pen-stack 6.10.3__tar.gz → 6.11.0__tar.gz

{pen_stack-6.10.3 → pen_stack-6.11.0}/CHANGELOG.md

@@ -3,6 +3,62 @@
 All notable changes to PEN-STACK are documented here. This file follows
 [Keep a Changelog](https://keepachangelog.com/) and the program's phase structure.
 
+## [6.11.0] - 2026-06-20 - WS-DELIVERY (PEN-DELIVER: cross-modality deliverability + learned capsid-fitness)
+
+**Series III, Stage D.** Delivery moves from an 8-vehicle rule palette to a cross-modality recommender with a
+**learned, benchmarked AAV capsid-fitness model**, a **grounded serotype→tissue tropism prior** from approved
+therapies, and an **immune-coupled dose tradeoff** — without fabricating tropism. All datasets independently verified.
+
+### Added
+- **Learned AAV capsid-fitness on FLIP-AAV** (`pen_stack/planner/delivery_predict.py`, `scripts/build_capsid_fitness.py`,
+  `benchmarks/delivery/`). Trained on the FLIP-AAV benchmark (Dallago 2021; Bryant 2021 packaging fitness,
+  `10.1038/s41587-020-00793-4`; Ogden 2019 `10.1126/science.aaw2900`). **Gate D-G1 PASSES** — beats a mutation-burden
+  baseline on **both** held-out splits: `sampled` Spearman **0.920 vs 0.522** (CI [0.387, 0.411]); `mut_des`
+  (mutant→designed) **0.814 vs 0.752** (CI [0.061, 0.064]). The 217 MB/split FLIP data stays on the VM; the derived
+  metrics + reproducible build script are committed. The ~3 MB model itself is gitignored — regenerated via
+  `scripts/build_capsid_fitness.py` and mounted into the deployed app (like `position_effect.pkl`); the axis abstains
+  gracefully when absent. Predicted fitness is a CANDIDATE for the measured packaging axis, NOT in-vivo tropism.
+- **Serotype→tissue tropism priors** (`configs/aav_serotype_tropism.yaml`) from APPROVED AAV therapies (AAV9→CNS,
+  **AAVrh74→skeletal muscle** vs **AAVRh74var→liver** — different capsids, kept separate; AAV5→liver; AAV2→retina/
+  putamen via local injection). A grounded prior for an approved serotype; a known-unknown (abstain) for a novel
+  capsid. Casgevy/Lyfgenia carry no serotype prior (not-AAV controls).
+- **Immune-coupled dose tradeoff** (`pen_stack/planner/delivery_immune.py`) — fuses deliverability with the Stage G
+  immune profile, surfaces dose↔immunogenicity per vehicle as a vector (`collapsed_score=None`, never fused).
+- **Verify-gated generative capsids** (`pen_stack/design/capsid_generate.py`) — fitness-filtered VP1 variants, 🔵
+  candidates only; abstains without the model.
+- **Surfaces:** REST `POST /delivery` + `POST /capsid_fitness` + `GET /delivery/tropism`, MCP `delivery_recommend`,
+  manifest `recommend_delivery` + `capsid_fitness`, a `capsid_fitness` scope card, a `delivery` Challenge task
+  (serotype→tissue), prereg `ws_delivery.yaml`.
+
+### Honest limits / deviations (see docs/DEVIATIONS_AND_DISCLOSURES.md)
+- The capsid-fitness model is a one-hot gradient-boosting model, not a protein-LM (disclosed; passes the gate anyway;
+  protein-LM is the upgrade path). Capsid-fitness covers AAV only (VLP/LNP data is thinner). In-vivo human tropism is
+  a known-unknown except the approved-therapy priors; novel capsids abstain on tropism. alpha-retro-VLP is exploratory.
+
+## [6.10.4] - 2026-06-20 - Chromatin incremental-value test (annotation, not a re-ranker) — chromatin work complete
+
+**PATCH — the final chromatin question, answered.** v6.10.3 validated accessibility as a moderate standalone
+predictor; v6.10.4 tests whether it adds **incremental** value *on top of* the CRISOT sequence score — i.e. whether
+it should re-rank.
+
+### Added
+- **Incremental-value analysis** (`scripts/offtarget_chromatin_incremental.py`; result in
+  `benchmarks/offtarget/chromatin_incremental.json`). On GUIDE-seq (HEK293T-matched), per off-target CRISOT +
+  accessibility: (A) a conditional logistic regression `active ~ z(CRISOT) + z(accessibility)` and (B) a
+  leave-one-guide-out held-out AUPRC of CRISOT-only vs a CRISOT+accessibility combiner — tested at two candidate
+  imbalances (1:16 and a realistic 1:123). **Result:** accessibility carries a **small, real conditional signal**
+  (coef ≈ 0.35, bootstrap CI excludes 0 at both imbalances) but adds **NO held-out ranking improvement** over CRISOT
+  (AUPRC gap CI includes 0 at both: −0.0025 [−0.011, +0.005] and +0.0027 [−0.014, +0.021]).
+
+### Decision
+- **Chromatin is a validated ANNOTATION, NOT a re-ranker.** The CRISOT sequence score already captures the
+  practically-relevant nomination ranking; a CRISOT+accessibility combiner does not improve held-out AUPRC, so it is
+  **not** wired into the numeric risk score (`CHROMATIN_VALIDATION.changes_numeric_risk_score = False`). The fitted
+  combiner coefficients are recorded for transparency but intentionally not applied. Ledger + docs updated.
+- **This completes the chromatin context work** (v6.10.1 wired → v6.10.2 cross-cell weak → v6.10.3 cell-type-matched
+  VALIDATED moderate → v6.10.4 incremental tested, annotation-only). Nothing about chromatin is now open or
+  undisclosed.
+
 ## [6.10.3] - 2026-06-20 - Chromatin context: cell-type-matched validation (VALIDATED, moderate)
 
 **PATCH — the definitive chromatin test.** v6.10.2's controlled experiment was ambiguous (GUIDE-seq positive, TTISS

{pen_stack-6.10.3 → pen_stack-6.11.0}/CITATION.cff

@@ -1,7 +1,7 @@
 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: 6.10.3
+version: 6.11.0
 date-released: 2026-06-20
 authors:
   - family-names: "Mahaboob Ali"

