQUIRE_AUTHORITATIVE_SOURCE=true
+NEUROFORGE_KB_STAGING_MIN_AUTHORITATIVE_SOURCES=1
+NEUROFORGE_KB_STAGING_AUTHORITATIVE_DOMAINS=
+NEUROFORGE_KB_STAGING_VERIFY_CLAIMS=true
+NEUROFORGE_KB_STAGING_MIN_CLAIM_COVERAGE=1.0
+NEUROFORGE_KB_STAGING_REQUIRE_AUTHORITATIVE_ACTIONS=true
+NEUROFORGE_KB_STAGING_MAX_VERIFICATION_STATEMENTS=24
+NEUROFORGE_KB_STAGING_VERIFICATION_REPAIR=true
 ```
 
 `NEUROFORGE_KB_STAGING_URL` and `NEUROFORGE_KB_STAGING_TOKEN` are container-internal values owned by the root Compose file. The token is derived from the existing `KB_INTEGRATION_TOKEN`; do not duplicate it under a second operator-managed secret name.
@@ -88,3 +96,11 @@ The production Compose does not use `env_file`. Agent and Knowledge receive only
 ### Staging quality gate (v1.5.1)
 
 `NEUROFORGE_KB_STAGING_SYNTHESIS_MODE=llm` is the production default. Drafts are published only after goal-relevance filtering, source diversification and successful LLM synthesis. `evidence` is a diagnostic mode that exposes the selected evidence bundle and must not be treated as an article. Off-topic search results are rejected before ingestion and do not count toward goal progress.
+
+### Production grounding gate (v1.5.5)
+
+Before an LLM-synthesized research article reaches human-review staging, NeuroForge now ranks selected evidence by source authority, requires the configured number of first-party/authoritative sources, validates critical identifiers (error codes, CVEs, KB IDs, versions and command switches) against the source bundle, and runs a second claim-verification pass. Every material draft statement must be accounted for. Prescriptive guidance must cite at least one evidence item classified as authoritative when `NEUROFORGE_KB_STAGING_REQUIRE_AUTHORITATIVE_ACTIONS=true`.
+
+The verifier is fail-closed. Unsupported or contradicted statements prevent publication. With `NEUROFORGE_KB_STAGING_VERIFICATION_REPAIR=true`, one evidence-only rewrite is attempted and the complete draft is then verified again. The repair may remove unsupported content but may not add outside knowledge. Source authority, claim-level evidence IDs, coverage, contradictions and whether a repair occurred are persisted in the staging JSON for human audit.
+
+Microsoft Q&A (`learn.microsoft.com/.../answers/...`) is intentionally treated as vendor-community rather than primary documentation. Operator-specific first-party domains can be added with `NEUROFORGE_KB_STAGING_AUTHORITATIVE_DOMAINS`.

docs/GO-LIVE-v1.5.0.md

@@ -25,3 +25,14 @@ Alternativ führt `./scripts/go-live.sh` genau diesen Ablauf ohne `make` aus. `I
 8. Vor GLPI-Schreibfreigabe einen vollständigen Ticketdurchlauf in `DRY_RUN=true` prüfen. Erst danach die gewünschten Automationen einzeln aktivieren.
 
 Docker, eine echte GLPI-Instanz, SearXNG und Ollama stehen in der Build-/Review-Umgebung nicht zur Verfügung; dieser Host-Smoke-Test ist deshalb ein bewusstes externes Release-Gate und darf nicht als lokal bestanden markiert werden.
+
+## v1.5.5 Production-Grounding Zusatzgate
+
+Vor Go-Live mit autonomem Research zusätzlich verifizieren:
+
+1. `NEUROFORGE_KB_STAGING_REQUIRE_AUTHORITATIVE_SOURCE=true` und `NEUROFORGE_KB_STAGING_VERIFY_CLAIMS=true` sind im aufgelösten Compose gesetzt.
+2. Ein Test-Goal mit explizitem Herstellerbezug erzeugt mindestens eine First-Party-Query (`site:`) und nutzt mindestens eine autoritative Quelle im finalen Staging-JSON.
+3. `claim_verification.verdict` ist `pass`, `claim_verification.coverage` ist `1`, `unsupported`/`contradictions` sind leer.
+4. `research_authoritative_sources >= 1` und `quality_gate_version=staging-v2` sind im Draft vorhanden.
+5. Ein absichtlich nicht belegter Versions-/Errorcode im Synthese-Test wird fail-closed abgewiesen.
+6. Ein Blog-/Forum-only Evidence-Set erzeugt bei aktiviertem Authority-Gate keinen Staging-Artikel.

docs/MIGRATION-v1.5.4-to-v1.5.5.md

@@ -0,0 +1,31 @@
+# Migration v1.5.4 → v1.5.5
+
+v1.5.5 is a drop-in quality-hardening update. Existing NeuroForge/Agent/Knowledge volumes, goals, research evidence and staging drafts are retained.
