pen-stack 6.9.1__tar.gz → 6.10.0__tar.gz

{pen_stack-6.9.1 → pen_stack-6.10.0}/CHANGELOG.md

@@ -3,6 +3,79 @@
 All notable changes to PEN-STACK are documented here. This file follows
 [Keep a Changelog](https://keepachangelog.com/) and the program's phase structure.
 
+## [6.10.0] - 2026-06-20 - WS-OFFTARGET (PEN-OFFTGT: cross-writer-family off-target nomination)
+
+**Series III, Stage E.** Off-target moves from a single-family bridge pseudosite scan to a **cross-writer-family,
+chromatin-aware NOMINATION engine** grounded in unbiased genome-wide assays — completing the safety triad
+(site B + writer C + off-target E). Nomination is scrupulously framed as **not a clearance**: every candidate
+ships with the empirical assay that would confirm it.
+
+### Added
+- **Off-Target-Bench** (`benchmarks/offtarget/`) — a real, leakage-controlled nomination benchmark over canonical
+  Cas9 guides (EMX1/VEGFA1-3/FANCF/HEK293) with **experimentally validated off-targets** from GUIDE-seq
+  (Tsai 2015, `10.1038/nbt.3117`) and CIRCLE-seq (Tsai 2017, `10.1038/nmeth.4278`). Held-out-guide split, per-assay
+  provenance, SHA256SUMS. **Gate E-G2 PASSES on real data + real tool:** the licensed **CRISOT-Score** predictor
+  (Chen et al., Nat Commun 2023, `10.1038/s41467-023-42695-4`; XGBoost RNA-DNA fingerprint) BEATS the sequence-
+  homology baseline — GUIDE-seq AUPRC **0.646 vs 0.467** (gap +0.179, CI [0.015, 0.340]); CIRCLE-seq **0.520 vs
+  0.266** (gap +0.253, CI [0.140, 0.361]); per-guide bootstrap CI excludes 0 on both assays.
+- `pen_stack/wgenome/offtarget_data.py` — validated assay/predictor provenance, a GROUNDED mismatch→active-fraction
+  risk calibration (real-data: GUIDE-seq 0-1mm→100% active, 2mm→76%, 3mm→23%, 4mm→3.3%), the bench fixture loader.
+- `pen_stack/wgenome/offtarget_predict.py` — `nominate_offtargets(writer_family, ...)`: **nuclease** (mismatch-
+  calibrated risk band + the real cached CRISOT score + a documented chromatin modifier, Lazzarotto 2020);
+  **serine integrase** (cryptic **pseudo-attB** scan on the real documented Bxb1 attB core GCGGTCTC/GT);
+  **bridge** (delegates to the existing Perry-DMS pseudosite engine). Abstains without inputs; never fabricates sites.
+- `pen_stack/wgenome/offtarget_assay.py` — validation-assay recommender (GUIDE/CHANGE/CIRCLE-seq for nucleases;
+  Cryptic-seq/HIDE-seq for integrases; **honest gap** for bridge recombinases — NO published genome-wide unbiased
+  off-target assay or predictor exists, verified).
+- **Surfaces:** REST `POST /offtarget` + `GET /offtarget/assay`, MCP `offtarget_scan`, manifest `nominate_offtargets`
+  (fabricates=False), an `offtarget_nomination` scope card, and a web **Off-Target** page.
+
+### Honest limits
+- Nomination is NOT a clearance; genome-wide candidate ENUMERATION needs the on-VM Cas-OFFinder/genome scan (this
+  engine SCORES + RANKS + risk-bands supplied candidates). The CRISOT predictor is CC-BY-NC — it runs only on the
+  VM and its weights are NEVER redistributed; only derived scores are cached (CI-safe). Bridge/integrase off-target
+  is data-thin/unmodeled and is flagged extrapolative; IntQuery (Tome Biosciences) is a paper-only reference.
+
+## [6.9.2] - 2026-06-19 - WS-IMMUNE2 real-tool rigor pass (no proxies / no heuristics across the immune + writer axes)
+
+**PATCH — a top-to-bottom audit replacing every remaining proxy/heuristic in the immune & writer-design stack with
+the real on-VM tool (gold-standard, licensed) or an honest abstention. No silent fallbacks.** Triggered by the
+directive *"use the real on-VM tool. No proxy or heuristics … check previous versions too … re-confirm the MHC-I
+capsid axis against NetMHCpan-4.1."* All licensed binaries stay on the VM — only derived numbers are cached.
+
+### Changed
+- **MHC-I capsid axis re-grounded on NetMHCpan-4.1** (`pen_stack/planner/capsid_epitope_oracle.py`). The PRIMARY
+  capsid CD8 predictor is now the gold-standard licensed **NetMHCpan-4.1** (%Rank_EL≤0.5, residue coverage, 12-allele
+  panel; `configs/mhc_epitope_oracle.yaml` `mhc1`); the v5.3 **MHCflurry** value is kept as an explicit, reported
+  **cross-check** (never silently substituted). Both agree AAV is the least CD8-immunogenic capsid
+  (AAV `capsid_immune_score` 0.4585 NetMHCpan / 0.2803 MHCflurry — the predictor disagreement is surfaced, not hidden).
+- **ADA-risk re-grounded** (`pen_stack/planner/ada_risk.py`). `ada_risk = real NetMHCIIpan-4.0 density × foreignness`,
+  where **foreignness is the protein ORIGIN** (authoritative central-tolerance signal). Unknown origin / uncached
+  density now **ABSTAINS** (no k-mer guess, no heuristic) — replacing the v6.9.0 albumin-only self-tolerance
+  heuristic. The **real human-proteome 9-mer self-match** (computed on the VM over the full UniProt reference
+  proteome, 20 431 proteins / 10.4 M 9-mers) is reported as a cross-check: human albumin **1.0** (self), the
+  foreign writers (SpCas9/ISCro4/Bxb1) and capsids **0.0** (non-self) — clean self/non-self separation.
+- **MHC-II axis no longer falls back to a production proxy** (`pen_stack/planner/immune_mhc2.py`). `mhc2_epitope_load`
+  uses the real NetMHCIIpan-4.0 cache by antigen name and otherwise **abstains**; the documented promiscuous-binder
+  estimate is retained as `mhc2_proxy_estimate` for **offline triage only** (explicitly labelled, not the production axis).
+- **Real Bxb1 attB written** (`pen_stack/atlas/guide_design.py`). The PASTE/PASSIGE pegRNA 3′ extension now writes the
+  **real documented Bxb1 minimal attB** verbatim (FlyBase FBto0000359; Ghosh, Kim & Hatfull, Mol Cell 2003;
+  8-bp common core GCGGTCTC around the central GT crossover) instead of the schematic poly-G/poly-C arms. Integrases
+  without a bundled documented site expose only the central core (never a fabricated full sequence).
+- `pen_stack/planner/immune_profile.py` — the writer-as-antigen card surfaces `foreignness`, the real
+  `self_match_human_proteome` cross-check, and the MHC-II/ADA backends (was the removed `self_tolerance` field).
+
+### Honest deviations from the v6.9 pre-registration (`prereg/ws_immune2.yaml`)
+- The pre-registered "1 − self_match_fraction from a human-proteome k-mer filter" foreignness *fallback* is
+  **dropped**: foreignness is the authoritative origin, and an unknown origin abstains rather than imputing from a
+  k-mer match (the real self-match is reported only as a cross-check). This is a strictly more honest rule.
+
+### Data / tests
+- `configs/mhc_epitope_oracle.yaml` — corrected `self_match` (the v6.9.x cache had a FASTA-keying bug that collapsed
+  the 20 431-protein proteome to 1 protein, zeroing every self-match incl. albumin); recomputed correctly on the VM.
+- `tests/unit/{test_ws_immune2,test_ws_epitope,test_ws_writer,test_ws_rel}.py` updated for the abstain semantics,
+  the NetMHCpan-4.1 primary capsid value + MHCflurry cross-check, the real attB, and the 6.9.2 version pins.
+
 ## [6.9.1] - 2026-06-20 - Real NetMHCIIpan-4.0 / NetMHCpan-4.1 MHC epitope load (replaces the v6.9.0 proxy)
 
