pen-stack 7.2.3__tar.gz → 8.0.1__tar.gz

{pen_stack-7.2.3 → pen_stack-8.0.1}/CHANGELOG.md

@@ -3,13 +3,460 @@
 All notable changes to PEN-STACK are documented here. This file follows
 [Keep a Changelog](https://keepachangelog.com/).
 
+## [8.0.1] - 2026-07-13
+
+A reproducibility and API-hardening patch. No scientific result and no public function signature changes.
+
+### Changed
+- **Result, verdict, and safety types are now immutable (frozen).** `OracleResult` and its nested
+  `Provenance`, `Verdict`, `SafetyVerdict`, `ScreenHit`, `Rule`, and `RuleResult` can no longer be mutated
+  after construction, so a result's provenance and scope, a verdict's legality, or a screen hit's severity
+  cannot be edited downstream. Build a changed copy with `model_copy(update=...)`.
+- **The REST, MCP, and Gymnasium-environment tests and the SBOL3/GenBank round-trip run by default.**
+  `fastapi`, `uvicorn`, `httpx`, `fastmcp`, `gymnasium`, `sbol3`, and `biopython` are in the `[dev]` extra, so
+  the full suite exercises these features rather than skipping them. Heavy compiled extras stay opt-in.
+- **The one-command reproduction recomputes a second headline from committed source.**
+  `scripts/pa_epridict_distinctness.py` (wired into `make repro`) re-derives the durability-vs-ePRIDICT
+  distinctness result, and `scripts/reproduce.py` now distinguishes recomputed headlines from echoed sealed
+  metrics.
+
+### Fixed
+- **An abstention is no longer reported as high confidence.** A design verified with no calibrated confidence
+  is now `not-computable`, not `grounded-confident`.
+- **The frozen-input integrity check works on a fresh checkout.** `python bench/run.py --verify` parses the
+  standard `sha256sum` manifest format (two-space, CRLF-tolerant); the manifest re-seals only on explicit
+  `--seal`; a `.gitattributes` pins the hashed inputs to LF for cross-platform stability.
+- **The reproduction harness can fail.** `scripts/reproduce.py` exits non-zero if any block fails instead of
+  always reporting success.
+- Corrected the ePRIDICT variance figure in the README (R-squared 0.041, matching the committed metrics).
+
+## [8.0.0] - 2026-07-13 - Stable release of record
+
+The first stable, citable release. The major version marks a committed public API surface (the SDK
+functions, the REST endpoints, and the twenty MCP tools) governed by a one-minor-version deprecation policy.
+It consolidates the gated 7.x development trail rather than replacing it: the pre-registrations, SHA-locked
+benchmarks, and disclosed negatives that trail documents remain intact and referenced. Stable surfaces
+carry calibrated uncertainty and validation status; the mechanism-based paths (serine-integrase, bridge, and
+CAST off-target) are marked experimental and may change.
+
+### Added
+- **A one-chromosome demo atlas** (`data/demo/atlas_k562.parquet`, chr19). A fresh clone can run the
+  safe-harbour quickstart without first fetching the full release: `PEN_ATLAS_DIR=data/demo pen-stack
+  writable --gene AAVS1 --ct k562` returns the ranked writable bins of the AAVS1 (PPP1R12C) locus. Opt-in via
+  `PEN_ATLAS_DIR`, so the atlas-gated test suite is unchanged. Full genome-wide atlases come from Zenodo via
+  `scripts/fetch_artifacts.sh`.
+- **An Anthropic (Claude) provider in the LLM layer.** `pen_stack/rag/llm.py` now supports three provider
+  styles (OpenAI-compatible, Ollama, Anthropic Messages), including multi-turn tool calling, so the
+  conversational agent and the evaluations run provider-agnostically across hosted-frontier and local models.
+  Keys are read from an env var or a gitignored file and are never committed.
+- **A measured, LLM-on grounding evaluation** (`benchmarks/grounding_llm_on/`). The ungrounded-baseline
+  contrast ships a committed metrics file and an adversarial probe set, run with the model live across three
+  model families: with no tools an ungrounded model emits concrete tool-only quantities (fabrication), while
+  the grounded agent, which routes every quantity through a validated tool, emits none. Transcripts are cached
+  and replay offline in CI.
+- **An agentic baseline** (`benchmarks/agentic_baseline/`). A real external LLM agent drives the validated
+  tools and every number in its trace is audited to a direct tool call. Verified across three models spanning
+  a small local model to hosted-frontier models; the no-fabrication audit passes for all.
+- **A one-command reproduction path.** `make repro` reproduces every result that runs on the committed data
+  and the demo atlas on a fresh clone, with no download and no API key (`scripts/reproduce.py`); `make fetch`
+  then `make repro-full` add the genome-wide atlas benchmarks from the Zenodo release. Verified end to end in
+  a clean `python:3.12-slim` container.
+- **Committed metrics for the three headline analyses** (`benchmarks/position_effect_human/`,
+  `offtarget/integrase_chalberg/`, `genotox_panel/`), so the paper's central numbers reproduce directly from
+  the repository.
+
+### Changed
+- Benchmark provenance references relative paths only. The license (MIT), MCP tool count (twenty), and
+  Verify-Bench rule count (eleven) are reconciled across the package metadata, the repository, and the
+  manuscript. The repository was confirmed free of hardcoded credentials, addresses, and absolute local paths.
+
+## [7.3.21] - 2026-07-11 - World-model graph: resolve any locus (not just exact GSH names)
+
+### Fixed
+- **The world-model graph query resolved only the exact curated safe-harbour node names.** The graph materialises
+  a `locus:` node for each curated GSH locus and matched the query string exactly, so `PPP1R12C` (AAVS1's anchor
+  gene), `ROSA26` (whose node is `hRosa26`), and real genes such as TRAC, HBB, and LMO2 all returned nothing. The
+  query now resolves in three tiers:
+  - A curated node is matched by **GSH name, anchor gene, or alias, case-insensitively** (AAVS1 ≡ aavs1 ≡
+    PPP1R12C; ROSA26 ≡ hRosa26; HIPP11 ≡ H11), returning the same DOI-provenanced answer.
+  - A query that is not a curated safe harbour but resolves to a **real gene** returns the **tier-1 reprogrammable
+    writers**, which engage any locus at the locus level (the documented crosslink property), flagged as a
+    candidate rather than a curated safe harbour (per-site reachability and safety remain the Planner's and
+    Guardian's responsibility).
+  - Anything else returns an explained empty result rather than an error.
+  The response carries `resolved_locus` and `is_curated_safe_harbour`, and the Closed Loop page shows the resolved
+  locus and a curated-vs-candidate badge.
+
+## [7.3.20] - 2026-07-11 - Stage J closed loop: a UI surface + /cloudlab 500-on-bad-provider fix
+
+### Fixed
+- **`POST /api/cloudlab` returned an unhandled 500 for any `provider` other than "mock".** `submit()` raises
+  `CloudLabError` for a provider outside the wired set, but `submit_gated()` caught only the safety refusal
+  (`ProtocolExportError`), so it propagated. The endpoint now returns a **422** with the valid-provider list for an
+  unknown provider, and a null/empty/missing provider falls back to the wired `mock` dry-run path.
+
+### Added
