pen-stack 6.10.0__tar.gz → 6.10.2__tar.gz

{pen_stack-6.10.0 → pen_stack-6.10.2}/CHANGELOG.md

@@ -3,6 +3,68 @@
 All notable changes to PEN-STACK are documented here. This file follows
 [Keep a Changelog](https://keepachangelog.com/) and the program's phase structure.
 
+## [6.10.2] - 2026-06-20 - Chromatin context: controlled validation (honest weak/inconsistent result)
+
+**PATCH — validate the chromatin axis for real, then scope it to what the data supports.** v6.10.1 wired a real
+Stage B accessibility read but had not *demonstrated* it carries signal. v6.10.2 runs a controlled experiment and
+reports the result honestly — it is **not** a clean win.
+
+### Added
+- **A controlled chromatin validation** (`benchmarks/offtarget/chromatin_validation.json`,
+  `scripts/offtarget_chromatin_validation.py`). Off-target protospacers mapped to hg38 (GRCh38.fa, exact match both
+  strands, 98.5% mapped); AUROC of the Stage B K562 ATAC/DNase accessibility for active-vs-inactive off-targets per
+  assay, with **in-vitro assays as a NEGATIVE control** (cell-free → no chromatin). Result: the in-vitro controls hug
+  0.5 (method sound — no spurious signal); the canonical cell-based assay **GUIDE-seq is a textbook modest positive
+  (AUROC 0.58, CI [0.550, 0.613])** even with a cross-cell-type K562 proxy; but the second cell-based assay **TTISS
+  reverses (0.346)**. **Verdict: WEAK / INCONSISTENT — chromatin is NOT a validated quantitative axis on this data**
+  (the cross-cell-type proxy is the likely cause; a cell-type-matched accessibility track would settle it — deferred).
+
+### Changed
+- The chromatin-accessibility modifier is now an explicit **annotation only**, labelled `validated=false`: it reads
+  the real Stage B track (or a caller-supplied scalar) and notes the qualitative Lazzarotto-2020 direction, but it
+  **does not change the numeric off-target risk score**. `offtarget_data.CHROMATIN_VALIDATION` carries the result.
+- Corrected the v6.10.1 disclosures-ledger entry that had marked chromatin "FIXED" — that was too strong;
+  it is now disclosed as *tested → weak/inconsistent → documented annotation, not a validated axis*.
+
+This is an honest negative-ish result reported in full rather than dropped or overstated.
+
+## [6.10.1] - 2026-06-20 - WS-OFFTARGET rigor pass + retroactive honesty audit (v6.7–v6.9)
+
+**PATCH — complete the Stage E plan and disclose every remaining deviation.** Closes the v6.10.0 gaps and adds a
+standing [DEVIATIONS_AND_DISCLOSURES.md](docs/DEVIATIONS_AND_DISCLOSURES.md) ledger; also retro-audited v6.7/v6.8/v6.9
+(plan vs shipped vs code) and fixed/disclosed the findings. No item is reported complete if it is a silent
+substitution/partial/deferral.
+
+### Added / changed (Stage E completion)
+- **Off-Target-Bench expanded to FOUR unbiased assays.** Added **CHANGE-seq** (Lazzarotto 2020, `10.1038/s41587-020-0555-7`)
+  + **SITE-seq** (Cameron 2017, `10.1038/nmeth.4284`) on **independent broad guide panels** (20 + 11 guides). The real
+  **CRISOT-Score** (MD-physics, **assay-agnostic → leakage-clean**) beats the homology baseline on **all four**:
+  AUPRC 0.646/0.520/0.541/0.521 vs 0.467/0.266/0.249/0.233; per-guide bootstrap CI excludes 0 on each.
+- **Chromatin wired to the real Stage B track + relabelled honestly.** `locus_accessibility(chrom, bin, ct)` reads
+  `phase_1/features/chromatin_{ct}.parquet` (ATAC/DNase) when present and **abstains** otherwise; the docs now say
+  "chromatin-accessibility modifier" (not "chromatin-aware engine") — the bare wheel / deployed atlas do not ship the
+  raw track, so the modifier is inactive there (disclosed).
+- **Off-target task added to the Genome-Writing Challenge** (`benchmarks/genome_writing_challenge/harness.py`) —
+  non-circular (label = wet-lab Active call), data-gated on the fixture; the reference solver nominates correctly.
+- **Disclosed (not eliminated):** CRISOT is used instead of the plan-named CCLMoff/CRISMER (CRISMER ships no license;
+  CCLMoff is GPU + trained on these assays → leakage; CRISOT is the leakage-clean, license-clean, CPU choice). A
+  genomic-coordinate **locus split** is not possible (harmonized data is coordinate-free) → held-out-guide + cross-assay.
+
+### Fixed (retroactive honesty audit)
+- **v6.9:** dropped the inaccurate "OOD-gated" claim in the MHC-II axis status (it is **coverage-gated**, abstains
+  when uncached — no distributional OOD gate); corrected the stale `AXIS_STATUS["mhc2_writer"].reason` + the
+  `immune_mhc2.py` module docstring (they still described the dropped v6.9.0 P1-anchor proxy as "the method" — the
+  production axis is the real NetMHCIIpan-4.0). Disclosed: the plan-named MHC-II tool ensemble + IEDB/ImmunoSeq/FVIII
+  ADA datasets were not used (single-tool NetMHCIIpan-4.0; the recovery bench is a 4-protein sanity check, not an
+  IEDB held-out leaderboard); the 6.9.1→6.9.2 MHC-II metric change (peptide-fraction → residue-coverage).
+- **v6.8:** disclosed that v6.8.0's attB was a poly-G/C **schematic** (fixed with the real Bxb1 attB in v6.9.2);
+  inter-curator agreement is N/A (single contributor); "≥1 external submission" is unmet (forward-looking).
+- **v6.7:** corrected the "chromosome-Mondrian" served-interval wording (per-chromosome Mondrian qhats are computed,
+  but the **global** qhat is served — correct, since a query has no chromosome at serve time); disclosed the prereg
+  consolidation + the deferred HEK293T-OOD demo / scope-manifest entry.
+
+See the full ledger in [docs/DEVIATIONS_AND_DISCLOSURES.md](docs/DEVIATIONS_AND_DISCLOSURES.md).
+
 ## [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,

