# ORCA6 v0.1-rc1 Release Notes Date: 2026-06-25 ## Summary ORCA6 v0.1-rc1 is the first local release-candidate checkpoint for the Qwen3 14B ORCA6 adapter. It is intended for guarded local testing of orchestration recommendations, RAG design advice, tool-selection tradeoffs, and safety-aware workflow planning. This is not a broad public-quality model release. The current best recipe is the refusal-tuned adapter plus the source-packet runtime guard documented in the runbook and evaluation report. ## Artifacts - Adapter: `qwen3_14b_orca_refusal_smoke/` - Merged Hugging Face model: `qwen3_14b_orca_refusal_smoke/merged/` - GGUF: `release/gguf/orca6-qwen3-14b-refusal-q8_0.gguf` - Model card: `release/MODEL_CARD.md` - Evaluation report: `release/evaluation_report.md` - Release manifest: `release/release_manifest.json` - Expanded release eval seed: `evals/orca_release_eval_prompts.jsonl` - Expanded release grounded outputs/report: `evals/qwen3_14b_orca_refusal_release_grounded_outputs.jsonl` and `evals/qwen3_14b_orca_refusal_release_grounded_report.json` - Machine-readable status: `release/status.json` - Training data audit: `data/training_data_audit.json` - Release dashboard: `release/release_dashboard.md` and `release/release_dashboard.json` - Manifest verifier: `scripts/verify_release_manifest.py` - Local guarded prompt CLI and smoke: `scripts/orca_adapter_chat.py` and `evals/qwen3_14b_orca_refusal_cli_smoke.json` - GitHub release asset sync helper: `scripts/sync_github_release_assets.py` - GGUF smoke wrapper/report: `scripts/smoke_test_gguf.py` and `release/gguf_smoke_report.json` - Hugging Face publish preflight and guarded publish sequence: `release/hf_publish_preflight.md` and `scripts/publish_hf_release.py` - Hugging Face post-publish verification: `release/hf_post_publish_verification.md` and `.json` - Release readiness rollup: `release/readiness.md` and `release/readiness.json` - Model artifact inventory: `release/model_artifact_inventory.md` and `release/model_artifact_inventory.json` - Release provenance: `release/provenance.md` and `release/provenance.json` - Release index: `release/index.md` and `release/index.json` - Tester prompt pack: `release/tester_prompt_pack.md`, `.json`, and `.jsonl` - Feedback review rubric: `release/feedback_review_rubric.md` and `.json` - v0.1 promotion checklist: `release/promotion_checklist.md` and `.json` - Feedback campaign runbook: `release/feedback_campaign_runbook.md` and `.json` - Feedback export snapshot: `release/feedback_export_snapshot.md` and `.json` - Feedback disposition log: `release/feedback_disposition_log.md` and `.json` - Tester feedback invite: `release/tester_feedback_invite.md` and `.json` - Draft final v0.1 release notes: `release/RELEASE_NOTES_v0.1_DRAFT.md` and `.json` - Final v0.1 promotion evidence map: `release/final_promotion_evidence_map.md` and `.json` - Final v0.1 promotion dry-run: `release/final_promotion_dry_run.md` and `.json` - Final v0.1 promotion guard: `scripts/promote_final_release.py` and `scripts/smoke_test_final_promotion.py` - Final v0.1 release verification: `release/final_release_verification.md` and `.json` - Final v0.1 recovery plan: `release/final_release_recovery_plan.md` and `.json` - Release issue closure guard: `release/issue_closure_guard.md` and `.json` - Release blocker matrix: `release/blocker_matrix.md` and `.json` - Release blocker trend: `release/release_blocker_trend.md` and `.json` - CI validation coverage: `release/ci_validation_coverage.md` and `.json` - Credential boundary audit: `release/credential_boundary_audit.md` and `.json` - Tracker issue audit: `release/tracker_issue_audit.md` and `.json` - External blocker handoff audit: `release/external_blocker_handoff_audit.md` and `.json` - Final gate recheck audit: `release/final_gate_recheck_audit.md` and `.json` - Release decision log audit: `release/release_decision_log_audit.md` and `.json` - Final release notes readiness audit: `release/final_release_notes_readiness_audit.md` and `.json` - Release artifact coverage audit: `release/release_artifact_coverage.md` and `.json` - External unblock packet: `release/external_unblock_packet.md` and `.json` - Status summary freshness guard: `release/status.md` count checks in `scripts/validate_release_artifacts.py` - Obsidian hardening note freshness guard: current-count checks in `scripts/validate_release_artifacts.py` - Release command audit: `release/release_command_audit.md` and `.json` - Release dependency graph: `release/release_dependency_graph.md` and `.json` - Release freshness matrix: `release/release_freshness_matrix.md` and `.json` - Release refresh order audit: `release/release_refresh_order_audit.md` and `.json` - Release invariant audit: `release/release_invariant_audit.md` and `.json` - Feedback intake: - GitHub issue forms for model feedback and groundedness failures - local Gradio app: `scripts/community_feedback_app.py` - Hugging Face Space scaffold: `spaces/orca6-feedback/` Large model artifacts are intentionally ignored by git. Use `release/release_manifest.json` to verify local sizes and SHA256 checksums. ## Validation - Retrieval smoke re-check: calibrated pass@3 20/20, calibrated pass@5 20/20, all expected@5 20/20. - Expanded guarded release eval: 54/54 overall, source packets 50/50, hard negatives 4/4. - Guarded grounded eval: 12/12 overall, source packets 8/8, hard negatives 4/4. - Unguarded grounded eval: 11/12; one empty-evidence high-risk payment prompt still invented a citation. - Local GGUF smoke: llama.cpp `b9784` CPU binary loaded the Q8_0 artifact and generated a conservative high-risk payment escalation answer. - Training data audit: 5 JSONL datasets checked with no missing prompts, missing answers, short answers, or duplicate prompts. - GitHub Actions `Validate` workflow is green for release artifact validation. - Manifest metadata verification is part of the `Validate` workflow; full local SHA256 verification is available with `scripts/verify_release_manifest.py`. - CI smoke-tests the Hugging Face publish dry-run path. - Hugging Face post-publish verification is green for `veroarc/ORCA6` and `veroarc/orca6-feedback`. - Release readiness reports local release readiness, publish readiness, and overall RC publish readiness as true. - GitHub release asset audit matches 109/109 expected attached assets. - Tester feedback invite provides a shareable