+- **A "Closed Loop" page (Stage J).** Three closed-loop capabilities previously reachable only via the API are now
+  surfaced (the fourth, the validation campaign, is already on Experiments):
+  - **Build & submit (cloud-lab)**: a design form plus provider select drives the verify-first flow: a cleared
+    design returns a DRAFT protocol preview and a mock/dry-run job receipt (Level 3, human-in-control); a hazardous
+    cargo (e.g. ricin) is refused with no protocol emitted.
+  - **SDL-brain benchmark**: the EIG/VOI designer versus random/greedy (curve-area gap with CI), BayBE/Atlas
+    cited, with a "BayBE not installed in this deployment" note when the optional dependency is absent.
+  - **World-model graph**: a locus/cargo-form query returns the provenanced writer→locus→vehicle paths.
+  Wired into the left-rail nav (Build & learn) and the router at `/closed-loop`.
+
+## [7.3.19] - 2026-07-11 - Experiment designer: expected information gain varies with cell-type coverage
+
+### Changed
+- **`expected_info_gain` now varies across candidates** instead of reducing to the twin's fixed ±0.20 uncertainty
+  band (which read a near-constant ~0.0202 for every candidate). EIG folds in the design-varying uncertainty the
+  fixed band understates: an unmeasured cell type (no atlas data, so more reducible uncertainty) and an
+  out-of-distribution design widen the effective standard deviation. EIG is ~0.02 in a fully-measured cell
+  (K562/HepG2), ~0.43 partially-measured (HSPC), ~0.71 unmeasured (iPSC/PBMC/…), and ~1.5 out-of-distribution; it
+  is constant across gene/vehicle/cargo within one cell type, since the mechanistic twin carries no per-gene
+  uncertainty. Coverage and OOD are folded into EIG rather than double-counted as separate acquisition terms.
+
+## [7.3.18] - 2026-07-11 - Experiment designer: acquisition rebuilt on design-varying signals
+
+### Changed
+- **The experiment-suggestion acquisition is rebuilt on design-varying signals.** Expected information gain derives
+  entirely from the twin's fixed ±0.20 uncertainty band, so the suggested batch scored a near-constant ~0.02 for
+  almost every candidate, and the pool varied only vehicle × cargo at one cell type, where relative expression is
+  per-copy (cargo-independent). The acquisition now combines:
+  - **Coverage novelty**: an experiment in an unmeasured cell type reduces more uncertainty; the measured-atlas
+    roster (K562/HepG2 full, HSPC partial, others none) drives it.
+  - **Feasibility gate**: an experiment that cannot be built as specified (cargo overflowing the vehicle) is
+    down-weighted.
+  - **Immune value-of-information**, counting only **in-scope** proxy axes, so it varies by vehicle (a PEGylated
+    LNP puts the anti-PEG axis in scope; AAV does not).
+  - **Marginal gain**: the submodular greedy objective (informativeness minus overlap with the picks already
+    scheduled), distinct per experiment and decreasing down the batch.
+  The candidate pool spans all cell types × 4 vehicles so the coverage/feasibility/immune signals differ across it;
+  the batch is scored per candidate cell, and each experiment is returned with its full breakdown. The Experiments
+  page shows the marginal-gain ranking with a "why" line (coverage, immune-VOI, buildability) and states that the
+  mechanistic twin cannot finely rank two experiments in the same regime, the ranking is coverage + buildability +
+  immune-VOI + diversity. `select_batch` scores each candidate once.
+- The "Load expression-validation campaign" plan is a fixed pre-registered 48-row campaign, independent of the
+  candidate-pool form; the scores panel notes this.
+
+## [7.3.17] - 2026-07-10 - App-wide input-validation sweep (out-of-domain values are flagged everywhere)
+
+### Added
+- **A consistent input-validation pass across every page.** Shared helpers in `lib/format.js`
+  (`invalidNucleotides`, `clampInt`) and a reusable `SeqWarning` component in `ui.jsx`:
+  - **Nucleotide sequence fields** warn on characters outside A/C/G/T/U/N (kept, not silently accepted):
+    Design Studio / Twin / Guardian cargo sequence, Off-Target guide / crRNA spacer / bridge target-core, Writer
+    Atlas guide-design target and donor.
+  - **Writer Atlas "Cargo bp"** now shows the same "clamped to X (allowed 1–300,000)" note Site Finder uses,
+    instead of clamping silently.
+  - **Design Studio multiplex edits**: the free-text chromosome flags an invalid value (chr1–chr22, chrX/Y/M)
+    and the genomic position clamps to a non-negative coordinate (was `parseInt || 0`, accepting negatives).
+  - Twin chromatin marks (0–1) and cargo-size vehicle cap + non-positive (from 7.3.16) complete the set.
+
+### Fixed
+- **Twin chromatin-mark fields no longer mangle a typed negative** (`"-5"` had become `"05"`). The value was
+  coerced with `parseFloat || 0` on every keystroke, collapsing a lone leading `-` to `0`. The fields now hold
+  the raw string while editing (coerced to a number only at predict time) and drop the native `min`/`max`, so an
+  out-of-range value, including a negative, can be entered and is flagged rather than silently blocked.
+  Free-text NL fields (co-scientist chat, WriteSpec brief), amino-acid / SMILES fields (capsid, oracles), and the
+  free-form model-feature fields are intentionally left unconstrained, since client-side validation would false-warn.
+
+## [7.3.16] - 2026-07-10 - Stage-H: vehicle-aware cargo hint, OOD-widen scope, input range warnings
+
+### Fixed
+- **The "Cargo size (bp)" hint reflects the selected delivery vehicle.** It was hardcoded to "AAV single-vector
+  payload caps near ~4.7 kb" for every vehicle; it now reflects the selected vehicle's real capacity (AAV dual
+  ~9 kb, lentivirus ~8 kb, LNP ~15 kb, HDAd ~35 kb, HSV ~100 kb, electroporation = no packaging limit), sourced
+  from a `VEHICLE_CAP_BP` map mirroring `configs/delivery_vehicles.yaml`, with an inline over-capacity warning when
+  the entered cargo exceeds the cap.
+- **The Stage-H "OOD widen" factor and its presentation.** The factor reads near its ×3 ceiling for most realistic
+  chromatin input. It is genuinely computed (near-zero input ≈ ×1, mid-range ≈ ×2.4–×2.8, cap ×3), but the
+  position-effect head was trained on normalized exact-site features, so typical 0–1 histone values at the app's
+  coarse 1 kb resolution sit outside that distribution and the widen flags them as out-of-distribution (near the
+  cap). The stat caption, the not-outcome-validated banner, and the scores panel now explain that a high widen at
+  this resolution is the signal working, the same cause as `outcome_validated: false`, rather than a fixed value.
+
+### Added
+- **Range warnings on out-of-domain inputs.** The five chromatin marks are documented normalized signal in [0,1];
+  the fields previously accepted e.g. -5 and fed it to the model unflagged. They are now bordered and warned when
+  outside 0–1 (value kept, not clamped, so the OOD signal still reflects it). A non-positive cargo size is flagged
+  (`feasibility.buildable=False` + `cargo_size_nonpositive` scope flag) instead of reading as buildable.
+- **A promoter/cell-type note.** Selecting a tissue-specific promoter (e.g. TBG = liver) surfaces that the