{pen_stack-6.10.3 → pen_stack-6.11.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pen-stack
-Version: 6.10.3
+Version: 6.11.0
 Summary: Open infrastructure for genome writing: the Writable Genome atlas, the Writer Atlas, and the Write Planner.
 Author-email: Anees Ahmed Mahaboob Ali <ahmedaneesm@gmail.com>
 License: MIT
@@ -90,7 +90,7 @@ every design against rule-grounded mechanism, reports calibrated confidence, cit
 [![codecov](https://codecov.io/gh/ahmedanees-m/pen-stack/branch/main/graph/badge.svg)](https://codecov.io/gh/ahmedanees-m/pen-stack)
 [![License: MIT](https://img.shields.io/badge/License-MIT-informational.svg)](LICENSE)
 [![Python 3.11+](https://img.shields.io/badge/python-3.11%2B-blue.svg)](https://www.python.org/)
-[![Version](https://img.shields.io/badge/version-6.10.3-blue.svg)](CHANGELOG.md)
+[![Version](https://img.shields.io/badge/version-6.11.0-blue.svg)](CHANGELOG.md)
 [![Status](https://img.shields.io/badge/status-1.0%20First%20Stable-success.svg)](docs/STABILITY.md)
 [![Tests](https://img.shields.io/badge/tests-378%20passing-success.svg)](tests/)
 [![Lint: ruff](https://img.shields.io/badge/lint-ruff-purple.svg)](https://github.com/astral-sh/ruff)
@@ -148,6 +148,27 @@ demonstrated (v5.12) and the benchmark went public (v5.13).
 > unknown funnel remains — made legible (scope flags, known-unknowns, honest baselines, no fabrication), not
 > hidden.
 
+## What is new in v6.11 — PEN-DELIVER (cross-modality deliverability + learned capsid-fitness)
+
+Delivery was an 8-vehicle rule palette that couldn't predict capsid fitness or tropism. v6.11 makes it a
+cross-modality recommender with a **learned, benchmarked AAV capsid-fitness model** and a **grounded serotype→tissue
+tropism prior** — without ever fabricating tropism.
+
+- **Learned capsid-fitness on FLIP-AAV** (Dallago 2021; Bryant 2021). The model **beats a mutation-burden baseline on
+  both held-out splits** — `sampled` Spearman **0.92 vs 0.52**, and the hard mutant→designed `mut_des` split
+  **0.81 vs 0.75** (per-split bootstrap CI excludes 0). The 217 MB FLIP data stays on the VM; the derived metrics +
+  reproducible build script are committed and the ~3 MB model is gitignored (regenerated via
+  `scripts/build_capsid_fitness.py` and mounted into the deployed app, like `position_effect.pkl`; the axis abstains
+  when absent). Predicted fitness is a **candidate** for the *measured packaging axis* — not an in-vivo tropism claim.
+- **Grounded serotype→tissue priors** from approved AAV therapies (AAV9→CNS, AAVrh74→skeletal muscle, AAV5→liver,
+  AAV2→retina/putamen via local injection) — and an honest **known-unknown (abstain)** for a novel capsid. The
+  AAVrh74 (muscle) vs AAVRh74var (liver) distinction is kept; ex-vivo therapies carry no serotype prior.
+- **Immune-coupled dose tradeoff** — fuses deliverability with the Stage G immune profile and surfaces the
+  dose↔immunogenicity tension per vehicle as a vector, **never collapsed**.
+- **Verify-gated generative capsids** (🔵 candidates), and surfaces: REST, MCP `delivery_recommend`, a `delivery`
+  Challenge task. *(Plan deviations — e.g. one-hot model vs a protein-LM — are disclosed in
+  [docs/DEVIATIONS_AND_DISCLOSURES.md](docs/DEVIATIONS_AND_DISCLOSURES.md).)*
+
 ## What is new in v6.10 — PEN-OFFTGT (cross-writer-family off-target nomination)
 
 Off-target prediction was a single-family bridge pseudosite scan that abstained for nucleases and integrases.

{pen_stack-6.10.3 → pen_stack-6.11.0}/README.md

@@ -15,7 +15,7 @@ every design against rule-grounded mechanism, reports calibrated confidence, cit
 [![codecov](https://codecov.io/gh/ahmedanees-m/pen-stack/branch/main/graph/badge.svg)](https://codecov.io/gh/ahmedanees-m/pen-stack)
 [![License: MIT](https://img.shields.io/badge/License-MIT-informational.svg)](LICENSE)
 [![Python 3.11+](https://img.shields.io/badge/python-3.11%2B-blue.svg)](https://www.python.org/)
-[![Version](https://img.shields.io/badge/version-6.10.3-blue.svg)](CHANGELOG.md)
+[![Version](https://img.shields.io/badge/version-6.11.0-blue.svg)](CHANGELOG.md)
 [![Status](https://img.shields.io/badge/status-1.0%20First%20Stable-success.svg)](docs/STABILITY.md)
 [![Tests](https://img.shields.io/badge/tests-378%20passing-success.svg)](tests/)
 [![Lint: ruff](https://img.shields.io/badge/lint-ruff-purple.svg)](https://github.com/astral-sh/ruff)
@@ -73,6 +73,27 @@ demonstrated (v5.12) and the benchmark went public (v5.13).
 > unknown funnel remains — made legible (scope flags, known-unknowns, honest baselines, no fabrication), not
 > hidden.
 