 **PATCH — rigour upgrade.** v6.9.0's MHC-II axis was a documented *heuristic proxy* (P1-anchor density). The

{pen_stack-6.9.1 → pen_stack-6.10.0}/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: 6.9.1
+version: 6.10.0
 date-released: 2026-06-20
 authors:
   - family-names: "Mahaboob Ali"

{pen_stack-6.9.1 → pen_stack-6.10.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pen-stack
-Version: 6.9.1
+Version: 6.10.0
 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-6.9.1-blue.svg)](CHANGELOG.md)
+[![Version](https://img.shields.io/badge/version-6.10.0-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 v6.10 — PEN-OFFTGT (cross-writer-family off-target nomination)
+
+Off-target prediction was a single-family bridge pseudosite scan that abstained for nucleases and integrases.
+v6.10 makes it a **cross-writer-family, chromatin-aware NOMINATION engine** — completing the safety triad
+(site + writer + off-target). It is scrupulously honest that **nomination is not a clearance**: every candidate
+ships with the empirical assay that would confirm it.
+
+- **A real, validated benchmark** (`benchmarks/offtarget/`) over canonical Cas9 guides with **experimentally
+  validated off-targets** from GUIDE-seq (Tsai 2015) and CIRCLE-seq (Tsai 2017). The licensed **CRISOT** predictor
+  (Chen et al. *Nat Commun* 2023) **beats the sequence-homology baseline** on held-out guides — GUIDE-seq AUPRC
+  **0.65 vs 0.47**, CIRCLE-seq **0.52 vs 0.27**; per-guide bootstrap CI excludes 0 on both. CRISOT is CC-BY-NC and
+  runs only on the VM — only derived scores are cached (like the licensed MHC tools).
+- **Grounded, calibrated risk** — the risk band IS the real-data fraction of candidates at *k* mismatches that were
+  experimentally validated-active (GUIDE-seq: 0–1 mm → 100%, 2 mm → 76%, 3 mm → 23%, 4 mm → 3.3%), not a guessed curve.
+- **Cross-family** — nucleases (mismatch-calibrated risk + the real CRISOT score), serine integrases (cryptic
+  **pseudo-attB** scan on the real documented Bxb1 attB core), bridge recombinases (the existing Perry-DMS engine).
+  Bridge/integrase off-target is **honestly flagged data-thin** — there is no published genome-wide unbiased assay
+  or predictor for bridge recombinases (verified).
+- **Surfaces** — REST `POST /offtarget`, MCP `offtarget_scan`, manifest `nominate_offtargets`, and an Off-Target
+  web page. Abstains without inputs; never fabricates sites.
+
 ## What is new in v6.9 — PEN-IMMUNE (MHC-II/CD4 + ADA + the writer enzyme as a distinct antigen)
 
 The immune profile did **CD8/MHC-I only** (capsid epitope load via MHCflurry) — but the **dominant** immunogenicity
@@ -161,16 +182,24 @@ are bacterial/phage) — yet Stage G scored only the capsid. v6.9 closes that ga
   never distributed), with a documented promiscuous-binder proxy as the offline/CI fallback. Population-level proxy
   (🟡), never a patient-HLA magnitude. *(v6.9.0 used the heuristic proxy; the real tool is more discriminating —
   human-self albumin 0.066 vs foreign writers 0.11–0.15.)*
-- **An ADA-risk axis** (`planner/ada_risk.py`) — **ADA-risk = MHC-II epitope density × foreignness**, with a
-  **self-tolerance filter** (JanusMatrix-style: self epitopes are tolerated; foreign drive ADA). It **recovers
+- **An ADA-risk axis** (`planner/ada_risk.py`) — **ADA-risk = MHC-II epitope density × foreignness**, where
+  **foreignness is the protein ORIGIN** (the authoritative central-tolerance signal). It **recovers
   immunogenic-vs-tolerated**: the foreign writers (real UniProt SpCas9 / ISCro4 / Bxb1) score **above** the human
-  self control (albumin), even without the origin label (the k-mer self-match tolerates the human protein).
+  self control (albumin). The **real full-human-proteome 9-mer self-match** (computed on the VM, 20 431 proteins) is
+  reported as a cross-check (albumin 1.0 self, foreign writers 0.0).
 - **The writer as a distinct antigen** — the profile now carries a `writer_as_antigen` card and a
   `writer_dominant_risk` flag: for **non-viral delivery of a bacterial writer there is no capsid antigen, so the
   WRITER is the dominant immunogen** — the insight the capsid-only profile missed.
 - **Immuno-Bench** (`benchmarks/immuno/`) — the immunogenic-vs-tolerated recovery track + an honest `calibrate_axis`
   ADA pass (it stays 🟡 at public-data power — no manufactured ✅, the standing wet-lab bottleneck).
 