+  mechanistic estimate is cell-type-agnostic and applies the same strength regardless of tissue match, so a
+  maximally-optimistic score for a biologically inappropriate promoter/cell-type pairing is not misread.
+
+## [7.3.15] - 2026-07-09 - Health check surfaces writability-atlas presence (deploy-mount visibility)
+
+### Fixed
+- **`/api/health` reports whether the per-cell-type writability atlases are present,** not just the bundled
+  writer atlas. A deploy that drops the runtime `-v ~/data/out:/work/data/out` mount leaves the container with no
+  per-cell-type atlas parquets, so Site Finder returns `atlas_<ct>.parquet not found` and the cell-type dropdown
+  shows "no atlas" for every cell type (including the measured K562/HepG2/HSPC), yet `/api/health` stayed green
+  because `atlas_present` checks only the bundled writer atlas. Health now includes `writability_atlas:
+  {measured_count, measured, ok}`; `measured_count == 0` makes a dropped-mount deploy immediately visible.
+
+## [7.3.14] - 2026-07-09 - Digital Twin: surface the Stage-H model's OWN OOD signal + disambiguate the two OOD checks
+
+### Fixed
+- **The Twin page's top-level "OOD" flag is disambiguated from the Stage-H model's own OOD signal.** The
+  top-level flag is bound to the **virtual-cell response oracle's** scope card (its list of in-distribution cell
+  types and perturbation kinds), and every cell type the dropdown offers is on that list, so it reads "no" through
+  this UI by construction, not a broken gate. The Stage-H **position-effect** model has its own separate
+  out-of-distribution signal (`ood_widen`, a Mahalanobis distance over the 5 supplied chromatin marks) that the
+  API returned but the UI never surfaced (×2.09 at typical marks, saturating at ×3.0 for marks pushed to 10–100).
+  Changes:
+  - The Stage-H card shows an **"OOD widen (this head)"** stat (×1 typical → ×3 cap), colour-graded, so the
+    position-effect model's own OOD signal is visible and moves with the chromatin context.
+  - The top-level stat is relabelled **"OOD (response model)"** with an inline note that it is the response
+    oracle's scope-card check (expected to read "no" for any dropdown cell type), a different signal from the
+    Stage-H widen factor.
+  - The "How to read these scores" panel documents both OOD checks separately.
+- A regression test locks in that `position_effect.ood_widen` widens for out-of-range chromatin marks.
+
+## [7.3.13] - 2026-07-09 - Guardian: FASTA-header robustness fix + Design Studio Proof detail parity
+
+### Fixed
+- **A sequence pasted straight from a database (a `>accession description` FASTA header plus wrapped sequence
+  lines) is now screened correctly.** `_looks_like_nucleotide()` ran on the raw pasted text, header included; the
+  header's non-nucleotide characters diluted the nucleotide fraction below the classification threshold, so the
+  header+sequence blob was misclassified as protein and scanned as one string, with no 6-frame translation. The
+  screen now strips FASTA/GenBank header lines (`>`, `;`) and whitespace before the nucleotide-vs-protein decision
+  (`_strip_headers_and_whitespace`) and cleans stray non-amino-acid characters in the protein branch. A
+  FASTA-formatted toxin sequence that previously returned zero hits now refuses.
+- **Design Studio's Proof object (`/api/verify/proof`) carries the full hazard detail.** A hazard hit was reduced
+  to a bare `{signature, kind, severity}` with no reason, citation, or (for a sequence-domain hit) the Pfam
+  accession/bit score that Guardian's `/api/safety` view already showed. `_biosecurity_axis()` now carries `reason`
+  (the matched signature name), `citation` (the control-list reference), and `pfam_accession`/`bit_score` when the
+  hit came from the sequence screen; the Proof panel renders them as Guardian's hit list does.
+
+## [7.3.12] - 2026-07-09 - Guardian: local Pfam/HMMER sequence-domain screen (the Cargo sequence box is now screened)
+
+### Added
+- **A raw Cargo sequence (DNA/RNA/protein) is screened for the curated toxin families, not only a declared
+  Cargo function.** `sequence_homology()` was previously a no-op in this deployment (its `external_hook` was an
+  unwired extension point). `HazardRegistry` now has a `.default()` registry that wires
+  `pen_stack/safety/pfam_scan.py` as that hook: a cargo sequence is 6-frame-translated (DNA/RNA) or scanned
+  directly (protein) against bundled public Pfam profile HMMs for the same 17 accessions already in
+  `configs/safety/hazard_registry.yaml` (ricin/RIP, botulinum, diphtheria, anthrax, cholera, staph/strep
+  enterotoxin, conotoxin), using each profile's Pfam gathering (GA) cutoff, the threshold Pfam's own annotation
+  pipeline uses to call a family match. No hazard sequences are bundled: profile HMMs are public statistical
+  models built from many aligned sequences, not any one organism's sequence, consistent with the registry's
+  existing no-hazard-sequences policy. The screen reaches every entry point unchanged (`/api/safety`,
+  `/api/verify`, `/api/generate`, the Co-Scientist chat's design lane), since all funnel through `screen_design()`
+  → the default registry: a sequence for a curated family refuses with no cargo_function declared, while benign
+  cargo (a GFP CDS fragment, a Factor IX fragment) stays clear. Scope: this covers the curated family list only
+  and is not a substitute for a full external homology screener (IBBIS Common Mechanism / SecureDNA) over a larger
+  reference set; an integrator can override via `HazardRegistry.load(external_hook=...)`.
+- `provenance.declared_signal` (v7.3.11) now counts a submitted `cargo_seq`/`cargo_sequence` as signal, since it
+  is genuinely screened.
+
+## [7.3.11] - 2026-07-09 - Guardian: distinguish an empty submission from a screened-clear result + Cargo function autofill fix
+
+### Fixed
+- **An empty Guardian submission is distinguished from a screened benign design.** With no cargo function, domain
+  tag, or taxon declared, `safety_gate` has no signal and correctly decides `clear`, but the badge and reason
+  ("no hazard signal") were indistinguishable from a benign design ("human factor IX") that was actually checked.
+  `SafetyVerdict.provenance` now carries `declared_signal` (whether anything screenable was submitted); the reason
+  text reads "nothing was screened" when it is `False`, and the Guardian page renders a caveat: *"Nothing was
+  screened. 'Clear' here means there was no input to check, not a verified-safe result."*
+- **Cargo function field text duplication.** The field had no `autoComplete` attribute, leaving it open to the
+  browser's autofill/suggestion heuristics interfering with a controlled input. `autoComplete="off"` is set on the
+  Cargo function and Gene/target inputs and the Cargo sequence textarea (the app's React state is the single source
+  of truth for these fields).
+
+## [7.3.10] - 2026-07-08 - Stage H: promoter selector + cargo-capacity feasibility + mechanistic/learned transparency; multiplex verify fix
+
+### Fixed
+- **The multiplex plan (v7.3.9) now triggers its rule.** A multi-edit plan was sent with the original