+## What is new in v6.11 — PEN-DELIVER (cross-modality deliverability + learned capsid-fitness)
+
+Delivery was an 8-vehicle rule palette that couldn't predict capsid fitness or tropism. v6.11 makes it a
+cross-modality recommender with a **learned, benchmarked AAV capsid-fitness model** and a **grounded serotype→tissue
+tropism prior** — without ever fabricating tropism.
+
+- **Learned capsid-fitness on FLIP-AAV** (Dallago 2021; Bryant 2021). The model **beats a mutation-burden baseline on
+  both held-out splits** — `sampled` Spearman **0.92 vs 0.52**, and the hard mutant→designed `mut_des` split
+  **0.81 vs 0.75** (per-split bootstrap CI excludes 0). The 217 MB FLIP data stays on the VM; the derived metrics +
+  reproducible build script are committed and the ~3 MB model is gitignored (regenerated via
+  `scripts/build_capsid_fitness.py` and mounted into the deployed app, like `position_effect.pkl`; the axis abstains
+  when absent). Predicted fitness is a **candidate** for the *measured packaging axis* — not an in-vivo tropism claim.
+- **Grounded serotype→tissue priors** from approved AAV therapies (AAV9→CNS, AAVrh74→skeletal muscle, AAV5→liver,
+  AAV2→retina/putamen via local injection) — and an honest **known-unknown (abstain)** for a novel capsid. The
+  AAVrh74 (muscle) vs AAVRh74var (liver) distinction is kept; ex-vivo therapies carry no serotype prior.
+- **Immune-coupled dose tradeoff** — fuses deliverability with the Stage G immune profile and surfaces the
+  dose↔immunogenicity tension per vehicle as a vector, **never collapsed**.
+- **Verify-gated generative capsids** (🔵 candidates), and surfaces: REST, MCP `delivery_recommend`, a `delivery`
+  Challenge task. *(Plan deviations — e.g. one-hot model vs a protein-LM — are disclosed in
+  [docs/DEVIATIONS_AND_DISCLOSURES.md](docs/DEVIATIONS_AND_DISCLOSURES.md).)*
+
 ## What is new in v6.10 — PEN-OFFTGT (cross-writer-family off-target nomination)
 
 Off-target prediction was a single-family bridge pseudosite scan that abstained for nucleases and integrases.

pen_stack-6.11.0/benchmarks/delivery/SHA256SUMS

@@ -0,0 +1,2 @@
+0384951e5158a1382f76a2264a486c5fb401729d5f6d87a8ac44776839e0fc50  split.json
+91041bd15e899d194e43070849d501a1b0efb4f9dff568587b03505ca0e96d3c  capsid_fitness_metrics.json

pen_stack-6.11.0/configs/aav_serotype_tropism.yaml

@@ -0,0 +1,81 @@
+# PEN-STACK v6.11 — AAV serotype -> tissue tropism PRIORS, grounded in APPROVED gene therapies (D-WS1).
+# Each prior is a DOCUMENTED, real serotype->tissue mapping evidenced by an FDA/EMA-approved product (the
+# authoritative source for "this capsid delivers to this tissue in humans"). Independently verified 2026-06-20.
+#
+# HONEST SCOPE: these are population-level PRIORS for the listed serotypes, strongest for the approved route.
+# - AAV2 tropism is ROUTE-DEPENDENT (retina via subretinal; CNS-putamen via intraputaminal) — local injection, not
+#   intrinsic systemic tropism. `route` is a covariate, not ignorable.
+# - AAVrh74 (Elevidys, skeletal muscle) and AAVRh74var (Beqvez, liver) are DIFFERENT capsids with DIFFERENT tropism —
+#   never collapsed into one "Rh74" prior.
+# - Ex-vivo cell therapies (Casgevy = non-viral CRISPR RNP; Lyfgenia = lentiviral) are NOT AAV and carry NO
+#   serotype->tissue prior (listed only as not-AAV controls).
+# - For a NOVEL / engineered capsid with no approved precedent, in-vivo human tropism is a known-unknown (⛔) — the
+#   recommender abstains on the tropism prior and never fabricates a tissue.
+
+version: "1.0"
+
+serotypes:
+  AAV9:
+    tissue: ["CNS", "motor_neuron"]
+    route: "systemic IV (blood-brain-barrier-crossing)"
+    example_product: "Zolgensma (onasemnogene abeparvovec)"
+    indication: "spinal muscular atrophy"
+    approval: "FDA 2019"
+    source: "FDA/Novartis PI; DailyMed setid 68cd4f06-70e1-40d8-bedb-609ec0afa471"
+    doi: "10.1056/NEJMoa1706198"          # Mendell et al. NEJM 2017 (AVXS-101 phase 1)
+    confidence: grounded
+  AAVrh74:
+    tissue: ["skeletal_muscle", "cardiac_muscle"]
+    route: "systemic IV"
+    note: "capsid serotype rh74; AAV2-derived ITRs"
+    example_product: "Elevidys (delandistrogene moxeparvovec)"
+    indication: "Duchenne muscular dystrophy"
+    approval: "FDA 2023"
+    source: "FDA PI; DailyMed setid 3525abcd-fbd2-46c5-9706-cffdf2e8a361"
+    confidence: grounded
+  AAV5:
+    tissue: ["liver"]
+    route: "systemic IV"
+    example_product: "Hemgenix (etranacogene dezaparvovec); Roctavian (valoctocogene roxaparvovec)"
+    indication: "hemophilia B; hemophilia A"
+    approval: "FDA 2022 / 2023"
+    source: "FDA Hemgenix + Roctavian product pages"
+    doi: "10.1056/NEJMoa2211644"          # Pipe et al. NEJM 2023 (HOPE-B, etranacogene)
+    confidence: grounded
+  AAVRh74var:
+    tissue: ["liver"]
+    route: "systemic IV"
+    note: "an ENGINEERED Rh74 VARIANT capsid — distinct from wild-type AAVrh74 (which is muscle-tropic)"
+    example_product: "Beqvez (fidanacogene elaparvovec)"
+    indication: "hemophilia B"
+    approval: "FDA 2024"
+    source: "FDA Beqvez PI"
+    confidence: grounded
+  AAV2:
+    tissue: ["retina", "CNS_putamen"]
+    route: "LOCAL injection only (subretinal for retina; intraputaminal for CNS) — route-dependent, NOT systemic"
+    example_product: "Luxturna (voretigene neparvovec, retina); Upstaza/Kebilidi (eladocagene exuparvovec, putamen)"
+    indication: "RPE65 inherited retinal dystrophy; AADC deficiency"
+    approval: "FDA 2017 (Luxturna); EMA 2022 / FDA 2024 (Upstaza/Kebilidi)"
+    source: "FDA Luxturna; EMA Upstaza EPAR"
+    doi: "10.1016/S0140-6736(17)31868-8"  # Russell et al. Lancet 2017 (voretigene phase 3)
+    confidence: grounded
+
+# NOT AAV — no serotype->tissue prior (ex-vivo cell therapies; tissue targeting comes from HSC selection +
+# autologous transplant + conditioning, NOT capsid tropism). Listed only as not-AAV controls.
+not_aav_controls:
+  Casgevy:
+    modality: "ex-vivo CRISPR/Cas9 RNP, non-viral electroporation of CD34+ HSCs"
+    serotype_prior: none
+    approval: "FDA 2023"
+  Lyfgenia:
+    modality: "ex-vivo lentiviral (BB305 LVV) modified CD34+ HSCs"
+    serotype_prior: none
+    approval: "FDA 2023"
+
+modeling_notes:
+  - "route is a covariate: AAV2's retina/putamen tropism is via LOCAL injection, not systemic IV"
+  - "AAVrh74 (muscle) vs AAVRh74var (liver) are separate capsids — never merge"
+  - "AAV5 -> liver is the strongest systemic-IV prior (two independent approved products)"
+  - "novel/engineered capsid with no approved precedent -> in-vivo human tropism is a known-unknown (abstain)"
+provenance_dois: ["10.1056/NEJMoa1706198", "10.1056/NEJMoa2211644", "10.1016/S0140-6736(17)31868-8"]