{pen_stack-6.10.0 → pen_stack-6.10.2}/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: 6.10.0
+version: 6.10.2
 date-released: 2026-06-20
 authors:
   - family-names: "Mahaboob Ali"

{pen_stack-6.10.0 → pen_stack-6.10.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pen-stack
-Version: 6.10.0
+Version: 6.10.2
 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-6.10.0-blue.svg)](CHANGELOG.md)
+[![Version](https://img.shields.io/badge/version-6.10.2-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)
@@ -151,23 +151,27 @@ demonstrated (v5.12) and the benchmark went public (v5.13).
 ## 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).
+v6.10 makes it a **cross-writer-family NOMINATION engine** with a real-data-calibrated risk band and a documented
+chromatin-accessibility modifier — 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 **experimentally validated off-targets** from
+  **four** unbiased assays — GUIDE-seq (Tsai 2015) + CIRCLE-seq (Tsai 2017) on canonical guides, and CHANGE-seq
+  (Lazzarotto 2020) + SITE-seq (Cameron 2017) on **independent broad guide panels**. The licensed **CRISOT** predictor
+  (Chen et al. *Nat Commun* 2023; MD-physics, assay-agnostic → leakage-clean) **beats the sequence-homology baseline
+  on all four** — AUPRC 0.65/0.52/0.54/0.52 vs 0.47/0.27/0.25/0.23; per-guide bootstrap CI excludes 0 on each.
+  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.
+- **Cross-family** — nucleases (mismatch-calibrated risk + the real CRISOT score + a chromatin-accessibility modifier
+  that reads the Stage B ATAC/DNase track when present, else abstains), 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`, an Off-Target web
+  page, and an off-target task in the Genome-Writing Challenge. Abstains without inputs; never fabricates sites.
+  *(Plan deviations — e.g. the CRISOT-for-CCLMoff/CRISMER substitution — are disclosed in
+  [docs/DEVIATIONS_AND_DISCLOSURES.md](docs/DEVIATIONS_AND_DISCLOSURES.md).)*
 
 ## What is new in v6.9 — PEN-IMMUNE (MHC-II/CD4 + ADA + the writer enzyme as a distinct antigen)
 

{pen_stack-6.10.0 → pen_stack-6.10.2}/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-6.10.0-blue.svg)](CHANGELOG.md)
+[![Version](https://img.shields.io/badge/version-6.10.2-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)
@@ -76,23 +76,27 @@ demonstrated (v5.12) and the benchmark went public (v5.13).
 ## 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).
+v6.10 makes it a **cross-writer-family NOMINATION engine** with a real-data-calibrated risk band and a documented
+chromatin-accessibility modifier — 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 **experimentally validated off-targets** from
+  **four** unbiased assays — GUIDE-seq (Tsai 2015) + CIRCLE-seq (Tsai 2017) on canonical guides, and CHANGE-seq
+  (Lazzarotto 2020) + SITE-seq (Cameron 2017) on **independent broad guide panels**. The licensed **CRISOT** predictor
+  (Chen et al. *Nat Commun* 2023; MD-physics, assay-agnostic → leakage-clean) **beats the sequence-homology baseline
+  on all four** — AUPRC 0.65/0.52/0.54/0.52 vs 0.47/0.27/0.25/0.23; per-guide bootstrap CI excludes 0 on each.
+  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.
+- **Cross-family** — nucleases (mismatch-calibrated risk + the real CRISOT score + a chromatin-accessibility modifier
+  that reads the Stage B ATAC/DNase track when present, else abstains), 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`, an Off-Target web
+  page, and an off-target task in the Genome-Writing Challenge. Abstains without inputs; never fabricates sites.
+  *(Plan deviations — e.g. the CRISOT-for-CCLMoff/CRISMER substitution — are disclosed in
+  [docs/DEVIATIONS_AND_DISCLOSURES.md](docs/DEVIATIONS_AND_DISCLOSURES.md).)*
 