+**v6.9.2 — real-tool rigor pass (no proxies / no heuristics).** A top-to-bottom audit: the **MHC-I capsid axis is
+re-grounded on the gold-standard NetMHCpan-4.1** (MHCflurry kept as a reported cross-check); **ADA foreignness is the
+authoritative origin** and **abstains** when unknown (no k-mer guess); the MHC-II axis **abstains** rather than
+falling back to a production proxy (the promiscuous-binder estimate is offline-triage only); and the PASTE pegRNA
+writes the **real documented Bxb1 minimal attB** (FlyBase / Ghosh 2003) instead of a schematic. Real tool or honest
+abstention — never a silent fallback.
+
 Real UniProt sequences only (no fabricated sequence); the axes are reported as a vector with `collapsed_score: None`;
 the realized CD4 response / ADA titer stay known-unknowns. See [docs/immune_profiler.md](docs/immune_profiler.md).
 

{pen_stack-6.9.1 → pen_stack-6.10.0}/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-6.9.1-blue.svg)](CHANGELOG.md)
+[![Version](https://img.shields.io/badge/version-6.10.0-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 v6.10 — PEN-OFFTGT (cross-writer-family off-target nomination)
+
+Off-target prediction was a single-family bridge pseudosite scan that abstained for nucleases and integrases.
+v6.10 makes it a **cross-writer-family, chromatin-aware NOMINATION engine** — completing the safety triad
+(site + writer + off-target). It is scrupulously honest that **nomination is not a clearance**: every candidate
+ships with the empirical assay that would confirm it.
+
+- **A real, validated benchmark** (`benchmarks/offtarget/`) over canonical Cas9 guides with **experimentally
+  validated off-targets** from GUIDE-seq (Tsai 2015) and CIRCLE-seq (Tsai 2017). The licensed **CRISOT** predictor
+  (Chen et al. *Nat Commun* 2023) **beats the sequence-homology baseline** on held-out guides — GUIDE-seq AUPRC
+  **0.65 vs 0.47**, CIRCLE-seq **0.52 vs 0.27**; per-guide bootstrap CI excludes 0 on both. CRISOT is CC-BY-NC and
+  runs only on the VM — only derived scores are cached (like the licensed MHC tools).
+- **Grounded, calibrated risk** — the risk band IS the real-data fraction of candidates at *k* mismatches that were
+  experimentally validated-active (GUIDE-seq: 0–1 mm → 100%, 2 mm → 76%, 3 mm → 23%, 4 mm → 3.3%), not a guessed curve.
+- **Cross-family** — nucleases (mismatch-calibrated risk + the real CRISOT score), serine integrases (cryptic
+  **pseudo-attB** scan on the real documented Bxb1 attB core), bridge recombinases (the existing Perry-DMS engine).
+  Bridge/integrase off-target is **honestly flagged data-thin** — there is no published genome-wide unbiased assay
+  or predictor for bridge recombinases (verified).
+- **Surfaces** — REST `POST /offtarget`, MCP `offtarget_scan`, manifest `nominate_offtargets`, and an Off-Target
+  web page. Abstains without inputs; never fabricates sites.
+
 ## What is new in v6.9 — PEN-IMMUNE (MHC-II/CD4 + ADA + the writer enzyme as a distinct antigen)
 
 The immune profile did **CD8/MHC-I only** (capsid epitope load via MHCflurry) — but the **dominant** immunogenicity
@@ -86,16 +107,24 @@ are bacterial/phage) — yet Stage G scored only the capsid. v6.9 closes that ga
   never distributed), with a documented promiscuous-binder proxy as the offline/CI fallback. Population-level proxy
   (🟡), never a patient-HLA magnitude. *(v6.9.0 used the heuristic proxy; the real tool is more discriminating —
   human-self albumin 0.066 vs foreign writers 0.11–0.15.)*
-- **An ADA-risk axis** (`planner/ada_risk.py`) — **ADA-risk = MHC-II epitope density × foreignness**, with a
-  **self-tolerance filter** (JanusMatrix-style: self epitopes are tolerated; foreign drive ADA). It **recovers
+- **An ADA-risk axis** (`planner/ada_risk.py`) — **ADA-risk = MHC-II epitope density × foreignness**, where
+  **foreignness is the protein ORIGIN** (the authoritative central-tolerance signal). It **recovers
   immunogenic-vs-tolerated**: the foreign writers (real UniProt SpCas9 / ISCro4 / Bxb1) score **above** the human
-  self control (albumin), even without the origin label (the k-mer self-match tolerates the human protein).
+  self control (albumin). The **real full-human-proteome 9-mer self-match** (computed on the VM, 20 431 proteins) is
+  reported as a cross-check (albumin 1.0 self, foreign writers 0.0).
 - **The writer as a distinct antigen** — the profile now carries a `writer_as_antigen` card and a
   `writer_dominant_risk` flag: for **non-viral delivery of a bacterial writer there is no capsid antigen, so the
   WRITER is the dominant immunogen** — the insight the capsid-only profile missed.
 - **Immuno-Bench** (`benchmarks/immuno/`) — the immunogenic-vs-tolerated recovery track + an honest `calibrate_axis`
   ADA pass (it stays 🟡 at public-data power — no manufactured ✅, the standing wet-lab bottleneck).
 
