pen-stack 3.1.0__tar.gz → 3.2.0__tar.gz

{pen_stack-3.1.0 → pen_stack-3.2.0}/CHANGELOG.md

@@ -1,202 +1,237 @@
-# Changelog
-
-All notable changes to PEN-STACK are documented here. This file follows
-[Keep a Changelog](https://keepachangelog.com/) and the program's phase structure.
-
-## [3.1.0] - 2026-06-04 - v3.1 release: publishable contributions + an adopted benchmark
-
-The v3.1 cycle completes (workstreams A-H). It hardens the honesty of the planning benchmark, surrounds the
-models with strong baselines, adds a predicted-structure safety axis, and ships the first benchmark and
-grounded agent for the genome-writing side. Every workstream is pre-registered (`prereg/ws_*.yaml`,
-SHA-locked) and reports its honest negatives.
-
-### Added
-- **WS-B - strong baselines + safety primary-metric switch.** Endogenous-expression baseline (TRIP-trained
-  Spearman 0.51 vs AlphaGenome ES-Bruce4 proxy 0.43), multi-mark ablation (all-marks >= best single), and a
-  published GSH rule-set: safe-harbour discrimination (learned 0.92, 95% CI [0.82, 0.98] vs distance-rule
-  0.38, delta CI excludes zero) is now the primary safety metric; the circular `genotoxic_cis` AUROC is a
-  labeled diagnostic. (`pen_stack.wgenome.gsh_baseline`, `pen_stack.validate.durability_baselines`.)
-- **WS-C - AlphaGenome integration.** Hosted-API provider with an offline cache; predicted-vs-measured track
-  validation (HepG2 ATAC Pearson 0.85) with an honest score-level low-confidence flag; a 3D structural-risk
-  axis from contact-map deltas (`pen_stack.wgenome.{providers,chromatin_seq,structure3d}`,
-  `pen_stack.validate.seq_vs_measured`).
-- **WS-D - Cargo Polish.** Cargo-sequence silencing-risk scan (`pen_stack.planner.cargo_polish`).
-- **WS-E - Genome-Writing Bench v0.1 + PEN-Agent.** The first writing-side benchmark (`benchmarks/`,
-  `bench/run.py`) with deterministic scorers, a leaderboard, and a real LLM-agent baseline; a grounded
-  write-planning state machine with a no-fabrication hard gate (`pen_stack.agent.pen_agent`).
-- **WS-F - local recalibration / private-data adaptation.** Gated recalibration / fine-tuning on private
-  data, in-container; the adapted model activates only if it beats the released model AND a no-skill
-  baseline; the released model is provably unchanged (`pen_stack.adapt`).
-- **WS-G - multiplex + guide QC.** A pairwise translocation-risk screen (`pen_stack.planner.multiplex`,
-  surfaced in PEN-Agent) and a bridge-RNA guide ranker (`pen_stack.bridge.guide_qc`).
-- **WS-H - release + dissemination.** README/badges updated for v3.1, `docs/quickstart.md`,
-  `docs/positioning.md`, the leaderboard submission guide, the dissemination log, and version 3.1.0.
-
-### Changed (honesty)
-- The planning benchmark's `recovery_at_k` ranking is now deterministic (stable sort + tie-breakers).
-- The LLM stack defaults to the local Ollama model on the compute tier with an automatic hosted-Nemotron
-  fallback, a cooldown cache, and bounded timeouts (no more multi-minute stalls when a provider is absent).
-
-## [3.1.0a0] - 2026-06-04 - v3.1 WS-A: de-circularize the planning benchmark (gate)
-
-The v3.1 cycle (publishable contributions + an adopted benchmark) opens with its gate: de-circularizing the
-Phase-3 planning benchmark before anything builds on it.
-
-### Changed (honesty)
-- **The Phase-3 "discriminating-stratum recovery@10 = 1.00 vs 0.00 (McNemar p, CI)" is now labeled
-  definitional, not predictive,** everywhere (README, manuscript abstract, `prereg/paper3.yaml`,
-  `validate/paper3_benchmark.py` docstring). An on-target identity term dominates the score, so the planner
-  ranks the goal's own gene first by construction. Documented in `docs/benchmark_circularity.md` (WS-A1).
-- The intent result is reframed as a **specification-compliance correctness table** (`validate/intent_specification.py`,
-  7/7), with no recovery/p-value/CI language (WS-A2).
-
-### Added (the honest, non-circular replacements)
-- **Blind safe-harbour site discovery (the new headline)**: `validate/blind_gsh_discovery.py` +
-  `configs/gsh_validated_heldout.yaml` (5 DOI-validated held-out GSH, gene-anchored to hg38) +
-  frozen/SHA-locked `data/gsh_matched_controls.parquet`. Run genome-wide (no on-target term), the planner's
-  writability separates validated GSH from matched-context controls at **AUROC 0.92** (safety-only 0.50)
-  (WS-A3).