 ## What is new in v6.9 — PEN-IMMUNE (MHC-II/CD4 + ADA + the writer enzyme as a distinct antigen)
 

pen_stack-6.10.2/benchmarks/offtarget/SHA256SUMS

@@ -0,0 +1,4 @@
+8752efa5b6ada36cdf07707999380150f749e819178008be04dcc6b8cd70c084  split.json
+e149d144c5b74499010b062cf9d3a7e2e290e71c8d3720d7b75811bf2ad7dbf3  offtarget_bench_fixture.csv
+e699f399c1c640968aea1d4fe6cc679f035db3a653a5766fed795ab5eec2cecb  offtarget_bench_metrics.json
+42556ae9915037d8a041259daafa93265a5c3eb9f6377245fdb425a4c3418448  offtarget_calibration.json

pen_stack-6.10.2/docs/DEVIATIONS_AND_DISCLOSURES.md

@@ -0,0 +1,57 @@
+# PEN-STACK — Deviations & Disclosures Ledger
+
+A standing, honest record of every place a shipped release **deviated** from its execution plan — a dataset/model
+that was substituted, a deliverable that was narrowed or deferred, or wording that overclaimed — together with the
+correction or the justification. Maintained from v6.10.1 onward (the rule: *no item is reported complete if it is a
+silent substitution/partial/deferral; it is either fixed or disclosed here*). Audited retroactively for v6.7–v6.9.
+
+Legend: **FIXED** = corrected in code/docs · **DISCLOSED** = a justified, permanent deviation · **N/A** = not
+applicable (e.g. single contributor).
+
+---
+
+## v6.10 / v6.10.1 — PEN-OFFTGT (Stage E, off-target nomination)
+
+| Item | Status | Detail |
+|---|---|---|
+| 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" | **DISCLOSED (tested v6.10.2 — weak/inconsistent)** | v6.10.0 overclaimed "chromatin-aware" with only a caller-supplied hook. v6.10.1 added a **real** `locus_accessibility(chrom, bin, ct)` reading the Stage B `chromatin_{ct}.parquet` ATAC/DNase track (verified on the VM), abstaining when absent, and relabelled "chromatin-accessibility modifier". **v6.10.2 ran a controlled validation** (off-targets mapped to hg38, AUROC of K562 accessibility for active-vs-inactive per assay, in-vitro negative controls): the in-vitro controls are ~0.5 (method sound), GUIDE-seq is a textbook modest positive (**0.58**, CI excludes 0.5) but TTISS **reverses** (0.346) — the cross-cell-type K562 proxy is the likely cause. **VERDICT: weak/inconsistent → NOT a validated quantitative axis.** Chromatin stays a documented **annotation only** (does not change the numeric risk score), labelled `validated=false`. Full result: `benchmarks/offtarget/chromatin_validation.json`. The definitive refinement (cell-type-matched ENCODE accessibility) is deferred. *This corrects the v6.10.1 ledger entry, which had marked it "FIXED" — that was too strong.* |
+| 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. |
+
+## v6.9 — PEN-IMMUNE (Stage G) — *audited retroactively, corrected in v6.10.1*
+
+| Item | Status | Detail |
+|---|---|---|
+| "OOD-gated" in the MHC-II axis status string | **FIXED (v6.10.1)** | No distributional OOD gate is computed for an in-panel sequence; the axis is **coverage-gated** (abstains when the antigen is uncached). Wording corrected in `immune_mhc2.py`. |
+| Stale `AXIS_STATUS["mhc2_writer"].reason` + module docstring described the dropped v6.9.0 P1-anchor proxy as "the method" | **FIXED (v6.10.1)** | Updated to state the production method is the real **NetMHCIIpan-4.0** (v6.9.1/6.9.2); the P1-anchor density is offline-triage only. |