+  `write_type` (e.g. `insertion`), whose rule categories exclude `multiplex`, so `multiplex.translocation_risk`
+  never ran (empty `soft_flags`). The plan now routes as `write_type="multiplex"`, so the translocation screen
+  fires (two DSB nucleases on different chromosomes → risk 0.64 flag; a DSB-free plan → ~zero).
+- **The Digital Twin's top-level "Relative expression" reflects promoter and vehicle changes; its
+  cell-type-agnostic, per-copy basis is now explicit.** The mechanistic estimate is
+  `promoter × copy-number × accessibility × pre-existing-NAb`, with accessibility neutral (no cell-type chromatin
+  track mounted), **cell-type-agnostic** (the cell-type position effect is the learned Stage-H head) and
+  **per-copy** (independent of cargo size), so it did not move with cell type or cargo size. The page now (a)
+  exposes the **31-promoter palette** as a selector, a live lever the estimate reacts to (`GET /twin/promoters`,
+  `design.promoter`); (b) surfaces the **learned head's cell-type-aware relative expression** next to the
+  mechanistic headline, so a cell-type change is visible; (c) flags **cargo-over-capacity** as not-buildable
+  (`feasibility` in `/predict`); (d) states the mechanistic basis inline.
+
+## [7.3.9] - 2026-07-08 - Design Studio: multiplex-plan UI + advisory (soft-penalty) flags surfaced
+
+### Added
+- **A "Plan multiple simultaneous edits" section on Design Studio** builds a 2–5-edit plan (chromosome / position /
+  writer-family rows) and sends it as `design.edits`, making the `multiplex.translocation_risk` legality rule
+  (implemented and rule-spec-published since v3.1, but with no UI path to `design.edits`) reachable. The default
+  two-row example matches the rule's own test fixture (two DSB nucleases on different chromosomes) and reproduces a
+  "flag" on first Verify.
+- **Advisory (soft-penalty) flags are rendered on the Verify result.** `Verdict.soft_flags` was always computed
+  (multiplex translocation risk, AAV packaging margin, delivery sequence constraints, fold cross-loop
+  complementarity) but never rendered, since soft penalties do not fail legality and the Proof object's legality
+  axis never carried them. A new "Advisory flags" panel shows the rule id, reason, and value for any soft-penalty
+  rule that fires, labelled non-blocking.
+
+## [7.3.8] - 2026-07-08 - Stage F: not_evaluated state under biosecurity refuse + standards concordance surfaced
+
+### Fixed
+- **The legality and confidence proof axes no longer report a verdict when biosecurity refuses a design.**
+  `verify.service.verify()` short-circuits on a biosecurity refuse and returns *before* the rules engine or
+  confidence calibration run (`legal=None`, `violations=[]`), a security-first fail-fast posture. `verify_proof()`'s
+  legality axis previously read that state as `status="fail"` with an empty violated list and no repair hint, and
+  the confidence axis read it as `status="abstain"` with a repair hint suggesting "supply the soft per-axis scores
+  to calibrate", which would not help. Both axes now report **`status="not_evaluated"`**, `ok=True` (an
+  unevaluated axis does not itself block; biosecurity already blocks `passable`), with an explanatory,
+  non-actionable repair hint. A genuine legality failure with no biosecurity hazard is unaffected (regression-tested).
+
+### Added
+- **Standards concordance (v6.12) is reachable via REST and rendered on the Guardian page.**
+  `concordance_report()` and `align_to_common_mechanism()` were implemented and unit-tested since v6.12 but had no
+  API endpoint or UI surface. `GET /safety/concordance` serves the labelled-probe-set report (8/8 concordant);
+  `POST /safety` additively carries a per-decision `standards` field (Common Mechanism ScreenStatus + SecureDNA
+  outcome), leaving the underlying decision unchanged. Guardian.jsx renders the per-screen standards alignment and
+  a concordance-benchmark card.
+
+## [7.3.7] - 2026-07-08 - Writer immunogenicity: surface the human self control so the ADA foreignness term is visible
+
+### Added
+- **The Writer Atlas immunogenicity table surfaces the human self control (albumin) as a labelled reference
+  row.** Every genome writer is foreign (foreignness = 1), so the shipped rows could not show that
+  `ada_risk = MHC-II density × foreignness(origin)` is multiplicative, the numbers were indistinguishable from
+  `ada_risk = MHC-II load`. The self control (origin = self → foreignness 0) has a non-zero MHC-II density (0.32)
+  yet ADA-risk 0, so ADA decouples from the load: the row demonstrates that the foreignness term does real work
+  (central tolerance). `writer_immunogenicity_table()` appends it with `is_control: True`; the UI highlights it and
+  explains the decoupling; a test asserts control ada_risk 0 versus foreign writers ada_risk > 0.
+
+## [7.3.6] - 2026-07-08 - Design Studio: immune-risk profile no longer shows stale numbers when the main vehicle changes
+
+### Fixed
+- **The Immune-risk profile card no longer desyncs from the main form's Delivery vehicle.** Changing the main-form
+  vehicle dropdown moved the profile card's highlighted vehicle chip (it tracked `design.delivery_vehicle`) while
+  the axis numbers stayed on the previously-profiled vehicle, so AAV numbers could be read as if they were LNP's.
+  The chip highlight now tracks the vehicle the profile was actually **computed for** (`immVehicle`), so chip and
+  numbers agree, and a banner appears when the main-form vehicle diverges ("These axes are for X … click its chip
+  or press Profile again to recompute for Y").
+
+## [7.3.5] - 2026-07-08 - Off-Target auto-scroll hotfix (instant, not smooth)
+
+### Fixed
+- **The v7.3.4 auto-scroll-to-result now moves the viewport.** `scrollIntoView({behavior:"smooth"})` no-ops on
+  this app's scroll container; dropping `behavior:"smooth"` makes the Off-Target result reliably scroll into view
+  on the first click.
+
+## [7.3.4] - 2026-07-08 - Off-Target page: scroll the result into view + guard rapid writer-class switches
+
+### Fixed
+- **The Off-Target finder result scrolls into view when it renders.** The page was the only finder missing the
+  shared auto-scroll-to-result pattern, so a result rendered below the tall explainer cards read as "no visible
+  output" on the first click. Adds a `resultsRef` and a `scrollIntoView` effect on busy/result/error, matching
+  Site Finder / Write Spec / Writer Atlas.
+- **A slow previous-writer-class response can no longer clobber the current one.** `run()` tags each request and
+  only the latest may write `res`/`error`/`busy` (a request-id guard), so switching writer class and clicking in
+  quick succession renders the current selection's result.
+
+## [7.3.3] - 2026-07-07 - Off-Target page copy: reflect SITE-seq, the chromatin annotation, and Cas12a in the on-page description and score guide
+
+### Changed
+- **Off-Target page `WhatYouGet` / `ScoreGuide` copy now documents the v7.3.1 capabilities** that were shipped in
+  the mechanics but not the descriptions: SITE-seq is named as the fourth grounding/confirming assay; the chromatin
+  accessibility annotation is described (validated, HEK293T DNase GUIDE-seq AUROC 0.671, annotation-only, not a