+**v6.9.2 — real-tool rigor pass (no proxies / no heuristics).** A top-to-bottom audit: the **MHC-I capsid axis is
+re-grounded on the gold-standard NetMHCpan-4.1** (MHCflurry kept as a reported cross-check); **ADA foreignness is the
+authoritative origin** and **abstains** when unknown (no k-mer guess); the MHC-II axis **abstains** rather than
+falling back to a production proxy (the promiscuous-binder estimate is offline-triage only); and the PASTE pegRNA
+writes the **real documented Bxb1 minimal attB** (FlyBase / Ghosh 2003) instead of a schematic. Real tool or honest
+abstention — never a silent fallback.
+
 Real UniProt sequences only (no fabricated sequence); the axes are reported as a vector with `collapsed_score: None`;
 the realized CD4 response / ADA titer stay known-unknowns. See [docs/immune_profiler.md](docs/immune_profiler.md).
 

pen_stack-6.10.0/benchmarks/offtarget/SHA256SUMS

@@ -0,0 +1,4 @@
+9a02d9fc0fb4fe305fc87510114df892027ac85f24a495bddccaf1c29e8decf7  split.json
+006642247dbb3dfe5ebefd0a9f0144e9e9d26b682e7b9173d13e2c4e1252bec8  offtarget_bench_fixture.csv
+6ec099c0d30f83bd09f7182c4f365ee7b1106e45f4ca80db46f4fdb0df2941ba  offtarget_bench_metrics.json
+626d383ab3a494d6899c1620fa7cb2e6626c055410c1fd57e6b7d57d308523f4  offtarget_calibration.json