{pen_stack-6.10.3 → pen_stack-6.11.0}/configs/oracles/scope_cards.yaml

@@ -139,6 +139,22 @@ oracles:
     generalizes_to_unseen_loci: false
     license: "open (this work; MHCflurry 10.1016/j.cels.2020.06.010, HLA-I supertypes 10.1186/1471-2172-9-1)"
 
+  capsid_fitness:               # v6.11 WS-DELIVERY: learned AAV capsid packaging-fitness + serotype tropism priors
+    family: protein_design
+    version: "flip-aav-gbr-2026"
+    output_kind: baseline       # scores/ranks capsid variants + tissue priors, not generative
+    valid_for: "RELATIVE AAV capsid PACKAGING-FITNESS prediction for VP1 variants (learned on FLIP-AAV / Bryant
+      2021; beats a mutation-burden baseline on held-out splits incl. mutant->designed); and GROUNDED serotype->
+      tissue tropism PRIORS from approved AAV gene therapies (AAV9->CNS, AAVrh74->skeletal muscle, AAV5->liver,
+      AAV2->retina/putamen via local injection)"
+    not_valid_for: "IN-VIVO HUMAN TROPISM of a capsid (the measured axis is packaging viability/assembly, not in-vivo
+      tissue targeting — a known-unknown except the approved-therapy priors); a NOVEL/engineered capsid's tissue
+      (abstains, never fabricated); manufacturing/CMC; immunogenicity magnitude (Stage G; a known-unknown);
+      VLP/LNP capsid-fitness (no equivalent learned model — thinner data)"
+    generalizes_to_unseen_loci: false
+    license: "open (this work; FLIP-AAV 10.1101/2021.11.09.467890, Bryant 2021 10.1038/s41587-020-00793-4,
+      Ogden 2019 10.1126/science.aaw2900; serotype priors from approved-therapy FDA approvals)"
+
   offtarget_nomination:         # v6.10 WS-OFFTARGET: cross-family off-target NOMINATION (not clearance)
     family: genome
     version: "crisot+homology-2026"