-- **Diversified writer-family recovery**: `validate/writer_recovery.py` + `data/writer_panel.csv` (8 writes,
-  4 families, DOIs). recovery@1 = **1.0** vs prevalence 0.25 (smallest-capacity DSB-free writer that fits
-  the cargo) (WS-A4).
-- **Within-locus ranking** (descriptive): `validate/within_locus_ranking.py` - AAVS1 documented bin at the
-  93rd within-locus percentile (top quartile); CLYBL at the 34th (honest negative) (WS-A5).
-- **Consolidated report** `scripts/p3_benchmark_report.py` -> `out/ws_a_report.md`; `prereg/ws_a.yaml` +
-  SHA lock. Gate G-A is met: blind AUROC reported, no circular claims remain (WS-A6).
-
-## [Unreleased] - 2026-06-03 - honest reframing, repository polish, coverage, hybrid LLM
-
-### Added
-- **Hybrid LLM backend** (`pen_stack/rag/llm.py`, `configs/llm.yaml`): a strong hosted model for
-  reasoning/agent/Q&A (NVIDIA Nemotron, OpenAI-compatible, free) with **automatic fallback** to the local
-  Ollama model, then to the deterministic no-LLM path. One `provider` switch. The agent and RAG were
-  refactored onto a single provider-agnostic `chat()` (NVIDIA tool-call IDs and Ollama native message
-  threading both handled). The LLM stays non-load-bearing - every number/citation still comes from
-  validated tools - so the model choice does not affect scientific reproducibility; it only improves
-  orchestration (Nemotron planned a goal in 2 tool calls vs the local 7B's 8-call loop). Core scientific
-  compute stays local/VM and uses no LLM. API keys are read from an env var or a **gitignored** file and
-  are never committed.
-
-### Changed
-- **Paper 4 reframed to its honest scope.** `pen-bridge` is positioned as the first measured-data-validated
-  tool that **nominates and ranks candidate off-target *locations*** for bridge recombinases - a
-  **screening tool, not a quantitative safety calculator**. The AUROC 0.77 vs 0.62 result is stated with
-  its caveat (favourable negative set; mostly tests core integrity), and the magnitude limitation
-  (sequence-risk does not rank recombination amount, rho ~0.30) is named as the single most important
-  limitation. Application-Note tier, first-of-its-kind for an unoccupied gap; the Writable Genome remains
-  the flagship. Manuscript + `prereg/paper4.yaml` + summaries updated.
-- **Variant-effect reframed:** the DMS recovers KNOWN enhancers (a catalogue feature), it is not a novel
-  variant-design method; EVOLVEpro is the engine to wrap when generating new variants.
-- **Repository made clean ASCII:** removed all decorative emojis and em/en dashes and other non-ASCII
-  punctuation across code, docs, configs, and manuscripts (box-drawing tree characters kept).
-
-### Added
-- 72-system ortholog characterisation (`bridge/ortholog_screen.py`) - explicitly DESCRIPTIVE (Table S1 has
-  no activity label): sequence-similarity organisation vs the validated standout ISCro4 (IS621 ranks most
-  similar, a sanity check). Exploratory secondary result, N ~72.
-- Coverage: CI runs `pytest --cov`, uploads to Codecov, and publishes a self-hosted coverage badge
-  (`tools/make_coverage_badge.py` -> `.github/badges/coverage.svg`). Unit-test coverage of the core logic
-  is **69%** (integration-only modules that need GPU/VM/network/LLM are excluded via `[tool.coverage.run]`).
-- Professional, emoji-free README with connected-repo badges (genome-atlas / mech-class / pen-score /
-  pen-assemble / pen-compare), an architecture diagram, and the problem/gaps explanation.
-
-## [3.0.0a5] - 2026-06-02 - Phase 1.5 (Bridge-recombinase off-target engine -> Paper 4, BEACHHEAD)
-
-The first public instrument: a bridge-recombinase off-target screening tool.
-
-### Added
-- **Off-target engine** (`pen_stack/bridge/offtarget.py` + `configs/bridge_offtarget_profile.yaml`):
-  genome-wide hg38 pseudosite scan (CT-core seed, per-chromosome, memory-bounded) + a position-weight
-  risk model grounded in the published mechanism. **Beats naive Hamming: AUROC 1.00 vs 0.59** at
-  separating core-preserving (real-risk) from core-disrupting (abolished) sites. Exposes
-  `predict_offtargets(family, site)` - completes the Phase-3 Planner cargo hook.
-- **Fold / cross-loop QC** (`bridge/fold_qc.py`): ViennaRNA fold (verified MFE on a 190-nt design) +
-  TBL/DBL cross-loop complementarity.
-- **Activity framework** (`bridge/activity.py`): exploratory DMS + 72-system trainer (deferred; data paywalled).
-- **`pen-bridge`** (`bridge/pipeline.py`, `bridge/cli.py`, `/bridge/design` API): **wraps** the Arc
-  BridgeRNADesigner (verified) and adds the off-target + QC layer.
-- `validate/paper4_validation.py` + `scripts/p4_genome_scan.py`; `prereg/paper4.yaml` + SHA lock.