+  re-ranker); and Cas12a is noted as enumeration-supported but SpCas9-scorer-scoped (the finder abstains rather than
+  mis-scoring). No behavior change, copy only.
+
+## [7.3.2] - 2026-07-07 - Digital Twin: surface the learned Stage-H position-effect head (served model + resolution-aware validation state)
+
+### Added
+- **Stage-H position-effect head surfaced on the Digital Twin page (`Twin.jsx`).** When a chromatin context is
+  supplied, the page shows which learned head served (human K562, Leemans 2019 / mESC TRIP), the served feature
+  resolution, and the `outcome_validated` flag, plus the learned log2-expression, its conformal interval and
+  p(silenced). An optional chromatin-context input (the head's five histone marks) enables the learned head from the
+  page. The human K562 head's validated regime (ρ≈0.57 at exact-site per-locus features) is distinguished from the
+  app's coarse 1 kb serving resolution, at which the head serves with `outcome_validated: false`.
+- The `stage_h_mode` (learned vs heuristic) is shown as a pill on the predicted-outcome card.
+- The human K562 position-effect head (`models/position_effect_human_k562.pkl`, Leemans 2019) and the mESC head
+  (`models/position_effect.pkl`) are staged into the served image so the v7.3.0 head loads at serve time (it was
+  wired in v7.3.0 but its artifact was not mounted); the artifacts are re-serialized in the serving image
+  (scikit-learn 1.9.0).
+
+### Changed
+- The Digital Twin `WhatYouGet` / `ScoreGuide` copy documents the learned Stage-H head and the exact-site vs coarse
+  1 kb resolution distinction.
+
+## [7.3.1] - 2026-07-07 - Off-Target finder: SITE-seq in the assay recommender, chromatin-annotation surfacing, and Cas12a scoping
+
+### Added
+- **SITE-seq in the nuclease validation-assay recommender.** `offtarget_assay.recommend_assay` now lists SITE-seq
+  (Cameron et al., Nat Methods 2017; DOI 10.1038/nmeth.4284) alongside GUIDE-/CHANGE-/CIRCLE-seq, completing the
+  four unbiased assays the mismatch-calibrated risk band is grounded on.
+- **Cas12a in the Off-Target finder writer-class selector (AsCas12a).** Genome-wide enumeration supports the Cas12a
+  TTTV 5' PAM, but the validated CRISOT-Score and the mismatch-calibrated risk band are SpCas9-specific; the finder
+  therefore reports that a validated Cas12a off-target scorer is not available rather than scoring Cas12a sites with
+  the SpCas9 model.
+- **Chromatin-accessibility annotation surfaced on the nuclease finder result**: its standalone validation
+  (cell-type-matched HEK293T DNase; GUIDE-seq AUROC 0.671) plus a per-site accessibility column when the Stage B
+  accessibility track is mounted.
+
+### Changed
+- **The nuclease off-target `method` string and the `nominate_offtargets` capability manifest now describe the
+  chromatin layer precisely**, a validated annotation surfaced when the Stage B track is mounted (annotation-only,
+  not a re-ranker), instead of describing it unconditionally.
+
+### Fixed
+- **The nuclease finder no longer returns a chromatin field with no context when the Stage B accessibility track is
+  absent.** The result now carries `chromatin_available` and a note stating the annotation is track-gated and the
+  CRISOT-driven ranking is unaffected, in place of a silent null.
+
+## [7.3.0] - 2026-07-04 - PEN-EXPRESS-HUMAN: a cell-type-matched Stage-H position-effect head for human K562 (Leemans 2019), replacing the non-transferring cross-species axis; served claim is resolution-aware
+
+### Added
+- **Human K562 position-effect head behind Stage H.** `twin.data.position_effect` now wires the **Leemans 2019**
+  human K562 dataset (Cell; PMID 30982597; DOI 10.1016/j.cell.2019.03.009) with a real loader (`_load_leemans`),
+  per-barcode coordinates fetched from the Cell supplement Dataset S2 joined to the van Steensel feature set.
+  `twin.position_effect.load_human_k562_model` + cell-type routing in `predict_stage_h` serve a `PositionEffectModel`
+  trained on human K562 for human K562 designs, falling back to the mESC model otherwise.
+- Held-out validation (chromosome-blocked 5-fold OOF; n=9,298; 5 shared histone marks): expression Spearman
+  **ρ ≈ 0.59**, split-conformal held-out coverage **0.90** (nominal 0.90, within tolerance). This beats the mESC
+  model applied cross-species, which is an honest null on human K562 (held-out ρ ≈ 0.13, does not beat the
+  lamina/heterochromatin baseline ρ ≈ 0.39), and the trivial "avoid heterochromatin" baseline.
+
+### Changed
+- **The v6.7 cross-cell-type transfer axis is no longer data-gated.** With a human cell type now fetched, a human
+  K562 design's Stage-H provenance/scope reflect **human K562 supervision** instead of "single_context_supervision
+  (mESC), cross-cell-type transfer data-gated". `predict_stage_h` gains a `cell_type` argument (default `None`;
+  backward compatible; `twin.outcome` passes the cell state through).
+
+### Serve-time honesty, resolution-aware `outcome_validated` scope
+- The head's held-out validation (ρ ≈ 0.57 / +0.065 over a learned-heterochromatin baseline, CI [0.026, 0.103])
+  holds at **exact-site per-integration** chromatin-feature resolution. On coarse 1 kb-binned features even a
+  human-trained model realizes only **ρ ≈ 0.16** (P2_MATCHED_TRANSFER). `predict_stage_h` therefore stamps the
+  `human_K562_outcome_validated` scope **only when the caller declares exact-site features** (`feature_resolution`
+  / `chromatin_features_resolution`, or `chromatin_ctx.resolution` ∈ exact-site set). Otherwise, including
+  undeclared resolution (conservative default), it returns `outcome_validated: False` and a downgraded flag
+  (`human_K562_head_served_UNVALIDATED_RESOLUTION …`) naming both numbers and the requirement. The prediction is
+  still served; only the validation claim is withheld until the served resolution matches the validated regime.
+  The output gains `served_feature_resolution` and `outcome_validated` fields.
+- **Deferred to v7.4:** exact-site chromatin-feature extraction *at serve time* (point-query ChIP for an arbitrary
+  locus), so the served path realizes the validated ρ ≈ 0.57 directly and auto-earns the `outcome_validated` scope.
+
+### Notes (honest scope)
+- The cassette factorization (`f_cassette + g_context`) adds no value on Leemans' 6 similar promoters (factored
+  ρ 0.588 vs context-only ρ 0.592); additive suffices, reported not hidden. The served silenced (escaper/repressed)
+  call is promoter-driven, not a position signal. The head's validated output is steady-state **position-effect
+  expression in K562**, not temporal durability or other cell types.
+- **Cross-species transfer is poor** (matched-source mouse↔human ρ ~0.08–0.20; P2_MATCHED_TRANSFER). The axis is
+  cell-type-specific; train a head per target cell type. An earlier draft over-read a feature-source-confounded
+  ρ 0.505 as transfer; corrected.
+- The model artifact + Leemans parquet are deployment data (git-ignored under `data/`/`models/`, per the data
+  policy), not shipped in the wheel. Absent them, the human loader returns `None` and Stage H is unchanged.
+
 ## [7.2.3] - 2026-07-02 - Off-Target: professional wording (self-contained, no external-tool comparisons) + a plain-language explainer
 