pen_stack-6.10.0/configs/mhc_epitope_oracle.yaml

@@ -0,0 +1,296 @@
+version: 6.9.2
+method:
+  mhc2: NetMHCIIpan-4.0 EL %Rank<=2, residue coverage
+  hla2_panel:
+  - DRB1_0101
+  - DRB1_0301
+  - DRB1_0401
+  - DRB1_0701
+  - DRB1_1101
+  - DRB1_1301
+  - DRB1_1501
+  mhc1: NetMHCpan-4.1 %Rank_EL<=0.5, residue coverage
+  hla1_panel:
+  - HLA-A01:01
+  - HLA-A02:01
+  - HLA-A03:01
+  - HLA-A11:01
+  - HLA-A24:02
+  - HLA-A26:01
+  - HLA-B07:02
+  - HLA-B08:01
+  - HLA-B15:01
+  - HLA-B40:01
+  - HLA-B44:03
+  - HLA-B58:01
+  metric: residue coverage by strong binders, union over panel (matches v5.3 MHCflurry)
+provenance_dois:
+- 10.1093/nar/gkac1029
+- 10.1093/nar/gkaa379
+note: population-level epitope load (frequent-HLA panel; NOT patient-HLA-specific).
+  Licensed NetMHC binaries run locally; only derived fractions cached.
+mhc2:
+  SpCas9:
+    epitope_fraction_strong: 0.636
+    immune_score: 0.364
+    metric: residue coverage by strong 15-mers (%Rank<=2.0), union over 7 alleles
+    length: 1368
+    n_covered: 870
+    alleles_used:
+    - DRB1_0101
+    - DRB1_0301
+    - DRB1_0401
+    - DRB1_0701
+    - DRB1_1101
+    - DRB1_1301
+    - DRB1_1501
+  ISCro4:
+    epitope_fraction_strong: 0.5644
+    immune_score: 0.4356
+    metric: residue coverage by strong 15-mers (%Rank<=2.0), union over 7 alleles
+    length: 326
+    n_covered: 184
+    alleles_used:
+    - DRB1_0101
+    - DRB1_0301
+    - DRB1_0401
+    - DRB1_0701
+    - DRB1_1101
+    - DRB1_1301
+    - DRB1_1501
+  Bxb1:
+    epitope_fraction_strong: 0.6467
+    immune_score: 0.3533
+    metric: residue coverage by strong 15-mers (%Rank<=2.0), union over 7 alleles
+    length: 501
+    n_covered: 324
+    alleles_used:
+    - DRB1_0101
+    - DRB1_0301
+    - DRB1_0401
+    - DRB1_0701
+    - DRB1_1101
+    - DRB1_1301
+    - DRB1_1501
+  HumanAlbumin:
+    epitope_fraction_strong: 0.3186
+    immune_score: 0.6814
+    metric: residue coverage by strong 15-mers (%Rank<=2.0), union over 7 alleles
+    length: 609
+    n_covered: 194
+    alleles_used:
+    - DRB1_0101
+    - DRB1_0301
+    - DRB1_0401
+    - DRB1_0701
+    - DRB1_1101
+    - DRB1_1301
+    - DRB1_1501
+  AAV2_VP1:
+    epitope_fraction_strong: 0.5592
+    immune_score: 0.4408
+    metric: residue coverage by strong 15-mers (%Rank<=2.0), union over 7 alleles
+    length: 735
+    n_covered: 411
+    alleles_used:
+    - DRB1_0101
+    - DRB1_0301
+    - DRB1_0401
+    - DRB1_0701
+    - DRB1_1101
+    - DRB1_1301
+    - DRB1_1501
+  Ad5_hexon:
+    epitope_fraction_strong: 0.6229
+    immune_score: 0.3771
+    metric: residue coverage by strong 15-mers (%Rank<=2.0), union over 7 alleles
+    length: 952
+    n_covered: 593
+    alleles_used:
+    - DRB1_0101
+    - DRB1_0301
+    - DRB1_0401
+    - DRB1_0701
+    - DRB1_1101
+    - DRB1_1301
+    - DRB1_1501
+  VSVg_Indiana:
+    epitope_fraction_strong: 0.3914
+    immune_score: 0.6086
+    metric: residue coverage by strong 15-mers (%Rank<=2.0), union over 7 alleles
+    length: 511
+    n_covered: 200
+    alleles_used:
+    - DRB1_0101
+    - DRB1_0301
+    - DRB1_0401
+    - DRB1_0701
+    - DRB1_1101
+    - DRB1_1301
+    - DRB1_1501
+  HSV1_gD:
+    epitope_fraction_strong: 0.5355
+    immune_score: 0.4645
+    metric: residue coverage by strong 15-mers (%Rank<=2.0), union over 7 alleles
+    length: 394
+    n_covered: 211
+    alleles_used:
+    - DRB1_0101
+    - DRB1_0301
+    - DRB1_0401
+    - DRB1_0701
+    - DRB1_1101
+    - DRB1_1301
+    - DRB1_1501
+  HSV1_gB:
+    epitope_fraction_strong: 0.4746
+    immune_score: 0.5254
+    metric: residue coverage by strong 15-mers (%Rank<=2.0), union over 7 alleles
+    length: 904
+    n_covered: 429
+    alleles_used:
+    - DRB1_0101
+    - DRB1_0301
+    - DRB1_0401
+    - DRB1_0701
+    - DRB1_1101
+    - DRB1_1301
+    - DRB1_1501
+mhc1:
+  AAV2_VP1:
+    epitope_fraction_strong: 0.5415
+    immune_score: 0.4585
+    metric: residue coverage by strong 9-mers (%Rank<=0.5), union over 12 alleles
+    length: 735
+    n_covered: 398
+    alleles_used:
+    - HLA-A01:01
+    - HLA-A02:01
+    - HLA-A03:01
+    - HLA-A11:01
+    - HLA-A24:02
+    - HLA-A26:01
+    - HLA-B07:02
+    - HLA-B08:01
+    - HLA-B15:01
+    - HLA-B40:01
+    - HLA-B44:03
+    - HLA-B58:01
+  Ad5_hexon:
+    epitope_fraction_strong: 0.7216
+    immune_score: 0.2784
+    metric: residue coverage by strong 9-mers (%Rank<=0.5), union over 12 alleles
+    length: 952
+    n_covered: 687
+    alleles_used:
+    - HLA-A01:01
+    - HLA-A02:01
+    - HLA-A03:01
+    - HLA-A11:01
+    - HLA-A24:02
+    - HLA-A26:01
+    - HLA-B07:02
+    - HLA-B08:01
+    - HLA-B15:01
+    - HLA-B40:01
+    - HLA-B44:03
+    - HLA-B58:01
+  VSVg_Indiana:
+    epitope_fraction_strong: 0.6223
+    immune_score: 0.3777
+    metric: residue coverage by strong 9-mers (%Rank<=0.5), union over 12 alleles
+    length: 511
+    n_covered: 318
+    alleles_used:
+    - HLA-A01:01
+    - HLA-A02:01
+    - HLA-A03:01
+    - HLA-A11:01
+    - HLA-A24:02
+    - HLA-A26:01
+    - HLA-B07:02
+    - HLA-B08:01
+    - HLA-B15:01
+    - HLA-B40:01
+    - HLA-B44:03
+    - HLA-B58:01
+  HSV1_gD:
+    epitope_fraction_strong: 0.6853
+    immune_score: 0.3147
+    metric: residue coverage by strong 9-mers (%Rank<=0.5), union over 12 alleles
+    length: 394
+    n_covered: 270
+    alleles_used:
+    - HLA-A01:01
+    - HLA-A02:01
+    - HLA-A03:01
+    - HLA-A11:01
+    - HLA-A24:02
+    - HLA-A26:01
+    - HLA-B07:02
+    - HLA-B08:01
+    - HLA-B15:01
+    - HLA-B40:01
+    - HLA-B44:03
+    - HLA-B58:01
+  HSV1_gB:
+    epitope_fraction_strong: 0.5962
+    immune_score: 0.4038
+    metric: residue coverage by strong 9-mers (%Rank<=0.5), union over 12 alleles
+    length: 904
+    n_covered: 539
+    alleles_used:
+    - HLA-A01:01
+    - HLA-A02:01
+    - HLA-A03:01
+    - HLA-A11:01
+    - HLA-A24:02
+    - HLA-A26:01
+    - HLA-B07:02
+    - HLA-B08:01
+    - HLA-B15:01
+    - HLA-B40:01
+    - HLA-B44:03
+    - HLA-B58:01
+self_match:
+  SpCas9:
+    fraction: 0.0
+    n: 1360
+    n_matched: 0
+  ISCro4:
+    fraction: 0.0
+    n: 318
+    n_matched: 0
+  Bxb1:
+    fraction: 0.0
+    n: 493
+    n_matched: 0
+  HumanAlbumin:
+    fraction: 1.0
+    n: 601
+    n_matched: 601
+  AAV2_VP1:
+    fraction: 0.0
+    n: 727
+    n_matched: 0
+  Ad5_hexon:
+    fraction: 0.0
+    n: 944
+    n_matched: 0
+  VSVg_Indiana:
+    fraction: 0.0
+    n: 503
+    n_matched: 0
+  HSV1_gD:
+    fraction: 0.0
+    n: 386
+    n_matched: 0
+  HSV1_gB:
+    fraction: 0.0
+    n: 896
+    n_matched: 0
+self_match_meta:
+  reference: UniProt reviewed human reference proteome UP000005640 (20431 proteins)
+  k: 9
+  tool: exact 9-mer identity vs full human proteome (JanusMatrix-style self-similarity
+    cross-check)