-
-### Notes
-- **Phase 1.5 COMPLETE** - pre-registered criteria met (or honestly gated): the off-target engine,
-  ViennaRNA fold, and designer wrap are verified on the VM (real hg38 scan: chr22 in ~21 s). The *blind
-  recall of Perry 2025's measured off-targets* and the DMS/activity model are gated on the paywalled
-  Perry 2025 supplementary (drop in via `ingest.load_offtarget_profile`). Completes the deferred Phase-2
-  Section 2.4 and Phase-3 Section 3.2 hooks. 68 tests green; ruff clean. **All program phases (0,1,1.5,2,3) now done.**
-
-## [3.0.0a4] - 2026-06-02 - Phase 3 (The Write Planner + agentic platform -> Paper 3, CAPSTONE)
-
-Inverse design + the paper-defining recovery@k benchmark + the agentic platform.
-
-### Added
-- **Inverse-design optimiser** (`pen_stack/planner/optimize.py`, `configs/intent_weights.yaml`): an
-  `edit_intent`-conditioned objective whose `target_gene_sign` flips whether hitting the target gene is
-  penalised or rewarded - the same TRAC site ranks #1 (knock-in) vs #101 (safe-harbour).
-- **Cargo/delivery** (`planner/cargo.py`, `planner/delivery.py`): donor spec + size check + delivery rule
-  table; bridge/seek off-target via an optional Phase-1.5 hook (pending until 1.5).
-- **End-to-end Planner** (`planner/pipeline.py`, `report.py`, `/plan` API, `pen-stack plan` CLI): ranked,
-  fully traceable plans with per-field provenance.
-- **Two-stratum recovery@k benchmark** (`validate/paper3_benchmark.py`, `data/benchmark_panel.csv`,
-  `prereg/paper3.yaml`): **discriminating stratum planner 1.00 vs baseline 0.00, McNemar p=0.0156, gap CI
-  [1.0,1.0] excludes zero; control tie 0.67=0.67**. Panel cited to Europe-PMC-verified sources.
-- **Forward hypotheses** (`validate/forward_hypotheses.py`): date-stamped novel F8/SERPINA1/CISH/HBA1
-  proposals + grounded cited ranking.
-- **Agentic platform**: `agent/tools.py` + `agent/orchestrator.py` (Ollama tool-calling, auditable trace,
-  no-fabrication, refusals), `agent/mcp_server.py` (fastmcp), `docker-compose.yml` + `docker/ui.Dockerfile`
-  + Streamlit **Agent** page + `docs/DEPLOY.md`/`docs/MCP.md`, `validate/agent_eval.py`.
-- Shipped `data/curated/gene_coords.parquet` (GENCODE-derived) so tools work in any container.
-
-### Notes
-- **Phase 3 COMPLETE** - pre-registered criteria met (`prereg/paper3.yaml` + `SHA256_LOCK_phase3.json`).
-  Agent verified on the VM in LLM mode (no-fabrication + plan-equivalence + refusals all pass). 63 tests
-  green; ruff clean. Wet-lab (3.7) skipped - non-gating. Bridge off-target hook completes with Phase 1.5.
-
-## [3.0.0a3] - 2026-06-02 - Phase 2 (Writer Atlas + Unified Stack -> Paper 2)
-
-The broad, cross-family Writer Atlas, the writer<->locus cross-link, and the installable platform.
-
-### Added
-- **Writer Atlas** (`pen_stack/atlas/expand.py`, `atlas.parquet`): **33,370 systems across 8 families**
-  (31,885 IS110/IS1111 orthologs + curated cores/reps), every row confidence-tagged + >=1 source DOI,
-  targeting metadata inherited from the WT-KB. `configs/atlas_families.yaml` drives the UniProt queries.
-- **Mechanism at scale** (`pen_stack/mech/`): ported audited 18-family Pfam whitelist v1.2.1; composite
-  co-occurrence rules; **core agreement 1.00** vs audited labels; conflicting calls -> review queue.
-- **Therapeutic readiness** (`pen_stack/score/therapeutic.py`): deliverability/cargo/human-cell axes,
-  components retained (ISCro4 326aa->AAV).
-- **Cross-link** (`pen_stack/atlas/crosslink.py`): bidirectional writer<->locus queries; AAVS1 held-out
-  check passes (0.90 writability + bridge-reachable). Per-family caches for k562/hepg2/hspc.
-- **Variant proposal** (`pen_stack/atlas/variant_propose.py`): point-mutation framework + retrospective
-  harness, no chimeras; DMS model pluggable (Phase 1.5).
-- **PEN-MONITOR** (`pen_stack/monitor/`): Europe PMC living-database engine; back-test surfaces ISPpu10;
-  never auto-edits the atlas; every candidate cited.
-- **Grounded RAG** (`pen_stack/rag/`, `pen_stack/agent/guardrails.py`): numbers from tool calls, claims
-  cited, clinical directives refused; optional Ollama/Qwen phrasing layer (presentation only).