 ### Changed
 - **Removed all external-tool comparisons** ("like CRISPOR / CHOPCHOP", "CRISPOR-class", "CRISPOR-comparable") from
-  every off-target surface — the Off-Target page, nav, code docstrings, REST docstring, the data card, `positioning.md`,
+  every off-target surface, the Off-Target page, nav, code docstrings, REST docstring, the data card, `positioning.md`,
   the prereg (re-SHA-locked), paper-4 prereg, and the v7.2.0 CHANGELOG entry. The capability is now described on its
-  own terms: a genome-wide off-target *finder* — given a guide/target it enumerates and ranks the off-target sites
+  own terms: a genome-wide off-target *finder*. Given a guide/target, it enumerates and ranks the off-target sites
   itself. No functional change.
 - **Off-Target page now explains off-target effects in plain language**: a new top card covers *what an off-target
   effect is*, *what you enter* (per writer class), and *what you get back* (ranked genome-wide sites with
@@ -19,25 +466,25 @@ All notable changes to PEN-STACK are documented here. This file follows
 
 ## [7.2.2] - 2026-07-02 - Integrase: precise capability framing (the negative is about a method, not the biology)
 
-A framing/presentation refinement of the v7.2.1 sealed-negative result — no science changes, the number is
+A framing/presentation refinement of the v7.2.1 sealed-negative result: no science changes; the number is
 unchanged and the negative stands.
 
 ### Changed
 - **Precise claim, everywhere:** what the sealed PhiC31 benchmark showed is that a specific *method* (attP
-  sequence-similarity scanning) is insufficient — **NOT** that pseudosites are unpredictable. The signal is real
+  sequence-similarity scanning) is insufficient, **NOT** that pseudosites are unpredictable. The signal is real
   (Chalberg's palindromic consensus; IntQuery's success on 410,776 cryptic attB), just not capturable by simple
   identity. So the honest, narrow claim is "sequence similarity is insufficient," and this negative is the
   empirical justification for a learned/DMS model (the indicated next step). Sharpened in `offtarget_integrase.py`,
   `phic31_recall_metrics.json`, and the data card so nothing reads as "unpredictable."
 - **App presentation:** `nominate_integrase` now returns a `capability` disclosure (badge + plain microcopy). The
   Off-Target page shows a badge reading **"mechanism-based · not predictive by sequence alone"** (not a bare red
-  flag) plus one line — *"φC31 pseudosite activity is sequence-guided but not predictable from similarity alone
-  (sealed benchmark, negative); confirm empirically"* — separating the grounded facts (verified att + documented
+  flag) plus one line, *"φC31 pseudosite activity is sequence-guided but not predictable from similarity alone
+  (sealed benchmark, negative); confirm empirically"*, separating the grounded facts (verified att + documented
   pseudo-attP with GenBank accessions) from the honest capability limit. A trust-building disclosure, not an alarm.
 
 ## [7.2.1] - 2026-07-02 - PhiC31 integrase: close the O-WS3/O-G2 data gap with verified data + a sealed recall benchmark (honest NEGATIVE)
 
-Closes the one genuine open item from v7.2.0 — the PhiC31 integrase data gap — with independently-verified,
+Closes the one genuine open item from v7.2.0, the PhiC31 integrase data gap, with independently-verified,
 open, citable sequences (no fabrication) and a **sealed recall benchmark whose result is reported verbatim**.
 
 ### Added
@@ -49,16 +496,16 @@ open, citable sequences (no fabrication) and a **sealed recall benchmark whose r
   DOI `10.1128/MCB.21.12.3926-3934.2001`; chr8/16/15). Corrected DOIs: Groth 2000 `10.1073/pnas.090527097`,
   Chalberg 2006 `10.1016/j.jmb.2005.11.098` (was `...11.108`).
 - **Sealed PhiC31 pseudo-attP recall benchmark** (`benchmarks/offtarget/integrase/`, O-G2): does attP-sequence-
-  similarity recover the documented pseudo-attP above length-matched GRCh38 background? **Result — NEGATIVE,
+  similarity recover the documented pseudo-attP above length-matched GRCh38 background? **Result: NEGATIVE,
   reported verbatim:** all three sites score 14 mm / 30 bp (53.3% identity), exactly the background **median**
   (60.6% / 77.0% / 82.4% of random windows are as-or-more attP-similar). Sequence identity to attP is **not** a
-  validated PhiC31 pseudo-attP predictor — φC31 recognition depends on palindromic architecture (Chalberg 2006) /
+  validated PhiC31 pseudo-attP predictor; φC31 recognition depends on palindromic architecture (Chalberg 2006) /
   a learned model (IntQuery), not raw identity.
 
 ### Changed
 - **Integrase status corrected: `semi_validated` → `mechanism_based_unvalidated` (with a sealed-NEGATIVE benchmark).**
   The v7.2.0 "semi-validated (documented pseudosites as partial ground truth)" implied the sites validated the
-  method; the sealed benchmark shows they do not. This is a *stronger, more honest* outcome — a tested negative
+  method; the sealed benchmark shows they do not. This is a *stronger, more honest* outcome, a tested negative
   rather than an untested claim. The verified att + documented pseudosites remain grounded facts and are surfaced
   as verified known off-target loci. `nominate_integrase` PhiC31 returns the documented pseudo-attP + the sealed
   benchmark; Bxb1 returns genome-wide similarity candidates carrying the same `similarity_ranking_validated: false`
@@ -67,33 +514,33 @@ open, citable sequences (no fabrication) and a **sealed recall benchmark whose r
 
 ## [7.2.0] - 2026-07-01 - PEN-OFFTGT v2: Stage E becomes a genome-wide, per-mechanism off-target FINDER (all 5 writer classes)
 
-Stage E was a candidate *scorer* — it ranked off-target sites you supplied. An off-target *finder* instead takes
+Stage E was a candidate *scorer*; it ranked off-target sites you supplied. An off-target *finder* instead takes
 a guide/target and enumerates the genome-wide off-target sites itself. PEN-OFFTGT v2 closes that enumeration gap
 AND applies the **correct off-target mechanism for each writer class**, each carrying a truthful per-mechanism
 validation status (O-WS0–O-WS9 of the plan).
 