{pen_stack-6.10.3 → pen_stack-6.11.0}/docs/DEVIATIONS_AND_DISCLOSURES.md

@@ -16,13 +16,24 @@ applicable (e.g. single contributor).
 |---|---|---|
 | Nuclease predictor named in plan (CCLMoff / CRISMER) vs shipped (CRISOT) | **DISCLOSED** | We use **CRISOT-Score** (Chen 2023, CC-BY-NC, CPU), not the plan's CCLMoff/CRISMER. Justified: **CRISMER ships no license** (cannot be redistributed/wrapped); **CCLMoff** is a GPU RNA-language-model stack AND was trained on these very assays (evaluating it on them would be **leakage**), whereas **CRISOT-Score is MD-physics, assay-agnostic → a leakage-clean held-out evaluation**. CRISOT is the more rigorous and license-appropriate choice. A genuine substitution, disclosed. |
 | Assay coverage (plan listed GUIDE/CIRCLE/CHANGE/SITE/SURRO/TTISS/Digenome-seq) | **FIXED (v6.10.1)** | v6.10.0 used GUIDE-seq + CIRCLE-seq only. v6.10.1 adds **CHANGE-seq + SITE-seq** (independent broad guide panels) — CRISOT beats homology on **all four** (per-guide bootstrap CI excludes 0). SURRO-seq is targeted/biased (kept as an orthogonal reference, not genome-wide truth); TTISS/Digenome-seq remain **DISCLOSED** as not-yet-folded-in. |
-| "Chromatin-aware via the Stage B feature store" | **VALIDATED (moderate, cell-type-matched; v6.10.3)** | v6.10.0 overclaimed "chromatin-aware" with only a caller hook → v6.10.1 added a real `locus_accessibility()` (verified on the VM) → v6.10.2 ran a controlled test with the cross-cell K562 proxy and got a *weak/inconsistent* result (GUIDE-seq 0.58, TTISS 0.346, in-vitro null) → **v6.10.3 settled it with a CELL-TYPE-MATCHED track** (ENCODE HEK293T DNase, ENCFF529BOG; matches GUIDE-seq/TTISS). **Cell-type matching lifts the canonical WT-Cas9 cell-based assay GUIDE-seq from AUROC 0.58 → 0.671 (CI [0.642, 0.701])**, the in-vitro control stays null (0.494). **VERDICT: VALIDATED (moderate, cell-type-matched) for WT-Cas9 cell-based off-target activity** — the cross-cell K562 proxy was dampening a real effect. Honest caveats: the effect is **moderate** (the sequence/CRISOT score still dominates); it does **not** transfer to TTISS (0.383, an expected outlier — a Cas9-*variant* specificity assay); and chromatin is surfaced as a validated **annotation** that does **not yet change the numeric risk score** (a calibrated CRISOT+accessibility combination is the remaining refinement, deferred). Full result: `benchmarks/offtarget/chromatin_validation.json`; reproducible scripts `scripts/offtarget_chromatin_{validation,matched}.py`. |
+| "Chromatin-aware via the Stage B feature store" | **VALIDATED (moderate, cell-type-matched; v6.10.3)** | v6.10.0 overclaimed "chromatin-aware" with only a caller hook → v6.10.1 added a real `locus_accessibility()` (verified on the VM) → v6.10.2 ran a controlled test with the cross-cell K562 proxy and got a *weak/inconsistent* result (GUIDE-seq 0.58, TTISS 0.346, in-vitro null) → **v6.10.3 settled it with a CELL-TYPE-MATCHED track** (ENCODE HEK293T DNase, ENCFF529BOG; matches GUIDE-seq/TTISS). **Cell-type matching lifts the canonical WT-Cas9 cell-based assay GUIDE-seq from AUROC 0.58 → 0.671 (CI [0.642, 0.701])**, the in-vitro control stays null (0.494). **VERDICT: VALIDATED (moderate, cell-type-matched) for WT-Cas9 cell-based off-target activity** — the cross-cell K562 proxy was dampening a real effect. Honest caveats: the effect is **moderate** (the sequence/CRISOT score dominates); it does **not** transfer to TTISS (0.383, an expected outlier — a Cas9-*variant* specificity assay). **v6.10.4 closed the last piece** — does accessibility add *incremental* value over CRISOT? On GUIDE-seq (HEK293T-matched), accessibility carries a **small real conditional signal** (logreg coef ~0.35, CI excludes 0 at two imbalances) but adds **NO held-out ranking improvement** over CRISOT (AUPRC gap CI includes 0 at both). **Decision: chromatin is a validated ANNOTATION, NOT a re-ranker** — it does **not** change the numeric risk score (CRISOT captures the ranking); the fitted combiner is recorded but intentionally not applied. Full results: `benchmarks/offtarget/chromatin_{validation,incremental}.json`; reproducible scripts `scripts/offtarget_chromatin_{validation,matched,incremental}.py`. **This is the final, complete state — nothing about chromatin is now open or undisclosed.** |
 | Held-out **guide** AND **locus** splits | **DISCLOSED** | Held-out-**guide** is implemented; a genomic-coordinate **locus** split is **not possible** — the harmonized assay data ships sequences, not coordinates. Instead we add **cross-assay generalization** (the assay-agnostic CRISOT-Score evaluated on GUIDE/CIRCLE canonical guides + CHANGE/SITE independent panels). This is the leakage-clean substitute and is stated in `benchmarks/offtarget/split.json`. |
 | Learned **integrase** off-target scorer (plan: "learned on HIDE/Cryptic-seq") | **DISCLOSED** | Shipped as a documented **pseudo-attB cryptic scan** on the real Bxb1 attB core, not a learned model — Cryptic-seq/HIDE-seq are recent preprints and IntQuery has **no public weights**, so a real learned integrase predictor is not groundable. Flagged extrapolative; IntQuery cited as paper-only. |
 | Bench "joins the Challenge" | **FIXED (v6.10.1)** | An `offtarget` nomination task is added to the Genome-Writing Challenge (`benchmarks/genome_writing_challenge/harness.py`) — non-circular (label = wet-lab Active call), data-gated on the fixture. |
 | "≥1 external use" (success criterion) | **DISCLOSED** | Forward-looking; not yet met (no external submission). The surfaces (REST/MCP/manifest/web) exist to enable it. |
 | HIDE-seq / Cryptic-seq integrase **data** | **DISCLOSED** | Validated + cited, **not used as data** (single-company Bxb1-centric preprint; bridge off-target is unmodeled). Used only to ground the assay recommender. |
 
+## v6.11 — PEN-DELIVER (Stage D, delivery)
+
+| Item | Status | Detail |
+|---|---|---|
+| Capsid-fitness model: plan named a "protein-LM" | **DISCLOSED** | Shipped a **windowed one-hot gradient-boosting** model (CPU, reproducible, ~3 MB), not a protein language model. It still **passes D-G1** (beats the mutation-burden baseline on both FLIP-AAV splits, CI excludes 0). A protein-LM (Wu et al. 2025, `10.1089/hum.2024.227`, reported r≈0.82) is the upgrade path — deferred. |
+| Capsid-fitness covers **AAV only** | **DISCLOSED** | No learned fitness model for VLP/LNP capsids — the public data is far thinner than AAV (per the dataset validation). VLP/LNP are ranked by the documented rule palette only; the recommender does not over-rank low-N modalities. |
+| FLIP-AAV: used `sampled` + `mut_des`; multi-trait (Eid 2024) not used | **DISCLOSED** | The bench uses the two most informative FLIP-AAV splits (in-distribution + the hard mutant→designed generalization). The other FLIP splits and the multi-trait Eid 2024 / Fit4Function data (tropism/manufacturability) are a documented extension, not used. |
+| alpha-retro-VLP | **DISCLOSED** | Not added as a grounded palette entry — its strongest in-vivo evidence is a **2025 conference abstract** (grey literature), the peer-reviewed anchor (Baron 2022) is ex-vivo. Flagged exploratory, not fabricated. |
+| In-vivo human tropism | **HONEST KNOWN-UNKNOWN** | Predicted capsid-fitness is for the **measured packaging axis**; in-vivo human tissue tropism is a known-unknown except the approved-therapy serotype priors. The recommender abstains on tropism for novel capsids — never fabricates a tissue. |
+| Model artifact (`models/capsid_fitness.pkl`, ~3 MB) **not committed to git** | **DISCLOSED** | The trained `.pkl` is **gitignored** (the `position_effect.pkl` pattern) — the **derived Delivery-Bench metrics + the reproducible build script** (`scripts/build_capsid_fitness.py`) are committed instead, and the model is **regenerated from FLIP-AAV on the VM and mounted into the deployed app**. CI stays lean and license-clean; the capsid-fitness axis **abstains gracefully** when the model file is absent (no fabricated scores). |
+
 ## v6.9 — PEN-IMMUNE (Stage G) — *audited retroactively, corrected in v6.10.1*
 