-- **Stack**: unified CLI subcommands, FastAPI server (`pen_stack/server/api.py`), Streamlit platform UI
-  (Writer Atlas + Ask pages), mkdocs site + 4 use-case tutorials. 46 tests green; ruff clean.
-
-### Notes
-- **Phase 2 COMPLETE** - pre-registered criteria met (`prereg/paper2.yaml` + `SHA256_LOCK_phase2.json`);
-  atlas Zenodo DOI pending author upload. Verified on the VM (Docker): API, UI (:8501), RAG with Qwen.
-
-## [3.0.0a0] - 2026-06-01 - Phase 0 (in progress)
-
-Fresh v3.0 monorepo. Supersedes the v1.0 platform repository (archived); consolidates the five prior
-repositories (`genome-atlas`, `mech-class`, `pen-score`, `pen-assemble`, `pen-compare`) as provenance.
-
-### Added
-- Monorepo scaffold: 13 modules (`atlas`, `mech`, `score`, `wgenome`, `planner`, `bridge`, `monitor`,
-  `rag`, `agent`, `ui`, `data`, `validate`, `server`), `pyproject.toml`, Docker image spec, `penctl`
-  laptop<->VM orchestrator, CI, `configs/`, `prereg/`.
-- `docs/INFRA.md` - three-tier (laptop / VM / Drive) Docker-only, SFTP-only workflow.
-- `configs/llm.yaml` - single LLM switch (Ollama + Qwen2.5-7B-Instruct, Apache-2.0).
-- `configs/datasets.yaml` - pinned dataset accessions + verified IDs (see VERIFICATION_REPORT_v3.0).
-
-- **WT-KB** (`pen_stack/atlas/`): 8 fully-sourced writer families with reachability tiers; schema enforces the >=1-DOI sourcing rule.
-- **Re-grounded axes** (`pen_stack/score/recalibrate.py`, `configs/score_axes.yaml`): `S_Cargo` from measured bp, `S_Prog` from targeting modality, `length_aa` backfilled - no per-enzyme overrides.
-- **Canonical universe** (`pen_stack/atlas/universe.py::assemble`): one path joining the 1,058-entity universe + WT-KB + crosswalk; cross-module consistency test.
-- **Descriptive scorecard** (`pen_stack/atlas/scorecard.py`): reframed from the circular certification; blind concordance recovers ISCro4 as the bridge standout without naming it. 21 tests green.
-
-### Notes
-- Independent verification of all datasets/IDs/DOIs/tools completed: no critical errors in the v3.0 plan
-  (full report in `Final_Part_v3.0/VERIFICATION_REPORT_v3.0.md`).
-- **Phase 0 COMPLETE** - all pre-registered success criteria met (`prereg/phase0.yaml` + SHA lock).
+# Changelog
+
+All notable changes to PEN-STACK are documented here. This file follows
+[Keep a Changelog](https://keepachangelog.com/) and the program's phase structure.
+
+## [3.2.0] - 2026-06-08 - v3.2 release: a calibrated, self-aware co-scientist
+
+The v3.2 cycle makes the genome-writing funnel **trustworthy**: every value carries a calibrated confidence,
+an extrapolation flag, and — where the biology is beyond any tool here — an explicit out-of-scope deferral.
+Workstreams UQ/EP/MC/BA, each pre-registered (`prereg/ws_{uq,ep,mc,ba}.yaml`, SHA-locked) and reporting its
+honest negatives. The Genome-Writing Bench bumps to **v0.2**.
+
+### Added
+- **WS-UQ - calibrated uncertainty + OOD.** Conformal prediction intervals (durability expression) and APS /
+  Mondrian prediction sets (safety, silenced) wrapping the existing heads with no retraining
+  (`pen_stack.wgenome.uncertainty`); an OOD detector that widens intervals out-of-distribution
+  (`pen_stack.wgenome.ood`); selective prediction + plan-level confidence
+  (`pen_stack.validate.selective_prediction`). Held-out coverage 0.895 vs 0.90 nominal; risk-coverage accuracy
+  0.739->0.930 under abstention. OOD across human cell types is weak (0.65-0.73) - reported as a heuristic.
+- **WS-EP - epistemic scoping.** A three-tier status (grounded-confident / grounded-extrapolating /
+  not-computable) on every agent output (`pen_stack.agent.epistemic`); a known-unknowns registry + scope
+  matcher (`configs/known_unknowns.yaml`, `pen_stack.agent.scope`, `docs/scope.md`) that defers out-of-scope
+  questions (deferral 1.0, false-defer 0.0); abstention in the agent. No-fabrication gate intact.
+- **WS-MC - mechanistic filters.** A hard target-site/PAM/att-site reachability reject
+  (`pen_stack.planner.target_site`, `configs/target_sites.yaml`; controls 9/9); vehicle-specific
+  delivery-sequence penalties (`pen_stack.planner.delivery_constraints`); and an off-target **energetics**
+  model (`pen_stack.bridge.offtarget_energetics`) that beats the 0.77 baseline at held-out AUROC 0.88 (robust
+  over 5 seeds) and ships as the default ranker.