 ### Per-mechanism paths (O-WS3–O-WS6)
-- **Serine integrase** (`offtarget_integrase.py`, O-WS3): a genome-wide **pseudo-attP** scan — a fixed-sequence
+- **Serine integrase** (`offtarget_integrase.py`, O-WS3): a genome-wide **pseudo-attP** scan, a fixed-sequence
   Cas-OFFinder scan of the attP core window over GRCh38, scored by att-arm similarity. **Bxb1** is fully encoded
   (FlyBase FBto0000359 / Ghosh 2003) and is highly specific. Status **semi-validated**. **PhiC31 is a disclosed
-  data gap** (its documented human pseudo-attP set — Chalberg 2006 — could not be verified from an open source in
+  data gap** (its documented human pseudo-attP set, Chalberg 2006, could not be verified from an open source in
   this build, so it abstains rather than fabricate). `data/curated/integrase_att.yaml`.
 - **Bridge** (`offtarget_bridge.py`, O-WS4): wraps the existing DMS-scored genome scan (measured Perry-2025
-  specificity, held-out ranking AUROC 0.88). Status **hard-locked mechanism-based-unvalidated** — the ranker is
+  specificity, held-out ranking AUROC 0.88). Status **hard-locked mechanism-based-unvalidated**; the ranker is
   DMS-validated but **no genome-wide cellular off-target assay exists** for bridge recombinases, so genomic
   recovery is unvalidated (stated in a no-ground-truth disclosure).
-- **CAST** (`offtarget_cast.py` + `data/curated/cast_systems.yaml`, O-WS5) — NEW: guide-directed spacer scan +
+- **CAST** (`offtarget_cast.py` + `data/curated/cast_systems.yaml`, O-WS5), NEW: guide-directed spacer scan +
   the distinctive **guide-independent untargeted-transposition** background per system (ShCAST/Type V-K high +
   AT-biased; VchCAST/Type I-F >95–99% on-target), all DOI-tagged. Status **mechanism-based-unvalidated**.
-- **PASTE / (ee)PASSIGE** (`offtarget_paste.py`, O-WS6) — NEW: composes the validated nuclease finder (Cas9
+- **PASTE / (ee)PASSIGE** (`offtarget_paste.py`, O-WS6), NEW: composes the validated nuclease finder (Cas9
   nickase, pegRNA) with the semi-validated integrase pseudo-attP scan (installed att); returns both component sets
   and recommends BOTH a nuclease assay and an integrase assay. Status **composite**.
 
 ### Status labels + assay recommender (O-WS7) + disclosures (O-WS8 / gate O-G3)
 - Every nomination carries a per-mechanism status label + the mechanism-appropriate confirm assay; `recommend_assay`
   now covers CAST (transposon insertion-site seq) and PASTE (dual). `docs/cards/offtarget_data.md` documents the
-  per-mechanism ground-truth status — **O-G3: every mechanism has either a sealed benchmark (nuclease) or an
-  explicit no-ground-truth / data-gap disclosure (integrase-PhiC31, bridge, CAST) — never a fabricated metric.**
+  per-mechanism ground-truth status, **O-G3: every mechanism has either a sealed benchmark (nuclease) or an
+  explicit no-ground-truth / data-gap disclosure (integrase-PhiC31, bridge, CAST), never a fabricated metric.**
 - Disclosed deviations from the plan (no silent substitution): `att_pwm.json` → an att-similarity model in
   `integrase_att.yaml` (a per-position PWM needs aligned pseudosites that do not exist at genome scale);
   `bridge_dms_specificity.parquet` → the existing measured `bridge_offtarget_profile_measured.parquet` is reused;
@@ -108,11 +555,11 @@ validation status (O-WS0–O-WS9 of the plan).
   mismatch count. Supports SpCas9 (NGG), SaCas9 (NNGRRT), and Cas12a (TTTV, 5' PAM). A full scan is heavy, so it
   runs **only on the VM** (`casoffinder:tools` Docker image, `docker/casoffinder.Dockerfile`); the live app
   **replays a committed coordinate cache** for the canonical guides or **abstains honestly** for a novel one
-  (never fabricates sites) — the same replay-or-abstain pattern as the heavy structure oracles. The enumerated
+  (never fabricates sites), the same replay-or-abstain pattern as the heavy structure oracles. The enumerated
   coordinates are facts from the public GRCh38 assembly (no license restriction), so the cache is committed
   (`data/offtarget/enumerated_cache.parquet`, 8 canonical guides, 40,268 sites).
 - **Nuclease off-target FINDER** (`pen_stack/wgenome/offtarget_nuclease.py`, O-WS2): chains enumeration into the
-  existing, validated scorer — real CRISOT-Score → mismatch-calibrated risk band → chromatin annotation — so a
+  existing, validated scorer, real CRISOT-Score → mismatch-calibrated risk band → chromatin annotation, so a
   guide returns the genome-wide ranked off-target set *with coordinates*. Scoring is unchanged from v6.10 (still
   validated: CRISOT beats homology on four unbiased assays); v2 adds the enumeration front end.
 - **O-G1 gate PASSED**: enumeration recovers **100%** of EMX1's documented GUIDE-seq off-targets within the ≤5-mm
@@ -141,14 +588,14 @@ validation status (O-WS0–O-WS9 of the plan).
     immune & delivery" action correctly scopes to the **delivery/vehicle** axes (genotoxicity, CD8 capsid, innate,
     pre-existing NAb, anti-PEG). `ImmuneProfileCard` hides the writer axes by default (`hideWriterAxes`).
   - A new **Writer immunogenicity** card on the **Writer Atlas** surfaces each genome writer's MHC-II/CD4 epitope
-    load + ADA risk + human self-match, read from the **committed NetMHCIIpan-4.0 cache — not recomputed**
+    load + ADA risk + human self-match, read from the **committed NetMHCIIpan-4.0 cache, not recomputed**
     (`GET /api/writer/immune`, `immune_profile.writer_immunogenicity_table()`).
   - **SpCas9 is removed** from the writer set: it is a nuclease/editor, not a large-cargo genome writer. The Writer
-    Atlas immunogenicity covers the actual writers — **Bxb1** (serine integrase) and **ISCro4** (bridge recombinase);
+    Atlas immunogenicity covers the actual writers, **Bxb1** (serine integrase) and **ISCro4** (bridge recombinase);
     the Cas9 nuclease and the human self control are excluded. (The engine retains SpCas9 in its cache for other
     uses; only the user-facing writer list drops it.)
 - The engine still returns the `mhc2_writer` / `ada_writer` axes from `immune_profile()` for API/MCP callers that
-  pass a `writer_family` — only the web UI moved the writer profiling to the Writer Atlas. No data was recomputed.
+  pass a `writer_family`; only the web UI moved the writer profiling to the Writer Atlas. No data was recomputed.
 
