1.5.1

docs/ENVIRONMENT.md

@@ -74,6 +74,7 @@ NEUROFORGE_KB_STAGING_MIN_EVIDENCE=4
 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_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.
@@ -83,3 +84,7 @@ This bridge can only create/update **human-review staging**. The Knowledge servi
 ## Production secret isolation
 
 The production Compose does not use `env_file`. Agent and Knowledge receive only explicit runtime variables. NeuroForge Admin/Worker/Metrics, Knowledge editor and Control Center credentials are therefore not broadly inherited by unrelated containers. Local source builds use the separate `docker-compose.dev.yml` override.
+
+### 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.

docs/MIGRATION-v1.5.0-to-v1.5.1.md

@@ -0,0 +1,47 @@
+# Migration v1.5.0 → v1.5.1
+
+v1.5.1 is a stop-the-line research/staging quality fix. It is intended to be a drop-in image/config upgrade from v1.5.0; persistent volumes are retained.
+
+## Required environment
+
+Keep the existing v1.5.0 secrets and add or confirm:
+
+```env
+IMAGE_TAG=1.5.1
+NEUROFORGE_KB_STAGING_SYNTHESIS_MODE=llm
+```
+
+For an external Ollama host, v1.5.1 again honors:
+
+```env
+OLLAMA_BASE_URL=http://your-ollama:11434
+OLLAMA_URLS=http://your-ollama:11434
+```
+
+If these variables are omitted, the root Compose defaults to the internal `http://ollama:11434` service.
+
+## Behaviour changes
+
+- Goal research rejects search/page material that has no subject anchor overlap with the goal before it is learned.
+- Research memories persist `provenance.goal_id`; legacy `goal:<id>` tags remain supported for reconciliation.
+- Goal evidence/source counters are recomputed from relevant persisted evidence, so counters can decrease after upgrade when old off-topic evidence is removed from the goal view.
+- Draft evidence is diversified across independent sources and limited to at most two chunks per source after the diversity pass.
+- Production staging uses LLM synthesis even when `autonomy.use_llm=false`. Raw evidence concatenation is no longer the default article path.
+- Failed, empty or off-topic synthesis is fail-closed: no staging article is created/updated.
+- Targets containing `Artikel`/`article` count created staging drafts instead of evidence chunks.
+- The standalone NeuroForge worker Compose no longer passes its token as a CLI argument.
+
+## Existing bad staging drafts
+
+Existing drafts are not automatically promoted or deleted. For a bad draft created by v1.5.0, leave it in staging or delete it manually. The next successful cycle for the same goal can update the same integration key with a newly synthesized draft.
+
+## Upgrade
+
+```sh
+docker compose --profile research pull
+docker compose --profile research up -d --force-recreate
+```
+
+Do not use `down -v`; the persistent NeuroForge/Agent/Ollama volumes must be retained.
+
+After startup, verify `/readyz`, then inspect the goal counters. Off-topic historical evidence may disappear from the per-goal counts by design.

docs/RELEASE-v1.5.1.md

@@ -0,0 +1,7 @@
+# GLPI NeuroForge Mega v1.5.1
+
+v1.5.1 fixes a research-quality failure discovered during the first real autonomous staging run: an NVIDIA goal could ingest an unrelated MDN WebRTC page and publish raw page chunks as a supposed knowledge article.
+
+The release adds a fail-closed relevance gate, durable goal provenance, relevant-evidence reconciliation, source-diverse draft selection, mandatory production LLM article synthesis, output relevance validation, correct article-target progress semantics, external Ollama Compose support, and the remaining worker-token Compose cleanup.
+
+Local release gates: all four Go modules pass `go test ./...`, `go vet ./...`, `go build ./...`, and the targeted race suites from `scripts/release-gate.sh`.