+- **WS-BA - bench v0.2 + uncertainty-aware agent.** Four trust tasks (T8 calibration, T9 selective prediction,
+  T10 OOD honesty, T11 out-of-scope refusal) contrasting the uncertainty-aware agent with an over-confident
+  baseline (4/4); PEN-Agent emits confidence + epistemic status + abstains; UI surfaces them. Bench re-SHA-locked.
+- **WS-OPT1 (optional) - Gymnasium interface.** A thin `gymnasium.Env` over the planner (`pen_stack.env`,
+  `[env]` extra) for agent-developer interoperability - interface only, no RL superiority claimed.
+- **Docs:** `docs/uncertainty.md`, `docs/scope.md`, `docs/mechanistic_constraints.md`; M-UQ methods note +
+  M1/M2 manuscript updates. WS-OPT2 (Opentrons) deferred to `docs/BACKLOG.md`.
+
+### Changed
+- Version 3.1.0 -> 3.2.0 (pyproject, `__init__`, CITATION.cff). README "what is new in v3.2"; badges + bench
+  v0.2. The bridge off-target default ranker is now the energetics model when its penalty table is present.
+
+## [3.1.0] - 2026-06-04 - v3.1 release: publishable contributions + an adopted benchmark
+
+The v3.1 cycle completes (workstreams A-H). It hardens the honesty of the planning benchmark, surrounds the
+models with strong baselines, adds a predicted-structure safety axis, and ships the first benchmark and
+grounded agent for the genome-writing side. Every workstream is pre-registered (`prereg/ws_*.yaml`,
+SHA-locked) and reports its honest negatives.
+
+### Added
+- **WS-B - strong baselines + safety primary-metric switch.** Endogenous-expression baseline (TRIP-trained
+  Spearman 0.51 vs AlphaGenome ES-Bruce4 proxy 0.43), multi-mark ablation (all-marks >= best single), and a
+  published GSH rule-set: safe-harbour discrimination (learned 0.92, 95% CI [0.82, 0.98] vs distance-rule
+  0.38, delta CI excludes zero) is now the primary safety metric; the circular `genotoxic_cis` AUROC is a
+  labeled diagnostic. (`pen_stack.wgenome.gsh_baseline`, `pen_stack.validate.durability_baselines`.)
+- **WS-C - AlphaGenome integration.** Hosted-API provider with an offline cache; predicted-vs-measured track
+  validation (HepG2 ATAC Pearson 0.85) with an honest score-level low-confidence flag; a 3D structural-risk
+  axis from contact-map deltas (`pen_stack.wgenome.{providers,chromatin_seq,structure3d}`,
+  `pen_stack.validate.seq_vs_measured`).
+- **WS-D - Cargo Polish.** Cargo-sequence silencing-risk scan (`pen_stack.planner.cargo_polish`).
+- **WS-E - Genome-Writing Bench v0.1 + PEN-Agent.** The first writing-side benchmark (`benchmarks/`,
+  `bench/run.py`) with deterministic scorers, a leaderboard, and a real LLM-agent baseline; a grounded
+  write-planning state machine with a no-fabrication hard gate (`pen_stack.agent.pen_agent`).
+- **WS-F - local recalibration / private-data adaptation.** Gated recalibration / fine-tuning on private
+  data, in-container; the adapted model activates only if it beats the released model AND a no-skill
+  baseline; the released model is provably unchanged (`pen_stack.adapt`).
+- **WS-G - multiplex + guide QC.** A pairwise translocation-risk screen (`pen_stack.planner.multiplex`,
+  surfaced in PEN-Agent) and a bridge-RNA guide ranker (`pen_stack.bridge.guide_qc`).
+- **WS-H - release + dissemination.** README/badges updated for v3.1, `docs/quickstart.md`,
+  `docs/positioning.md`, the leaderboard submission guide, the dissemination log, and version 3.1.0.
+
+### Changed (honesty)
+- The planning benchmark's `recovery_at_k` ranking is now deterministic (stable sort + tie-breakers).
+- The LLM stack defaults to the local Ollama model on the compute tier with an automatic hosted-Nemotron
+  fallback, a cooldown cache, and bounded timeouts (no more multi-minute stalls when a provider is absent).
+
+## [3.1.0a0] - 2026-06-04 - v3.1 WS-A: de-circularize the planning benchmark (gate)
+
+The v3.1 cycle (publishable contributions + an adopted benchmark) opens with its gate: de-circularizing the
+Phase-3 planning benchmark before anything builds on it.
+
+### Changed (honesty)
+- **The Phase-3 "discriminating-stratum recovery@10 = 1.00 vs 0.00 (McNemar p, CI)" is now labeled
+  definitional, not predictive,** everywhere (README, manuscript abstract, `prereg/paper3.yaml`,
+  `validate/paper3_benchmark.py` docstring). An on-target identity term dominates the score, so the planner
+  ranks the goal's own gene first by construction. Documented in `docs/benchmark_circularity.md` (WS-A1).