reviewer handoff with release, prompt-pack, issue-form, label, and safety instructions. - Feedback status currently reports 0 labeled feedback/groundedness issues and 0 structured feedback submissions; the campaign tracker uses the `feedback-campaign` label. - Feedback export snapshot dry-runs the live GitHub export path and currently reports 0 exportable feedback rows and 0 approved-for-training rows. - Feedback disposition log records that no tester feedback has been submitted for review, so the reviewed-feedback promotion gate remains open. - Local and Hugging Face Space feedback intake now preserve tester prompt IDs in metadata and require a secrets/private-data check before training approval. - Stable local manifest verification currently checks 126 artifacts. - Final promotion evidence map ties every final `v0.1` gate to its evidence file, refresh command, and current blocked/passed status. - Final promotion dry-run is intentionally blocked until tester feedback, reviewed feedback disposition, final release notes, and final release verification are complete. - Final promotion execution is guarded by `scripts/promote_final_release.py`, which defaults to dry-run and refuses `--execute` while gates are blocked. - Final release verification is prepared and remains blocked until the final `v0.1` tag and GitHub release exist. - Final release recovery planning is prepared and contains only read-only or gated repair commands, with destructive recovery work reserved for manual release-owner escalation. - Release issue closure guard is prepared and keeps tracker issues open until their evidence reports and final release verification are ready. - Tracker issue audit verifies issues #1 and #2 remain open, labeled, and aligned with the current keep-open external blocker evidence. - External blocker handoff audit maps the completed HF publish and remaining tester feedback blockers to owners, tracker issues, evidence, commands, and closure conditions. - Final gate recheck audit confirms local gates, assets, manifest verification, invariants, and handoffs are ready while final promotion remains externally blocked. - Release decision log audit records HF publish as completed, local-safe actions as approved, and remaining gated release actions as deferred. - Final release notes readiness audit confirms final notes remain draft-blocked until feedback disposition and final promotion gates are ready. - Large artifact distribution audit confirms the adapter is published while merged Qwen3 14B shards and GGUF export remain optional large-artifact uploads behind explicit flags. - External execution dry-run audit confirms safe handoff commands are present while final promotion and issue-close execution remain gated. - Final evidence completeness audit verifies final-promotion gate statuses, source reports, refresh commands, and deferred external execution evidence. - Release provenance boundary audit verifies stable provenance fields while documenting commit and CI snapshot boundaries for checkpoint commits. - Final notes publication guard confirms RC notes are current while final `v0.1` notes remain deferred until external gates close. - Tracker closure preconditions audit confirms tracker issue close commands remain documented but blocked until final evidence and verification pass. - Release blocker matrix consolidates remaining external blockers across final promotion, feedback, and issue closure reports. - Release artifact coverage audit verifies key release reports are registered in the manifest, GitHub release asset sync list, and release index. - External unblock packet provides the ordered tester feedback, final promotion, and tracker closure sequence without executing those gates. - Status summary freshness guard checks the human-readable status count lines against the current manifest and GitHub release asset audit. - Obsidian hardening note freshness guard checks the current project-note count lines against the same machine-readable reports. - Release command audit checks handoff command references for missing scripts and records which commands are gated before final execution. - Release dependency graph records refresh dependencies and commands for generated release reports. - Release freshness matrix records generated timestamps across active and externally blocked release reports. - Release refresh order audit verifies required ordering constraints inside `scripts/refresh_release_artifacts.py`. - CI validation coverage records the release checks represented in the GitHub Actions `Validate` workflow. - Release invariant audit checks the intended locally ready, externally blocked, non-promoted release state across generated reports. - Release blocker trend classifies the remaining blocker set and records the current externally blocked baseline. - Credential boundary audit checks release-facing files for obvious credential values and records auth-sensitive commands. - Tracker issue audit checks live GitHub tracker state against release evidence before any issue closure can proceed. - External blocker handoff audit checks that each remaining external blocker has owner-action evidence before final promotion can proceed. - Final gate recheck audit verifies the final `v0.1` tag/release has not been created before external evidence is ready. - Release decision log audit keeps HF publish, feedback training merge, final promotion, and tracker closure deferred until their evidence gates change. - Final release notes readiness audit verifies `release/RELEASE_NOTES_v0.1.md` has not been created before final evidence is ready. ## Known Limitations - Requires source-packet runtime guard for citation-heavy or high-risk prompts. - Training data is still a small bootstrap set: 41 train rows and 3 validation rows. - Auto-graded preferences should be supplemented with human-reviewed labels. - Hugging Face adapter and feedback Space are published under the `veroarc` namespace; merged model and GGUF hosting remain optional large-artifact decisions. - Live GitHub release asset audit is a post-sync check and is intentionally outside the release manifest to avoid circular manifest churn. ## Next Gates - Decide whether to upload merged model and GGUF as large hosted artifacts. - Run community feedback intake on real orchestration prompts. - Convert reviewed feedback into future SFT/preference examples.