 | Item | Status | Detail |

{pen_stack-6.10.3 → pen_stack-6.11.0}/docs/cards/offtarget_data.md

@@ -61,10 +61,16 @@ AUROC of accessibility for active-vs-inactive off-targets per assay, with in-vit
 DNase track (`ENCFF529BOG`, matching GUIDE-seq's HEK293 and TTISS's HEK293T): **cell-type matching lifts the
 canonical WT-Cas9 cell-based assay from 0.58 → 0.671** (the cross-cell proxy was dampening a real effect), with the
 in-vitro control still null. **Verdict: VALIDATED (moderate, cell-type-matched)** for WT-Cas9 cell-based off-target
-activity. Honest caveats: the effect is **moderate** (the sequence/CRISOT score still dominates nomination); it does
-**not** transfer to TTISS (a Cas9-variant specificity assay, the expected outlier); and chromatin is surfaced as a
-validated **annotation** that does **not yet change the numeric risk score** (a calibrated CRISOT+accessibility
-combination is the remaining refinement). Reproducible: `scripts/offtarget_chromatin_{validation,matched}.py`.
+activity. Honest caveats: the effect is **moderate** (the sequence/CRISOT score dominates nomination); it does
+**not** transfer to TTISS (a Cas9-variant specificity assay, the expected outlier).
+
+**Incremental value over CRISOT (v6.10.4, `chromatin_incremental.json`):** on GUIDE-seq (HEK293T-matched),
+accessibility carries a **small real conditional signal** (logistic-regression coefficient ~0.35, bootstrap CI
+excludes 0 at both 1:16 and 1:123 candidate imbalance) but adds **NO held-out ranking improvement** over CRISOT
+(leave-one-guide-out AUPRC gap CI includes 0 at both). **Decision: chromatin is a validated ANNOTATION, NOT a
+re-ranker** — it does **not** change the numeric risk score (CRISOT already captures the practically-relevant ranking
+signal); the fitted CRISOT+accessibility combiner is recorded but intentionally not applied. Reproducible:
+`scripts/offtarget_chromatin_{validation,matched,incremental}.py`.
 
 ## Risk calibration (grounded)
 The nomination risk band IS the empirical fraction of candidates at *k* mismatches that were validated-active

pen_stack-6.11.0/docs/delivery_recommender.md

@@ -0,0 +1,54 @@
+# Stage D — the deliverability engine (v6.11 PEN-DELIVER)
+
+Stage D upgrades delivery from an 8-vehicle **rule palette** to a cross-modality recommender with a **learned,
+benchmarked AAV capsid-fitness model**, a **grounded serotype→tissue tropism prior** from approved therapies, and an
+**immune-coupled dose tradeoff** — without ever fabricating tropism.
+
+## Learned capsid-fitness (`planner/delivery_predict.capsid_fitness`)
+
+A model trained on the **FLIP-AAV** benchmark (Dallago et al. 2021; built on **Bryant et al. 2021** packaging
+fitness, `10.1038/s41587-020-00793-4`; **Ogden et al. 2019** is the foundational landscape, `10.1126/science.aaw2900`).
+A windowed one-hot gradient-boosting model over the mutagenized VP1 561–588 region. **Gate D-G1 PASSES** — it beats a
+mutation-burden baseline on held-out FLIP-AAV on **both** splits (`benchmarks/delivery/`):
+
+| Split | learned Spearman | baseline | gap 95% CI |
+|---|---|---|---|
+| `sampled` (random 80/20, in-distribution) | **0.920** | 0.522 | [0.387, 0.411] |
+| `mut_des` (mutant→designed, hard generalization) | **0.814** | 0.752 | [0.061, 0.064] |
+
+The licensed FLIP-AAV data (217 MB/split) stays on the VM; the **derived metrics + reproducible build script**
+(`scripts/build_capsid_fitness.py`) are committed. The trained model (`models/capsid_fitness.pkl`, ~3 MB) is
+**gitignored** — regenerated from the build script and **mounted into the deployed app** (the `position_effect.pkl`
+pattern), so CI stays lean and the axis **abstains gracefully** when the model is absent. **Honest scope:** predicted
+fitness is a **candidate** for the *measured* packaging axis — it is **not** an in-vivo human-tropism claim (a
+known-unknown).
+
+## Serotype → tissue tropism priors (`configs/aav_serotype_tropism.yaml`)
+
+Real serotype↔tissue mappings evidenced by **approved AAV gene therapies** (independently verified): AAV9→CNS
+(Zolgensma), **AAVrh74→skeletal muscle** (Elevidys), AAV5→liver (Hemgenix/Roctavian), **AAVRh74var→liver** (Beqvez —
+a *different* capsid from AAVrh74), AAV2→retina/putamen via **local injection** (Luxturna/Upstaza). A grounded prior
+for an approved serotype; a **known-unknown (abstain)** for a novel/engineered capsid — never a fabricated tissue.
+Ex-vivo cell therapies (Casgevy = non-viral CRISPR; Lyfgenia = lentiviral) carry **no** serotype prior.
+
+## Immune-coupled tradeoff (`planner/delivery_immune.delivery_immune_tradeoff`)
+
+Fuses deliverability with the **Stage G** per-axis immune profile and surfaces the **dose↔immunogenicity tradeoff**
+per vehicle as a vector — `collapsed_score` is always `None`. The realized in-vivo magnitude / patient titer are
+declared known-unknowns; this is decision-support, not a clinical dosing directive.
+
+## Generative candidates (`design/capsid_generate.generate_capsid_candidates`)
+
+Verifier-as-discriminator (v5.8): propose VP1-region variants, score with the learned capsid-fitness, keep survivors
+above the WT fitness. Every survivor is a **🔵 candidate** — packaging fitness only; assembly, in-vivo tropism, and
+immunogenicity are not claimed. Abstains when the model is absent (no fabricated capsids).
+
+## Surfaces
+REST `POST /delivery` + `POST /capsid_fitness` + `GET /delivery/tropism`, MCP `delivery_recommend`, manifest
+`recommend_delivery` + `capsid_fitness`, the `capsid_fitness` scope card, and a `delivery` task in the
+Genome-Writing Challenge (serotype→tissue, label = approved-therapy registry).
+
+## Honest limits
+Predicted capsid-fitness is strongest for the **measured** packaging axis and extrapolative for in-vivo human tissue
+(⛔). VLP/LNP coverage is thinner than AAV (alpha-retro-VLP is exploratory — its strongest in-vivo evidence is a
+conference abstract). Manufacturing/CMC is out of scope. See `docs/DEVIATIONS_AND_DISCLOSURES.md` for plan deviations.