+- The intent result is reframed as a **specification-compliance correctness table** (`validate/intent_specification.py`,
+  7/7), with no recovery/p-value/CI language (WS-A2).
+
+### Added (the honest, non-circular replacements)
+- **Blind safe-harbour site discovery (the new headline)**: `validate/blind_gsh_discovery.py` +
+  `configs/gsh_validated_heldout.yaml` (5 DOI-validated held-out GSH, gene-anchored to hg38) +
+  frozen/SHA-locked `data/gsh_matched_controls.parquet`. Run genome-wide (no on-target term), the planner's
+  writability separates validated GSH from matched-context controls at **AUROC 0.92** (safety-only 0.50)
+  (WS-A3).
+- **Diversified writer-family recovery**: `validate/writer_recovery.py` + `data/writer_panel.csv` (8 writes,
+  4 families, DOIs). recovery@1 = **1.0** vs prevalence 0.25 (smallest-capacity DSB-free writer that fits
+  the cargo) (WS-A4).
+- **Within-locus ranking** (descriptive): `validate/within_locus_ranking.py` - AAVS1 documented bin at the
+  93rd within-locus percentile (top quartile); CLYBL at the 34th (honest negative) (WS-A5).
+- **Consolidated report** `scripts/p3_benchmark_report.py` -> `out/ws_a_report.md`; `prereg/ws_a.yaml` +
+  SHA lock. Gate G-A is met: blind AUROC reported, no circular claims remain (WS-A6).
+
+## [Unreleased] - 2026-06-03 - honest reframing, repository polish, coverage, hybrid LLM
+
+### Added
+- **Hybrid LLM backend** (`pen_stack/rag/llm.py`, `configs/llm.yaml`): a strong hosted model for
+  reasoning/agent/Q&A (NVIDIA Nemotron, OpenAI-compatible, free) with **automatic fallback** to the local
+  Ollama model, then to the deterministic no-LLM path. One `provider` switch. The agent and RAG were
+  refactored onto a single provider-agnostic `chat()` (NVIDIA tool-call IDs and Ollama native message
+  threading both handled). The LLM stays non-load-bearing - every number/citation still comes from
+  validated tools - so the model choice does not affect scientific reproducibility; it only improves
+  orchestration (Nemotron planned a goal in 2 tool calls vs the local 7B's 8-call loop). Core scientific
+  compute stays local/VM and uses no LLM. API keys are read from an env var or a **gitignored** file and
+  are never committed.
+
+### Changed
+- **Paper 4 reframed to its honest scope.** `pen-bridge` is positioned as the first measured-data-validated
+  tool that **nominates and ranks candidate off-target *locations*** for bridge recombinases - a
+  **screening tool, not a quantitative safety calculator**. The AUROC 0.77 vs 0.62 result is stated with
+  its caveat (favourable negative set; mostly tests core integrity), and the magnitude limitation
+  (sequence-risk does not rank recombination amount, rho ~0.30) is named as the single most important
+  limitation. Application-Note tier, first-of-its-kind for an unoccupied gap; the Writable Genome remains
+  the flagship. Manuscript + `prereg/paper4.yaml` + summaries updated.
+- **Variant-effect reframed:** the DMS recovers KNOWN enhancers (a catalogue feature), it is not a novel
+  variant-design method; EVOLVEpro is the engine to wrap when generating new variants.
+- **Repository made clean ASCII:** removed all decorative emojis and em/en dashes and other non-ASCII
+  punctuation across code, docs, configs, and manuscripts (box-drawing tree characters kept).
+
+### Added
+- 72-system ortholog characterisation (`bridge/ortholog_screen.py`) - explicitly DESCRIPTIVE (Table S1 has
+  no activity label): sequence-similarity organisation vs the validated standout ISCro4 (IS621 ranks most
+  similar, a sanity check). Exploratory secondary result, N ~72.
+- Coverage: CI runs `pytest --cov`, uploads to Codecov, and publishes a self-hosted coverage badge
+  (`tools/make_coverage_badge.py` -> `.github/badges/coverage.svg`). Unit-test coverage of the core logic
+  is **69%** (integration-only modules that need GPU/VM/network/LLM are excluded via `[tool.coverage.run]`).
+- Professional, emoji-free README with connected-repo badges (genome-atlas / mech-class / pen-score /
+  pen-assemble / pen-compare), an architecture diagram, and the problem/gaps explanation.
+
+## [3.0.0a5] - 2026-06-02 - Phase 1.5 (Bridge-recombinase off-target engine -> Paper 4, BEACHHEAD)
+
+The first public instrument: a bridge-recombinase off-target screening tool.
+
+### Added
+- **Off-target engine** (`pen_stack/bridge/offtarget.py` + `configs/bridge_offtarget_profile.yaml`):
+  genome-wide hg38 pseudosite scan (CT-core seed, per-chromosome, memory-bounded) + a position-weight
+  risk model grounded in the published mechanism. **Beats naive Hamming: AUROC 1.00 vs 0.59** at
+  separating core-preserving (real-risk) from core-disrupting (abolished) sites. Exposes
+  `predict_offtargets(family, site)` - completes the Phase-3 Planner cargo hook.