+
+1. Build/publish the v1.5.5 images through the normal Gitea pipeline.
+2. Set `IMAGE_TAG=1.5.5`.
+3. Keep the new production defaults enabled unless you intentionally run a diagnostic environment:
+
+```env
+NEUROFORGE_KB_STAGING_REQUIRE_AUTHORITATIVE_SOURCE=true
+NEUROFORGE_KB_STAGING_MIN_AUTHORITATIVE_SOURCES=1
+NEUROFORGE_KB_STAGING_AUTHORITATIVE_DOMAINS=
+NEUROFORGE_KB_STAGING_VERIFY_CLAIMS=true
+NEUROFORGE_KB_STAGING_MIN_CLAIM_COVERAGE=1.0
+NEUROFORGE_KB_STAGING_REQUIRE_AUTHORITATIVE_ACTIONS=true
+NEUROFORGE_KB_STAGING_MAX_VERIFICATION_STATEMENTS=24
+NEUROFORGE_KB_STAGING_VERIFICATION_REPAIR=true
+```
+
+4. Pull/recreate NeuroForge and the worker (or the complete stack):
+
+```bash
+docker compose --profile research pull
+docker compose --profile research up -d --force-recreate neuroforge neuroforge-worker
+```
+
+5. Do not delete volumes. Existing low-quality staging drafts created by older releases should be reviewed/deleted manually; v1.5.5 does not silently rewrite previously stored draft content.
+
+## Expected behavioral change
+
+A goal may now remain at `0/1 Staging-Artikel` even with many learned web memories if the selected bundle has no authoritative source or if the verifier finds unsupported/contradicted claims. That is the intended fail-closed production behavior. The goal's `last_staging_error` explains the blocked gate.

docs/RELEASE-v1.5.5.md

@@ -0,0 +1,36 @@
+# GLPI NeuroForge Mega v1.5.5
+
+v1.5.5 turns the autonomous Research → Staging path into a production-grounded pipeline. The release follows a live Windows 11 DISM `0x800f081f` test where the technical pipeline completed but the synthesized draft recommended the wrong SFC/DISM order and relied mostly on non-primary web sources.
+
+## Production quality gates
+
+- **Source authority ranking:** selected evidence is ordered by first-party/vendor authority before confidence/recency while retaining source diversity.
+- **Authoritative-source requirement:** production Compose requires at least one authoritative source in the evidence actually supplied to the synthesizer. Operator-specific domains can be added without replacing the built-in vendor set.
+- **Vendor-community distinction:** Microsoft Q&A under `learn.microsoft.com/.../answers/...` remains useful corroboration but is not counted as primary Microsoft documentation.
+- **Authority-aware query planning:** deterministic goal research reserves a first-party `site:` query for recognized vendors such as Microsoft, Fortinet, NVIDIA, Cisco, Broadcom/VMware, Red Hat, Ubuntu and Apple.
+- **Critical identifier guard:** newly synthesized error codes, CVEs, Microsoft KB IDs, versions and command switches must already exist in the selected evidence bundle.
+- **Claim → Evidence verification:** a separate critic pass receives numbered draft statements (`S*`) and numbered evidence (`E*`) and must account for every material statement with explicit evidence IDs.
+- **Authoritative action gate:** commands/recommendations require at least one authoritative evidence item when the production default is enabled.
+- **Contradiction handling:** unsupported or contradicted statements fail closed. One evidence-only grounding rewrite may remove/correct unsupported content; the entire rewritten draft is then verified again.
+- **Auditable JSON:** Staging metadata now persists source authority, independent origins, authoritative-source count, claim coverage, statement evidence IDs, contradictions and repair status.
+- **Independent corroboration semantics:** multiple URLs/pages on the same registrable source origin no longer inflate `research_corroborations` as independent confirmation.
+
+## Defaults
+
+The root Compose enables the production gates by default:
+
+```env
+NEUROFORGE_KB_STAGING_REQUIRE_AUTHORITATIVE_SOURCE=true
+NEUROFORGE_KB_STAGING_MIN_AUTHORITATIVE_SOURCES=1
+NEUROFORGE_KB_STAGING_VERIFY_CLAIMS=true
+NEUROFORGE_KB_STAGING_MIN_CLAIM_COVERAGE=1.0
+NEUROFORGE_KB_STAGING_REQUIRE_AUTHORITATIVE_ACTIONS=true
+NEUROFORGE_KB_STAGING_MAX_VERIFICATION_STATEMENTS=24
+NEUROFORGE_KB_STAGING_VERIFICATION_REPAIR=true
+```
+
+`NEUROFORGE_KB_STAGING_AUTHORITATIVE_DOMAINS` is additive. Keep it empty unless the deployment uses first-party documentation hosts not covered by the built-ins.
+
+## Governance
+
+The bridge still writes **only** human-review staging. The Knowledge integration enforces `auto_reply=false`; production Knowledge promotion remains a human action.