{pen_stack-6.9.1 → pen_stack-6.10.0}/configs/oracles/scope_cards.yaml

@@ -139,6 +139,23 @@ 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)"
 
+  offtarget_nomination:         # v6.10 WS-OFFTARGET: cross-family off-target NOMINATION (not clearance)
+    family: genome
+    version: "crisot+homology-2026"
+    output_kind: baseline       # ranks/risk-bands CANDIDATE sites, not generative
+    valid_for: "RELATIVE off-target NOMINATION for a writer: rank candidate sites so validated off-targets
+      surface first, with a mismatch-CALIBRATED empirical risk band (real GUIDE-seq/CIRCLE-seq active fractions).
+      Nucleases use the real CRISOT-Score (beats a homology baseline on held-out guides, CI excludes 0); serine
+      integrases use a documented pseudo-attB cryptic scan; bridge recombinases delegate to the Perry-DMS engine"
+    not_valid_for: "a safety CLEARANCE (nomination != validation; every result ships with the empirical assay
+      that would confirm it); genome-wide candidate ENUMERATION (needs the on-VM Cas-OFFinder/genome scan; this
+      engine SCORES supplied candidates); a per-site cleavage/recombination PROBABILITY; bridge-recombinase
+      off-target magnitude (NO published genome-wide unbiased assay/predictor exists -> extrapolative/abstains);
+      structural variants/translocations beyond nominated sites"
+    generalizes_to_unseen_loci: false
+    license: "open (this work; GUIDE-seq 10.1038/nbt.3117, CIRCLE-seq 10.1038/nmeth.4278, CHANGE-seq
+      10.1038/s41587-020-0555-7, CRISOT 10.1038/s41467-023-42695-4 CC-BY-NC run on the VM only)"
+
   innate_sensing:               # v5.4 WS-INNATE: computed nucleic-acid innate-sensing motif load
     family: genome
     version: "cpg-oe+dsrna-2026"