+- **Fold / cross-loop QC** (`bridge/fold_qc.py`): ViennaRNA fold (verified MFE on a 190-nt design) +
+  TBL/DBL cross-loop complementarity.
+- **Activity framework** (`bridge/activity.py`): exploratory DMS + 72-system trainer (deferred; data paywalled).
+- **`pen-bridge`** (`bridge/pipeline.py`, `bridge/cli.py`, `/bridge/design` API): **wraps** the Arc
+  BridgeRNADesigner (verified) and adds the off-target + QC layer.
+- `validate/paper4_validation.py` + `scripts/p4_genome_scan.py`; `prereg/paper4.yaml` + SHA lock.
+
+### Notes
+- **Phase 1.5 COMPLETE** - pre-registered criteria met (or honestly gated): the off-target engine,
+  ViennaRNA fold, and designer wrap are verified on the VM (real hg38 scan: chr22 in ~21 s). The *blind
+  recall of Perry 2025's measured off-targets* and the DMS/activity model are gated on the paywalled
+  Perry 2025 supplementary (drop in via `ingest.load_offtarget_profile`). Completes the deferred Phase-2
+  Section 2.4 and Phase-3 Section 3.2 hooks. 68 tests green; ruff clean. **All program phases (0,1,1.5,2,3) now done.**
+
+## [3.0.0a4] - 2026-06-02 - Phase 3 (The Write Planner + agentic platform -> Paper 3, CAPSTONE)
+
+Inverse design + the paper-defining recovery@k benchmark + the agentic platform.
+
+### Added
+- **Inverse-design optimiser** (`pen_stack/planner/optimize.py`, `configs/intent_weights.yaml`): an
+  `edit_intent`-conditioned objective whose `target_gene_sign` flips whether hitting the target gene is
+  penalised or rewarded - the same TRAC site ranks #1 (knock-in) vs #101 (safe-harbour).
+- **Cargo/delivery** (`planner/cargo.py`, `planner/delivery.py`): donor spec + size check + delivery rule
+  table; bridge/seek off-target via an optional Phase-1.5 hook (pending until 1.5).
+- **End-to-end Planner** (`planner/pipeline.py`, `report.py`, `/plan` API, `pen-stack plan` CLI): ranked,
+  fully traceable plans with per-field provenance.
+- **Two-stratum recovery@k benchmark** (`validate/paper3_benchmark.py`, `data/benchmark_panel.csv`,
+  `prereg/paper3.yaml`): **discriminating stratum planner 1.00 vs baseline 0.00, McNemar p=0.0156, gap CI
+  [1.0,1.0] excludes zero; control tie 0.67=0.67**. Panel cited to Europe-PMC-verified sources.
+- **Forward hypotheses** (`validate/forward_hypotheses.py`): date-stamped novel F8/SERPINA1/CISH/HBA1
+  proposals + grounded cited ranking.
+- **Agentic platform**: `agent/tools.py` + `agent/orchestrator.py` (Ollama tool-calling, auditable trace,
+  no-fabrication, refusals), `agent/mcp_server.py` (fastmcp), `docker-compose.yml` + `docker/ui.Dockerfile`
+  + Streamlit **Agent** page + `docs/DEPLOY.md`/`docs/MCP.md`, `validate/agent_eval.py`.
+- Shipped `data/curated/gene_coords.parquet` (GENCODE-derived) so tools work in any container.
+
+### Notes
+- **Phase 3 COMPLETE** - pre-registered criteria met (`prereg/paper3.yaml` + `SHA256_LOCK_phase3.json`).
+  Agent verified on the VM in LLM mode (no-fabrication + plan-equivalence + refusals all pass). 63 tests
+  green; ruff clean. Wet-lab (3.7) skipped - non-gating. Bridge off-target hook completes with Phase 1.5.
+
+## [3.0.0a3] - 2026-06-02 - Phase 2 (Writer Atlas + Unified Stack -> Paper 2)
+
+The broad, cross-family Writer Atlas, the writer<->locus cross-link, and the installable platform.
+
+### Added
+- **Writer Atlas** (`pen_stack/atlas/expand.py`, `atlas.parquet`): **33,370 systems across 8 families**
+  (31,885 IS110/IS1111 orthologs + curated cores/reps), every row confidence-tagged + >=1 source DOI,
+  targeting metadata inherited from the WT-KB. `configs/atlas_families.yaml` drives the UniProt queries.
+- **Mechanism at scale** (`pen_stack/mech/`): ported audited 18-family Pfam whitelist v1.2.1; composite
+  co-occurrence rules; **core agreement 1.00** vs audited labels; conflicting calls -> review queue.
+- **Therapeutic readiness** (`pen_stack/score/therapeutic.py`): deliverability/cargo/human-cell axes,
+  components retained (ISCro4 326aa->AAV).