+| Plan-named MHC-II tools (MixMHC2pred, Graph-pMHC, HLAIIPred, FIONA, iTope) + ADA datasets (IEDB held-out, ImmunoSeq-217, FVIII/ATHN/MLOF inhibitor sets) | **DISCLOSED** | The MHC-II axis is **NetMHCIIpan-4.0 only** (a single real tool, not the named ensemble); the Immuno-Bench panel is the 4 bundled origin-labelled proteins, **not** an IEDB/ADA held-out leaderboard. The named datasets/tools were **not used**. The single-tool result is real; the broader ensemble/benchmark is deferred. |
+| Immuno-Bench "recovery" circularity | **DISCLOSED** | `recovery()` scores 3 foreign + 1 self protein; because `ada_risk = density × foreignness(origin)` and `foreignness(self)=0`, the self control is **zeroed by construction** — "foreign > self" cannot fail. It is a sanity check, not the IEDB/ADA held-out leaderboard the plan named. The ADA-incidence absence (stays 🟡) was already disclosed; this circularity note is new. |
+| MHC-II metric changed (peptide-fraction → residue-coverage) between 6.9.1 and 6.9.2 | **DISCLOSED** | The 4× difference between the §6b (0.153…) and §6c/shipped (0.636…) numbers is a **metric-definition change** (now residue coverage), not a contradiction. No live surface emits the old numbers. |
+
+## v6.8 — PEN-WRITER (Stage C) — *audited retroactively*
+
+| Item | Status | Detail |
+|---|---|---|
+| attB was a poly-G/C **schematic** at v6.8.0, surfaced as a design candidate | **FIXED (v6.9.2)** | v6.8.0 emitted `("G"*18)+core+("C"*18)` and only tested the central dinucleotide, never disclosing the attB sequence was a placeholder. The **real documented Bxb1 minimal attB** (FlyBase FBto0000359; Ghosh 2003) replaced it in v6.9.2. Disclosed here as a v6.8.0 limitation. |
+| Inter-curator agreement (pre-registered C-WS1 acceptance) | **N/A** | Single contributor (`ahmedanees-m`) — inter-curator agreement is structurally inapplicable; the curation is single-author with per-row DOI + verbatim quote. |
+| "≥1 external submission" (success criterion) | **DISCLOSED** | Not met; forward-looking. |
+| Writer predictor gate (beats baseline on held-out **family**) | **OK (already disclosed)** | It wins on held-out **locus** only, **not** family; KB ranking retained. Stated verbatim in code/docs — no fix needed. |
+
+## v6.7 — PEN-EXPRESS (Stage H expression) — *audited retroactively*
+
+| Item | Status | Detail |
+|---|---|---|
+| TPE-Bench headline: held-out-**cell-type** (plan) vs held-out-**chromosome** (shipped) | **OK (already disclosed)** | The live track is labelled `chrom_holdout` and the cell-type track is labelled `DATA-GATED` everywhere — an honestly-disclosed scope narrowing, not covert. |
+| "chromosome-Mondrian" served interval | **FIXED (v6.10.1)** | Per-chromosome Mondrian qhats are computed, but the **served** band uses the global qhat (a query has no chromosome at serve time — which is correct). Wording corrected in `position_effect.py` to avoid implying per-query Mondrian serving. |
+| Prereg consolidation (`ws_expr_{d,m,u,cal}` → single `ws_expr2.yaml`) | **DISCLOSED** | The four planned preregs were consolidated into one; the `cal` (wet-lab) prereg was dropped with the user-scoped wet-lab omission. |
+| HEK293T-OOD acceptance demo + scope-manifest entry | **DISCLOSED** | Deferred with the data-gated cross-cell-type transfer track (needs a second cell-type epigenome). The OOD detector itself ships and is wired; titer/patient state are registered known-unknowns stack-wide. |
+
+---
+
+*This ledger is updated on every release. If you find a deviation not listed here, it is a bug in our honesty, not
+an intended omission — please flag it.*