{pen_stack-6.10.3 → pen_stack-6.11.0}/pen_stack/__init__.py

@@ -1,2 +1,2 @@
 """PEN-STACK v3.0 - open infrastructure for genome writing."""
-__version__ = "6.10.3"
+__version__ = "6.11.0"

{pen_stack-6.10.3 → pen_stack-6.11.0}/pen_stack/agent/cite.py

@@ -52,6 +52,16 @@ def curated_dois() -> frozenset[str]:
                     dois.add(_rec["doi"])
     except Exception:  # noqa: BLE001
         pass
+    # v6.11 delivery provenance (FLIP-AAV capsid-fitness + approved-therapy serotype->tissue priors)
+    try:
+        trop = yaml.safe_load(resource("configs/aav_serotype_tropism.yaml").read_text(encoding="utf-8")) or {}
+        dois.update(trop.get("provenance_dois", []) or [])
+        for _rec in (trop.get("serotypes") or {}).values():
+            if _rec.get("doi"):
+                dois.add(_rec["doi"])
+        dois.update(["10.1101/2021.11.09.467890", "10.1038/s41587-020-00793-4", "10.1126/science.aaw2900"])
+    except Exception:  # noqa: BLE001
+        pass
     # v5.5 anti-vector seroprevalence provenance (serosurveys)
     try:
         sp = yaml.safe_load(resource("configs/seroprevalence.yaml").read_text(encoding="utf-8"))

{pen_stack-6.10.3 → pen_stack-6.11.0}/pen_stack/agent/mcp_server.py

@@ -103,6 +103,16 @@ def offtarget_scan(writer_family: str, guide: str | None = None, candidate_sites
                                sequence=sequence, assay=assay)
 
 
+@mcp.tool()
+def delivery_recommend(cargo_form: str, cargo_bp: int | None = None, target_tissue: str | None = None,
+                       safety_weight: float = 0.5, in_vivo: bool | None = None) -> dict:
+    """v6.11 PEN-DELIVER cross-modality delivery recommender: rank vehicles by cargo-form + safety<->efficacy + a
+    GROUNDED serotype->tissue tropism prior (from approved AAV therapies; a known-unknown for novel capsids), plus
+    the learned FLIP-AAV capsid-fitness capability. Never fabricates tropism; abstains without inputs."""
+    from pen_stack.planner.delivery_predict import recommend_delivery_plus
+    return recommend_delivery_plus(cargo_form, cargo_bp, target_tissue, safety_weight=safety_weight, in_vivo=in_vivo)
+
+
 @mcp.tool()
 def generate_designs(goal: dict | None = None, candidates: list | None = None, keep: int = 25) -> dict:
     """v5.8 generative designer (verifier-as-discriminator): hazardous/illegal candidates are DISCARDED;

{pen_stack-6.10.3 → pen_stack-6.11.0}/pen_stack/api/manifest.py

@@ -52,6 +52,17 @@ _TOOLS = [
      "input": "writer family + guide/candidate sites (nuclease) or locus sequence (integrase)",
      "output": "ranked off-target candidates + calibrated risk + recommended validation assay",
      "entrypoint": "pen_stack.wgenome.offtarget_predict.nominate_offtargets", "fabricates": False},
+    {"name": "recommend_delivery", "summary": "cross-modality delivery recommender: rank vehicles by cargo-form + "
+     "safety<->efficacy + a grounded serotype->tissue tropism prior (approved therapies) + the learned FLIP-AAV "
+     "capsid-fitness capability; tropism is a known-unknown for novel capsids (v6.11 PEN-DELIVER)",
+     "input": "cargo form/size + target tissue (+ safety weight, in/ex-vivo)",
+     "output": "ranked vehicles + serotype tropism prior + capsid-fitness bench; never fabricates tropism",
+     "entrypoint": "pen_stack.planner.delivery_predict.recommend_delivery_plus", "fabricates": False},
+    {"name": "capsid_fitness", "summary": "learned AAV capsid packaging-fitness for a VP1 sequence (FLIP-AAV-"
+     "trained; beats a mutation-burden baseline on held-out splits); a CANDIDATE for the measured packaging axis, "
+     "NOT an in-vivo tropism claim (v6.11 PEN-DELIVER)",
+     "input": "AAV VP1 capsid amino-acid sequence", "output": "predicted packaging-fitness (candidate) or abstain",
+     "entrypoint": "pen_stack.planner.delivery_predict.capsid_fitness", "fabricates": False},
 ]
 
 _POLICY = ("outputs outside scope are returned as `out_of_scope` (known-unknown) or `extrapolating` (OOD) and are "