+- **Cross-link** (`pen_stack/atlas/crosslink.py`): bidirectional writer<->locus queries; AAVS1 held-out
+  check passes (0.90 writability + bridge-reachable). Per-family caches for k562/hepg2/hspc.
+- **Variant proposal** (`pen_stack/atlas/variant_propose.py`): point-mutation framework + retrospective
+  harness, no chimeras; DMS model pluggable (Phase 1.5).
+- **PEN-MONITOR** (`pen_stack/monitor/`): Europe PMC living-database engine; back-test surfaces ISPpu10;
+  never auto-edits the atlas; every candidate cited.
+- **Grounded RAG** (`pen_stack/rag/`, `pen_stack/agent/guardrails.py`): numbers from tool calls, claims
+  cited, clinical directives refused; optional Ollama/Qwen phrasing layer (presentation only).
+- **Stack**: unified CLI subcommands, FastAPI server (`pen_stack/server/api.py`), Streamlit platform UI
+  (Writer Atlas + Ask pages), mkdocs site + 4 use-case tutorials. 46 tests green; ruff clean.
+
+### Notes
+- **Phase 2 COMPLETE** - pre-registered criteria met (`prereg/paper2.yaml` + `SHA256_LOCK_phase2.json`);
+  atlas Zenodo DOI pending author upload. Verified on the VM (Docker): API, UI (:8501), RAG with Qwen.
+
+## [3.0.0a0] - 2026-06-01 - Phase 0 (in progress)
+
+Fresh v3.0 monorepo. Supersedes the v1.0 platform repository (archived); consolidates the five prior
+repositories (`genome-atlas`, `mech-class`, `pen-score`, `pen-assemble`, `pen-compare`) as provenance.
+
+### Added
+- Monorepo scaffold: 13 modules (`atlas`, `mech`, `score`, `wgenome`, `planner`, `bridge`, `monitor`,
+  `rag`, `agent`, `ui`, `data`, `validate`, `server`), `pyproject.toml`, Docker image spec, `penctl`
+  laptop<->VM orchestrator, CI, `configs/`, `prereg/`.
+- `docs/INFRA.md` - three-tier (laptop / VM / Drive) Docker-only, SFTP-only workflow.
+- `configs/llm.yaml` - single LLM switch (Ollama + Qwen2.5-7B-Instruct, Apache-2.0).
+- `configs/datasets.yaml` - pinned dataset accessions + verified IDs (see VERIFICATION_REPORT_v3.0).
+
+- **WT-KB** (`pen_stack/atlas/`): 8 fully-sourced writer families with reachability tiers; schema enforces the >=1-DOI sourcing rule.
+- **Re-grounded axes** (`pen_stack/score/recalibrate.py`, `configs/score_axes.yaml`): `S_Cargo` from measured bp, `S_Prog` from targeting modality, `length_aa` backfilled - no per-enzyme overrides.
+- **Canonical universe** (`pen_stack/atlas/universe.py::assemble`): one path joining the 1,058-entity universe + WT-KB + crosswalk; cross-module consistency test.
+- **Descriptive scorecard** (`pen_stack/atlas/scorecard.py`): reframed from the circular certification; blind concordance recovers ISCro4 as the bridge standout without naming it. 21 tests green.
+
+### Notes
+- Independent verification of all datasets/IDs/DOIs/tools completed: no critical errors in the v3.0 plan
+  (full report in `Final_Part_v3.0/VERIFICATION_REPORT_v3.0.md`).
+- **Phase 0 COMPLETE** - all pre-registered success criteria met (`prereg/phase0.yaml` + SHA lock).

{pen_stack-3.1.0 → pen_stack-3.2.0}/CITATION.cff

@@ -1,19 +1,19 @@
-cff-version: 1.2.0
-message: "If you use PEN-STACK, please cite it as below."
-title: "PEN-STACK: open infrastructure for genome writing"
-version: 3.1.0
-date-released: 2026-06-01
-authors:
-  - family-names: "Mahaboob Ali"
-    given-names: "Anees Ahmed"
-    affiliation: "VIT University, Vellore"
-    email: ahmedaneesm@gmail.com
-repository-code: "https://github.com/ahmedanees-m/pen-stack"
-license: MIT
-keywords:
-  - genome writing
-  - writable genome
-  - writer atlas
-  - bridge recombinase
-  - safe harbor
-  - write planner
+cff-version: 1.2.0
+message: "If you use PEN-STACK, please cite it as below."
+title: "PEN-STACK: open infrastructure for genome writing"
+version: 3.2.0
+date-released: 2026-06-01
+authors:
+  - family-names: "Mahaboob Ali"
+    given-names: "Anees Ahmed"
+    affiliation: "VIT University, Vellore"
+    email: ahmedaneesm@gmail.com
+repository-code: "https://github.com/ahmedanees-m/pen-stack"
+license: MIT
+keywords:
+  - genome writing
+  - writable genome
+  - writer atlas
+  - bridge recombinase
+  - safe harbor
+  - write planner