pen_stack-6.10.2/docs/cards/offtarget_data.md

@@ -0,0 +1,85 @@
+# 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; 4 assays as of v6.10.1)
+| Assay | Setting | Guide panel | Citation | DOI |
+|---|---|---|---|---|
+| GUIDE-seq | cell-based, unbiased | canonical (8) | Tsai et al., *Nat Biotechnol* 2015 | `10.1038/nbt.3117` |
+| CIRCLE-seq | in vitro (cell-free), unbiased | canonical (8) | Tsai et al., *Nat Methods* 2017 | `10.1038/nmeth.4278` |
+| CHANGE-seq | in vitro, high-throughput | independent broad (20) | Lazzarotto et al., *Nat Biotechnol* 2020 | `10.1038/s41587-020-0555-7` |
+| SITE-seq | in vitro biochemical | independent broad (11) | Cameron et al., *Nat Methods* 2017 | `10.1038/nmeth.4284` |
+
+Canonical Cas9 guides: **EMX1, VEGFA site 1/2/3, FANCF, HEK293 site 2/3/4**. CHANGE/SITE-seq use **independent broad
+guide panels** (not the canonical 8) — a cross-assay generalization test. 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)
+CRISOT-Score is the MD-physics, **assay-agnostic** scorer (not fit on these labels) → a leakage-clean held-out
+evaluation on every assay.
+
+| Assay | guide panel | CRISOT AUPRC | homology AUPRC | gap | 95% CI (held-out-guide bootstrap) | beats |
+|---|---|---|---|---|---|---|
+| GUIDE-seq | canonical (8) | 0.646 | 0.467 | +0.179 | [0.014, 0.329] | ✅ |
+| CIRCLE-seq | canonical (8) | 0.520 | 0.266 | +0.253 | [0.146, 0.370] | ✅ |
+| CHANGE-seq | independent (20) | 0.541 | 0.249 | +0.292 | [0.235, 0.348] | ✅ |
+| SITE-seq | independent (11) | 0.521 | 0.233 | +0.287 | [0.239, 0.335] | ✅ |
+
+The learned predictor beats the homology baseline on **all four** assays (per-guide bootstrap CI excludes 0),
+including two independent broad guide panels (cross-assay generalization).
+
+## Chromatin-accessibility modifier (honest scope — tested, NOT validated)
+The nominator applies a documented chromatin modifier (open chromatin raises realized off-target activity;
+Lazzarotto 2020). It reads the **real Stage B accessibility track** (`phase_1/features/chromatin_{ct}.parquet`,
+ATAC/DNase) when a candidate's genomic locus + cell type are supplied AND the feature store is present (verified on
+the VM), or accepts a caller-supplied scalar; it **abstains** otherwise (the bare wheel / CI / deployed atlas do not
+ship the raw track).
+
+**v6.10.2 controlled validation** (`benchmarks/offtarget/chromatin_validation.json`): off-targets mapped to hg38
+(98.5%), AUROC of K562 accessibility for active-vs-inactive off-targets per assay, with in-vitro assays as a
+NEGATIVE control. Result — **weak / inconsistent**:
+
+| Assay | modality | accessibility AUROC | 95% CI |
+|---|---|---|---|
+| GUIDE-seq | cell-based | **0.58** | [0.550, 0.613] — excludes 0.5 |
+| TTISS | cell-based | 0.346 | [0.322, 0.368] — *reversed* |
+| CHANGE-seq | in-vitro control | 0.496 | [0.480, 0.513] — null ✓ |
+| CIRCLE-seq | in-vitro control | 0.519 | [0.501, 0.535] — ~null ✓ |
+| SITE-seq | in-vitro control | 0.469 | [0.452, 0.489] — ~null ✓ |
+
+The in-vitro controls hug 0.5 (the method has no spurious signal); the canonical cell-based assay (GUIDE-seq) shows
+the textbook modest positive even with a **cross-cell-type** K562 proxy, but TTISS reverses. **Verdict: NOT a
+validated quantitative axis** — chromatin is a documented **annotation only** (it does not change the numeric risk
+score), labelled `validated=false`. A cell-type-matched accessibility track is the definitive refinement (deferred).
+
+## 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.

{pen_stack-6.10.0 → pen_stack-6.10.2}/docs/offtarget.md

@@ -10,9 +10,16 @@ A nominated off-target is a candidate, and every result ships with the empirical
 `nominate_offtargets(writer_family, ...)` dispatches by writer family:
 
 - **Nuclease (Cas9):** given a guide + candidate sites (e.g. from a Cas-OFFinder scan), each candidate gets a
-  **mismatch-calibrated empirical risk** (the real GUIDE-seq/CIRCLE-seq active fraction at *k* mismatches — not a
-  guessed curve), the **real CRISOT-Score** when the (guide, site) is in the cached bench, and a documented
-  **chromatin modifier** (open chromatin raises realized off-target activity; Lazzarotto 2020, CHANGE-seq).
+  **mismatch-calibrated empirical risk** (the real active fraction at *k* mismatches — not a guessed curve), the
+  **real CRISOT-Score** when the (guide, site) is in the cached bench, and a documented **chromatin-accessibility
+  modifier** (open chromatin raises realized off-target activity; Lazzarotto 2020). The modifier reads the **real
+  Stage B accessibility track** (`phase_1/features/chromatin_{ct}.parquet`) when a candidate's genomic locus + cell
+  type are supplied and the store is present, accepts a caller-supplied scalar otherwise, and **abstains** when
+  neither is available (the bare wheel / current deployed atlas do not ship the raw track). A **v6.10.2 controlled
+  validation** (off-targets mapped to hg38; AUROC of accessibility for active-vs-inactive off-targets, in-vitro
+  negative controls) found the signal **weak/inconsistent** (GUIDE-seq 0.58 but TTISS 0.346; in-vitro controls
+  ~null) — so chromatin is an **annotation only** (`validated=false`), it does **not** change the numeric risk
+  score. Full result: `benchmarks/offtarget/chromatin_validation.json`.
 - **Serine integrase (Bxb1):** a cryptic **pseudo-attB** scan that seeds on the *real documented* Bxb1 attB core
   (`GCGGTCTC`, central GT; FlyBase FBto0000359, Ghosh 2003) and reports candidate cryptic sites by arm mismatches.
 - **Bridge recombinase:** delegates to the existing Perry-DMS pseudosite engine (`pen_stack.bridge.offtarget`).
@@ -22,11 +29,15 @@ candidate ENUMERATION needs the on-VM scan; this engine SCORES + RANKS + risk-ba
 