pen_stack-6.10.0/docs/cards/offtarget_data.md

@@ -0,0 +1,52 @@
+# Data card — Off-Target-Bench (v6.10 PEN-OFFTGT)
+
+## Summary
+A real, leakage-controlled benchmark for cross-writer-family off-target **nomination**: given a guide and its
+candidate sites, rank the candidates so the experimentally validated off-targets surface first. Labels are the
+wet-lab assay calls (NON-circular — the label is the experiment, not a predictor).
+
+## Ground truth (independently verified 2026-06-19)
+| Assay | Setting | Citation | DOI |
+|---|---|---|---|
+| GUIDE-seq | cell-based, unbiased, genome-wide | Tsai et al., *Nat Biotechnol* 2015 | `10.1038/nbt.3117` |
+| CIRCLE-seq | in vitro (cell-free), unbiased | Tsai et al., *Nat Methods* 2017 | `10.1038/nmeth.4278` |
+
+Canonical Cas9 guides: **EMX1, VEGFA site 1/2/3, FANCF, HEK293 site 2/3/4**. The harmonized candidate/label tables
+are sourced from the CRISOT data release (Zenodo `10.5281/zenodo.8420032`), which redistributes the public assay
+supplements; PEN-STACK cites the **original assay papers** as the ground-truth provenance.
+
+## Learned predictor (real tool, VM-only)
+**CRISOT-Score** — Chen et al., *Nat Commun* 2023, `10.1038/s41467-023-42695-4`; an XGBoost RNA-DNA interaction
+fingerprint. **License: CC-BY-NC** → it runs only on the VM (`crisot:tools` Docker, `xgboost`/`pandas`/`numpy`);
+its weights are NEVER redistributed. Only DERIVED scores are cached/committed (CI-safe), exactly like the licensed
+NetMHC tools.
+
+## Baseline (pre-registered)
+Sequence-homology nomination = ascending **mismatch count** (Hamming over the 20-nt protospacer).
+
+## Result (E-G2, full real data, on the VM)
+| Assay | CRISOT AUPRC | homology AUPRC | gap | 95% CI (held-out-guide bootstrap) | beats homology |
+|---|---|---|---|---|---|
+| GUIDE-seq | 0.646 | 0.467 | +0.179 | [0.015, 0.340] | ✅ |
+| CIRCLE-seq | 0.520 | 0.266 | +0.253 | [0.140, 0.361] | ✅ |
+
+The learned predictor beats the homology baseline on both assays (per-guide bootstrap CI excludes 0).
+
+## Risk calibration (grounded)
+The nomination risk band IS the empirical fraction of candidates at *k* mismatches that were validated-active
+(full real data): GUIDE-seq 0–1 mm → 1.00, 2 mm → 0.765, 3 mm → 0.231, 4 mm → 0.033, 5 mm → 0.0028, 6 mm → 0.00014.
+Mismatch counts outside the calibrated range abstain rather than extrapolate.
+
+## Files
+- `benchmarks/offtarget/offtarget_bench_fixture.csv` — real validated off-targets + cached CRISOT scores (CI-safe;
+  inactives downsampled with a fixed seed for a small committed file).
+- `benchmarks/offtarget/offtarget_bench_metrics.json` — the AUTHORITATIVE full-data metrics.
+- `benchmarks/offtarget/offtarget_calibration.json` — the full-data mismatch / CRISOT-decile calibration.
+- `benchmarks/offtarget/split.json`, `SHA256SUMS` — split definition + checksums.
+
+## Honest limits
+Nomination is **not** a clearance — every result ships with the empirical assay that would confirm it. Genome-wide
+candidate ENUMERATION needs the on-VM Cas-OFFinder/genome scan; this benchmark covers SCORING + RANKING of supplied
+candidates. Bridge/integrase off-target is data-thin: there is **no published genome-wide unbiased off-target assay
+or predictor for bridge recombinases** (verified), and the large-serine-integrase assays (Cryptic-seq/HIDE-seq) and
+predictor (IntQuery) are recent single-company preprints with no public weights.