pen_stack-6.11.0/pen_stack/design/capsid_generate.py

@@ -0,0 +1,62 @@
+"""Verify-gated generative AAV capsid candidates (v6.11 PEN-DELIVER, D-WS3).
+
+Proposes engineered AAV capsid variants in the mutagenized VP1 region, scores them with the learned FLIP-AAV
+capsid-fitness model, and keeps only candidates above a fitness threshold — the verifier-as-discriminator pattern
+(v5.8): GENERATE proposes, the fitness model + biosecurity context DISPOSE. Every survivor is a CANDIDATE (🔵),
+never asserted to assemble or to have any in-vivo tropism. Abstains when the fitness model is not present (no
+fabricated capsids).
+
+Honest: a generated capsid's predicted fitness is for the MEASURED packaging axis only; assembly, in-vivo tropism,
+and immunogenicity are not claimed (the latter routes through Stage G). This is a design proposer, not a validation.
+"""
+from __future__ import annotations
+
+_AAS = "ACDEFGHIKLMNPQRSTVWY"
+
+
+def _mutate(seq: str, region: tuple[int, int], n_mut: int, rng) -> str:
+    s = list(seq)
+    lo, hi = region
+    positions = rng.sample(range(lo, min(hi, len(s))), min(n_mut, max(0, min(hi, len(s)) - lo)))
+    for p in positions:
+        s[p] = _AAS[rng.randrange(len(_AAS))]
+    return "".join(s)
+
+
+def generate_capsid_candidates(wt_vp1: str, n: int = 200, max_mut: int = 4, fitness_threshold: float | None = None,
+                               seed: int = 7, top: int = 20) -> dict:
+    """Propose AAV capsid variants (random substitutions in the VP1 555-595 mutagenized window), score with the
+    learned capsid-fitness model, and return the top survivors above ``fitness_threshold`` (defaults to the WT's
+    predicted fitness). Candidates are flagged 🔵; abstains when the fitness model is absent — never fabricates."""
+    import random
+    from pen_stack.planner.delivery_predict import capsid_fitness
+    wt = capsid_fitness(wt_vp1)
+    if not wt.get("available"):
+        return {"available": False, "abstain": True, "candidates": [], "output_kind": "candidate",
+                "note": "capsid-fitness model not present -> cannot score generated capsids; no fabrication.",
+                "bench": wt.get("bench")}
+    rng = random.Random(seed)
+    thr = fitness_threshold if fitness_threshold is not None else wt["predicted_fitness"]
+    region = (555, 595)
+    seen = set()
+    scored = []
+    for _ in range(max(1, n)):
+        nm = rng.randint(1, max_mut)
+        cand = _mutate(wt_vp1, region, nm, rng)
+        if cand == wt_vp1 or cand in seen:
+            continue
+        seen.add(cand)
+        f = capsid_fitness(cand)["predicted_fitness"]
+        scored.append({"vp1_window": cand[region[0]:region[1]], "predicted_fitness": f,
+                       "n_mut_in_region": sum(1 for a, b in zip(cand[region[0]:region[1]],
+                                                                wt_vp1[region[0]:region[1]]) if a != b),
+                       "passes_threshold": bool(f >= thr), "output_kind": "candidate"})
+    survivors = sorted([c for c in scored if c["passes_threshold"]],
+                       key=lambda c: c["predicted_fitness"], reverse=True)[:top]
+    return {"available": True, "abstain": False, "wt_predicted_fitness": wt["predicted_fitness"],
+            "fitness_threshold": round(float(thr), 4), "n_proposed": len(scored), "n_survivors": len(survivors),
+            "candidates": survivors, "output_kind": "candidate",
+            "verifier": "learned FLIP-AAV capsid-fitness (verifier-as-discriminator, v5.8); fitness >= WT kept",
+            "honesty": "🔵 CANDIDATES — predicted packaging fitness only; assembly, in-vivo tropism, and "
+                       "immunogenicity are NOT claimed (immunogenicity routes through Stage G). A design proposer, "
+                       "not a validation. Run a biosecurity/verify() gate before any synthesis."}

pen_stack-6.11.0/pen_stack/planner/delivery_immune.py

@@ -0,0 +1,60 @@
+"""Immune-coupled delivery selection (v6.11 PEN-DELIVER, D-WS4).
+
+Fuses the cross-modality deliverability recommendation (Stage D) with the per-axis immune-risk profile (Stage G) and
+surfaces the **dose <-> immunogenicity tradeoff** as a VECTOR — never collapsed into one number. The point: a vehicle
+that is most deliverable may carry the highest immune liability (e.g. AAV: efficient in vivo, but pre-existing NAbs +
+capsid CD8 + dose-dependent innate sensing), and that tension must be visible, not hidden.
+
+Honest: the immune axes are population-level proxies (Stage G); the realized in-vivo magnitude / patient titer is a
+declared known-unknown. The dose<->immune tradeoff is informative, NOT a clinical dosing directive.
+"""
+from __future__ import annotations
+
+
+def _highest_risk_axis(design: dict) -> dict | None:
+    """The in-scope immune axis with the LOWEST score (= highest risk) from the Stage G profile, with its value."""
+    from pen_stack.planner.immune_profile import immune_profile
+    axes = immune_profile(dict(design)).get("axes", {})
+    in_scope = {k: a for k, a in axes.items() if a.get("in_scope") and a.get("value") is not None}
+    if not in_scope:
+        return None
+    k = min(in_scope, key=lambda x: in_scope[x]["value"])
+    return {"axis": k, "score": in_scope[k]["value"], "note": in_scope[k].get("note")}
+
+
+def delivery_immune_tradeoff(cargo_form: str, cargo_bp: int | None = None, target_tissue: str | None = None,
+                             *, writer_family: str | None = None, serotype: str | None = None,
+                             safety_weight: float = 0.5, in_vivo: bool | None = None) -> dict:
+    """Rank deliverability AND attach each vehicle's Stage G immune profile, surfacing the dose<->immune tradeoff per
+    vehicle as a vector. Never collapses deliverability and immunogenicity into one score."""
+    from pen_stack.planner.delivery_predict import recommend_delivery_plus
+    rec = recommend_delivery_plus(cargo_form, cargo_bp, target_tissue, safety_weight=safety_weight, in_vivo=in_vivo)
+    coupled = []
+    for prof in rec.get("ranked", []) or rec.get("eligible", []) or []:
+        veh = prof.get("vehicle") or prof.get("name")
+        if not veh:
+            continue
+        design = {"delivery_vehicle": veh, "serotype": serotype, "writer_family": writer_family}
+        risk = _highest_risk_axis(design)
+        coupled.append({
+            "vehicle": veh,
+            "deliverability_balance": prof.get("balance"),
+            "efficacy_score": prof.get("efficacy_score"),
+            "immune_highest_risk_axis": risk,                 # the dominant immune liability (Stage G)
+            "tradeoff": (f"deliverability {prof.get('balance')} vs dominant immune liability "
+                         f"'{risk['axis']}' (score {risk['score']})" if risk else
+                         "deliverability ranked; immune axes abstain for this vehicle"),
+        })
+    return {
+        "cargo_form": cargo_form, "cargo_bp": cargo_bp, "target_tissue": target_tissue,
+        "serotype_tropism_prior": rec.get("serotype_tropism_prior"),
+        "coupled": coupled,
+        "collapsed_score": None,           # deliberately None — dose<->immune is a vector, NEVER fused
+        "dose_immune_note": ("AAV/viral vehicles are efficient in vivo but carry the highest immune liability "
+                             "(pre-existing NAbs, capsid CD8, dose-dependent innate sensing); higher dose raises both "
+                             "transduction AND immunogenicity. The tradeoff is surfaced per vehicle, never collapsed."),
+        "known_unknowns": ["in_vivo_immunogenicity_magnitude", "patient_specific_titer", "realized_dose_response"],
+        "honesty": "immune axes are population-level proxies (Stage G); realized magnitude/titer is a known-unknown; "
+                   "this is decision-support, NOT a clinical dosing directive. No collapsed score.",
+        "no_fabrication": True,
+    }