 ## The benchmark (E-G2, real data + real tool)
 
-`benchmarks/offtarget/` is a held-out-guide nomination benchmark over canonical Cas9 guides with **experimentally
-validated off-targets** (GUIDE-seq, CIRCLE-seq). The licensed **CRISOT** predictor (CC-BY-NC, run on the VM) beats
-the sequence-homology baseline on both assays — GUIDE-seq AUPRC **0.65 vs 0.47**, CIRCLE-seq **0.52 vs 0.27**, with
-the per-guide bootstrap CI on the gap excluding 0. Only derived CRISOT scores are cached/committed; the weights are
-never redistributed. See `docs/cards/offtarget_data.md` for full provenance.
+`benchmarks/offtarget/` is a held-out-guide nomination benchmark over **four** unbiased assays — GUIDE-seq +
+CIRCLE-seq (canonical guides) and CHANGE-seq + SITE-seq (**independent broad guide panels**, a cross-assay test).
+The licensed **CRISOT-Score** predictor (CC-BY-NC, run on the VM; MD-physics and **assay-agnostic** → leakage-clean)
+beats the sequence-homology baseline on **all four** — AUPRC 0.65/0.52/0.54/0.52 vs 0.47/0.27/0.25/0.23, with the
+per-guide bootstrap CI on the gap excluding 0 on each. The Off-Target-Bench also contributes a nomination task to the
+**Genome-Writing Challenge**. Only derived CRISOT scores are cached/committed; the weights are never redistributed.
+A genomic-coordinate locus split is not possible (the harmonized data ships sequences, not coordinates) — held-out-
+guide + cross-assay is the leakage-clean evaluation. See `docs/cards/offtarget_data.md` and
+`docs/DEVIATIONS_AND_DISCLOSURES.md` for full provenance and disclosed plan deviations.
 
 ## Validation-assay recommendation (`pen_stack/wgenome/offtarget_assay.py`)
 

{pen_stack-6.10.0 → pen_stack-6.10.2}/pen_stack/__init__.py

@@ -1,2 +1,2 @@
 """PEN-STACK v3.0 - open infrastructure for genome writing."""
-__version__ = "6.10.0"
+__version__ = "6.10.2"

{pen_stack-6.10.0 → pen_stack-6.10.2}/pen_stack/planner/immune_mhc2.py

@@ -6,13 +6,16 @@ protein itself is immunogenic** (Cas9 elicits MHC-II-presented CD4 responses, Si
 10.1038/s41467-021-25414-9; bridge recombinases / serine integrases are bacterial/phage) — yet Stage G scored
 only the capsid. v6.9 adds an MHC-II epitope-load axis over the **writer** as a distinct antigen.
 
-Method (grounded, dependency-free, allele-agnostic PROMISCUOUS-binder density): MHC-II presents a 9-mer core in an
-open groove; the **P1 pocket is deep and hydrophobic**, the single dominant anchor (M/F/Y/W/L/I/V) (Stern & Wiley,
-Nature 1994; Jardetzky 1996), with secondary pockets at P4/P6/P9. A *promiscuous* epitope (binds many HLA-DR) is
-exactly one with a strong P1 anchor + favorable secondaries (Southwood et al., J Immunol 1998). We count
-promiscuous-binder cores -> an epitope-density proxy. This is a **population-level, sequence-intrinsic PROXY (🟡)**,
-NOT a trained allele-specific predictor and NOT a patient-HLA-specific magnitude (a known-unknown). Whether those
-epitopes drive ADA depends on self-tolerance (foreign vs self) — handled in `ada_risk`.
+PRODUCTION method (v6.9.1/6.9.2): the **real, licensed NetMHCIIpan-4.0** eluted-ligand %Rank over a frequent HLA-II
+panel (residue-coverage metric), run on the VM with only the derived fractions cached (`configs/mhc_epitope_oracle.yaml`);
+`mhc2_epitope_load(seq, name)` returns the real value for a cached antigen and otherwise **abstains** (no production
+proxy). A population-level proxy (🟡), never a patient-HLA-specific magnitude (a known-unknown). Whether the epitopes
+drive ADA depends on self-tolerance (foreign vs self) — handled in `ada_risk`.
+
+OFFLINE-ONLY fallback (`mhc2_proxy_estimate`, NOT the production axis): a documented, dependency-free PROMISCUOUS-
+binder density — MHC-II presents a 9-mer core in an open groove whose **P1 pocket** is the dominant hydrophobic
+anchor (M/F/Y/W/L/I/V; Stern & Wiley, Nature 1994) with secondary pockets at P4/P6/P9 (Southwood 1998). Provided for
+offline triage where NetMHCIIpan cannot be run; the production axis uses the real tool and abstains otherwise.
 """
 from __future__ import annotations
 
@@ -75,8 +78,9 @@ def real_mhc2_load(name: str) -> dict | None:
             "n_covered": rec.get("n_covered"), "length": rec.get("length"),
             "method": f"NetMHCIIpan-4.0 EL %Rank<=2, {rec.get('metric', 'residue coverage')}, frequent HLA-II "
                       f"panel ({len(panel)} alleles): {', '.join(panel)}",
-            "status": "population-level (frequent-HLA panel; NetMHCIIpan-4.0 eluted-ligand), OOD-gated; NOT a "
-                      "patient-HLA-specific magnitude (known-unknown)", "backend": "netmhciipan_cache"}
+            "status": "population-level (frequent-HLA panel; NetMHCIIpan-4.0 eluted-ligand), coverage-gated "
+                      "(abstains for uncached antigens — NOT a distributional OOD gate); NOT a patient-HLA-specific "
+                      "magnitude (known-unknown)", "backend": "netmhciipan_cache"}
 
 
 def mhc2_epitope_load(seq: str, name: str | None = None) -> dict:

{pen_stack-6.10.0 → pen_stack-6.10.2}/pen_stack/twin/position_effect.py

@@ -150,7 +150,8 @@ class PositionEffectModel:
         lo, hi = self.conformal.interval(yhat, sigma=widen)
         return {"yhat": yhat, "lo": lo, "hi": hi, "ood_widen": widen,
                 "nominal_coverage": 1 - self.conformal.alpha,
-                "interval_kind": "trained split-conformal (chromosome-Mondrian), OOD-widened"}
+                "interval_kind": "trained split-conformal (chromosome-blocked OOF; per-chromosome Mondrian qhats "
+                                 "computed, GLOBAL qhat served — a query has no chromosome at serve time), OOD-widened"}
 
     # ---- persistence ----------------------------------------------------------------------------
     def save(self, path: str | Path) -> str:

{pen_stack-6.10.0 → pen_stack-6.10.2}/pen_stack/validate/immune_calibration.py

@@ -96,9 +96,10 @@ AXIS_STATUS: dict[str, dict] = {
                   "rates (and induced post-dose-1 anti-PEG is a separate dynamic)."},
     "mhc2_writer": {        # v6.9 G-WS1 — CD4/MHC-II epitope load over the writer enzyme
         "status": "mechanistic_proxy",
-        "label": "mhc2_writer: mechanistic/population proxy — NOT outcome-validated",
-        "reason": "sequence-intrinsic promiscuous MHC-II binder density (P1-anchor; Stern & Wiley 1994) is a "
-                  "presentation potential, NOT a trained allele-specific predictor or a patient-HLA magnitude."},
+        "label": "mhc2_writer: population proxy — NOT outcome-validated",
+        "reason": "real NetMHCIIpan-4.0 eluted-ligand MHC-II epitope load over a frequent-HLA panel (v6.9.2; the "
+                  "v6.9.0 P1-anchor proxy was replaced) — a population-level presentation potential, NOT a "
+                  "patient-HLA-specific magnitude or an observed-ADA-validated number."},
     "ada_writer": {        # v6.9 G-WS2 — ADA-risk (MHC-II x foreignness, self-tolerance filtered)
         "status": "mechanistic_proxy",
         "label": "ada_writer: mechanistic/population proxy — NOT outcome-validated",

{pen_stack-6.10.0 → pen_stack-6.10.2}/pen_stack/wgenome/offtarget_data.py

@@ -48,20 +48,51 @@ CANONICAL_GUIDES = {
 }
 
 # ---- DERIVED, real-data risk calibration: empirical active fraction by mismatch count -----------
-# (computed on the VM over ALL canonical-guide candidates; the off-target nomination risk band is grounded
-# in how often a candidate at k mismatches was actually validated-active, not a guessed curve.)
+# (computed on the VM over ALL candidates per assay; the off-target nomination risk band is grounded in how often a
+# candidate at k mismatches was actually validated-active, not a guessed curve. Missing k -> abstain, not extrapolate.)
 MISMATCH_ACTIVE_FRACTION = {
     "guideseq": {0: 1.0, 1: 1.0, 2: 0.76471, 3: 0.23129, 4: 0.033, 5: 0.00276, 6: 0.00014},
     "circleseq": {0: 1.0, 1: 1.0, 2: 1.0, 3: 0.67146, 4: 0.26566, 5: 0.05924, 6: 0.00985},
+    "changeseq": {0: 0.95, 2: 0.80851, 3: 0.56625, 4: 0.231, 5: 0.05505, 6: 0.00751},
+    "siteseq": {0: 1.0, 1: 1.0, 2: 1.0, 3: 0.67188, 4: 0.2491, 5: 0.03554, 6: 0.00478},
+}
+
+# ---- chromatin-accessibility validation (v6.10.2): a CONTROLLED test of whether Stage B accessibility predicts
+# off-target activity (Lazzarotto 2020). VERDICT: WEAK / INCONSISTENT -> NOT a validated quantitative axis.
+# (off-targets mapped to hg38; AUROC of K562 accessibility for active-vs-inactive per assay; full result in
+# benchmarks/offtarget/chromatin_validation.json). The in-vitro controls are ~null (method sound); GUIDE-seq is a
+# textbook modest positive (0.58) but TTISS reverses (0.346) — the cross-cell-type K562 proxy is the likely cause.
+CHROMATIN_VALIDATION = {
+    "verdict": "weak/inconsistent — NOT a validated quantitative axis on this data",
+    "auroc_accessibility_for_activity": {
+        "guideseq": {"modality": "cell-based", "auroc": 0.58, "ci95": [0.550, 0.613]},
+        "ttiss": {"modality": "cell-based", "auroc": 0.346, "ci95": [0.322, 0.368]},
+        "changeseq": {"modality": "in_vitro_control", "auroc": 0.496, "ci95": [0.480, 0.513]},
+        "circleseq": {"modality": "in_vitro_control", "auroc": 0.519, "ci95": [0.501, 0.535]},
+        "siteseq": {"modality": "in_vitro_control", "auroc": 0.469, "ci95": [0.452, 0.489]},
+    },
+    "validated": False,
+    "note": "documented biological effect (Lazzarotto 2020); on our cross-cell-type data the signal is weak and "
+            "inconsistent (GUIDE-seq 0.58 supports it, TTISS 0.346 contradicts, in-vitro controls ~null). Used as a "
+            "qualitative annotation only — it does NOT change the numeric risk score. A cell-type-matched "
+            "accessibility track would be needed to settle it.",
 }
 
 # ---- Off-Target-Bench headline (REAL full-data result; per-guide AUPRC, held-out-guide bootstrap CI) -----
+# CRISOT-Score is the MD-physics, assay-AGNOSTIC scorer (not fit on these labels) -> a leakage-clean held-out
+# evaluation on every assay. GUIDE/CIRCLE-seq use canonical guides; CHANGE/SITE-seq use INDEPENDENT broad guide
+# panels (cross-assay generalization). CRISOT beats the homology baseline on ALL FOUR (per-guide bootstrap CI > 0).
 BENCH_SUMMARY = {
-    "guideseq": {"n_guides": 8, "crisot_auprc": 0.6458, "homology_auprc": 0.4668,
-                 "auprc_gap": 0.179, "gap_ci95": [0.0154, 0.3396], "crisot_beats_homology": True},
-    "circleseq": {"n_guides": 8, "crisot_auprc": 0.5197, "homology_auprc": 0.2664,
-                  "auprc_gap": 0.2533, "gap_ci95": [0.1404, 0.3608], "crisot_beats_homology": True},
-    "metric": "per-guide AUPRC; baseline = ascending mismatch count; learned = real CRISOT-Score (VM, cached)",
+    "guideseq": {"n_guides": 8, "guide_panel": "canonical", "crisot_auprc": 0.6458, "homology_auprc": 0.4668,
+                 "auprc_gap": 0.179, "gap_ci95": [0.0136, 0.3292], "crisot_beats_homology": True},
+    "circleseq": {"n_guides": 8, "guide_panel": "canonical", "crisot_auprc": 0.5197, "homology_auprc": 0.2664,
+                  "auprc_gap": 0.2533, "gap_ci95": [0.1463, 0.3704], "crisot_beats_homology": True},
+    "changeseq": {"n_guides": 20, "guide_panel": "independent_broad", "crisot_auprc": 0.541, "homology_auprc": 0.2486,
+                  "auprc_gap": 0.2924, "gap_ci95": [0.2349, 0.3477], "crisot_beats_homology": True},
+    "siteseq": {"n_guides": 11, "guide_panel": "independent_broad", "crisot_auprc": 0.5207, "homology_auprc": 0.2332,
+                "auprc_gap": 0.2874, "gap_ci95": [0.2388, 0.3354], "crisot_beats_homology": True},
+    "metric": "per-guide AUPRC; baseline = ascending mismatch count; learned = real CRISOT-Score (assay-agnostic, "
+              "VM, cached); CRISOT beats homology on all 4 assays (bootstrap CI excludes 0)",
 }