 ### Added
 - `test_ws_profile.py`: `test_writer_immunogenicity_table_for_the_writer_atlas` (covers Bxb1 + ISCro4 from the
@@ -163,11 +610,11 @@ validation status (O-WS0–O-WS9 of the plan).
   grounded administration modifier (`immune_profile._administration_modifier` / `_apply_administration`):
   - **Pre-existing NAb is muted to "no barrier" (1.0) ex vivo.** Ex-vivo delivery (cells transduced in a dish and
     washed before transplant) never exposes the vector to the patient's circulating antibodies, so pre-existing
-    anti-vector NAb does not gate eligibility — the eligibility fraction is 1.0 regardless of titer. The original
+    anti-vector NAb does not gate eligibility; the eligibility fraction is 1.0 regardless of titer. The original
     seroprevalence value is preserved (`pre_admin_value`) and the muting is explained in the axis note. This is the
     same in-vivo/ex-vivo distinction the v5.1 `delivery_immunology` profile already encodes (`computed_ex_vivo_muted`).
   - **CD8 capsid is flagged muted ex vivo** (systemic anti-capsid response minimal), but its intrinsic presentability
-    value is kept — transduced cells can still present capsid epitopes, a real residual concern, so no number is
+    value is kept; transduced cells can still present capsid epitopes, a real residual concern, so no number is
     fabricated.
   - **Genotoxicity, innate, anti-PEG and the writer axes are unchanged** by administration context (insertional risk
     and cargo/writer immunogenicity are intrinsic, not a function of circulating-antibody exposure).
@@ -182,7 +629,7 @@ validation status (O-WS0–O-WS9 of the plan).
 ## [7.1.6] - 2026-06-30 - Immune axes wired end-to-end + Delivery folded into Design Studio
 
 ### Fixed
-- **Four immune-risk axes were unreachable from the web UI and showed as out-of-scope** — innate sensing, anti-PEG,
+- **Four immune-risk axes were unreachable from the web UI and showed as out-of-scope**: innate sensing, anti-PEG,
   and the writer-as-antigen MHC-II + ADA axes abstained on every web design because the design form never collected
   their inputs (the engine computed them correctly when called directly; this was a wiring gap, not a model gap):
   - **`innate`** needs a cargo sequence. The design form now has an optional **Cargo sequence** field; the cargo
@@ -216,19 +663,19 @@ validation status (O-WS0–O-WS9 of the plan).
 ### Added
 - The `chrom` design field was a free-text pass-through with no validation and no downstream effect. It is now
   meaningful (`pen_stack/planner/chromosome.py`):
-  - **Validation** — the web form's Chromosome input is a controlled dropdown (chr1–chr22, chrX, chrY, chrM), so
+  - **Validation**: the web form's Chromosome input is a controlled dropdown (chr1–chr22, chrX, chrY, chrM), so
     an impossible value like `chrZZZ` can no longer be entered; `canonical_chromosome()` normalises/validates on
     the backend (`chrX`/`X`/`x`, `chr1`/`1`, `chrM`/`chrMT`/`MT`, `23`→X, `24`→Y; rejects the rest).
-  - **Gene/chromosome concordance** — `verify()` now flags `chromosome_mismatch` when the entered chromosome does
+  - **Gene/chromosome concordance**: `verify()` now flags `chromosome_mismatch` when the entered chromosome does
     not match the named gene's canonical location (e.g. *BRCA1 is on chr17, not chr1*), `chromosome_invalid` for a
     non-standard value, and `chromosome_unverifiable` for an unknown gene. The web form shows the warning inline
     with a one-click fix, driven by the new `GET /api/gene/location` endpoint. This closes the silent-mismatch gap.
-  - **Chromosome context** — `chromosome_context()` adds grounded, chromosome-driven advisories surfaced as scope
-    flags: **chrM** is not addressable by nuclear genome-writing tools (mtDNA needs DdCBE/TALED — out of scope),
+  - **Chromosome context**: `chromosome_context()` adds grounded, chromosome-driven advisories surfaced as scope
+    flags: **chrM** is not addressable by nuclear genome-writing tools (mtDNA needs DdCBE/TALED, out of scope),
     **chrY** is male-specific with ampliconic repeats, **chrX** is hemizygous in 46,XY vs X-inactivated in 46,XX.
 - The Design Studio **cell-type dropdown now shows writability-atlas coverage** (from `/api/celltypes`): K562 /
   HepG2 / HSPC are labelled measured (full / full / partial), while h1_hesc / iPSC / CD8 T / PBMC are labelled
-  "no atlas" — so a data-gated cell type is no longer silently indistinguishable from a measured one (the same
+  "no atlas", so a data-gated cell type is no longer silently indistinguishable from a measured one (the same
   honest-field principle as the chromosome field). The 7 cell types are unchanged and already matched the backend.
 
 ### Note (honest scoping)
@@ -274,12 +721,12 @@ validation status (O-WS0–O-WS9 of the plan).
     receptor-binding / tropism, high severity) and `FUNC-ONCOGENIC-SUPPRESSOR` (dominant-negative TP53 / RB / PTEN,
     apoptosis-checkpoint ablation, high severity), both DURC / HHS-P3CO categories.
   - **Oncogenic-manipulation PATTERN screen** (`oncogenic_manipulation` in the registry +
-    `HazardRegistry.oncogenic_flags`): a flat keyword list is brittle to paraphrase — a red-team pass found the
+    `HazardRegistry.oncogenic_flags`): a flat keyword list is brittle to paraphrase; a red-team pass found the
     Guardian caught only **1/8** mechanism/synonym phrasings ("R175H p53 abolishing transactivation", "PTEN
     knockout", "RAS G12D constitutive activation", "hTERT immortalization", "APC frameshift", "EGFR
     ligand-independent activation", "NF1+BAX/BAK knockout"). The pattern screen flags the *combination*
     `(tumor-suppressor + disruptive verb)` OR `(oncogene + activating signature)` OR `immortalization`, which
-    catches **8/8** while the deliberate asymmetry spares therapy with **no allow-list** — restoring a suppressor or
+    catches **8/8** while the deliberate asymmetry spares therapy with **no allow-list**: restoring a suppressor or
     silencing an oncogene matches neither, so "p53 correction to restore apoptosis", "TRAC/CCR5 knockout",
     "knock-down of mutant KRAS" stay clear (**11/11** benign). Disposition is escalate (dual-use → human review).
   - **Keyword matcher hardened** (`_kw_match`): a plain substring test made the ricin abbreviation `"rip"` fire
@@ -301,7 +748,7 @@ validation status (O-WS0–O-WS9 of the plan).
   pbmc / h1_hesc / ipsc) returned an empty table, because the planner-backed `candidate_space` needs `plan_write`.
   `generate_designs` now falls back to `space.vehicle_sweep(goal)`: it sweeps the capacity-compatible vehicles,
   carries `cargo_function` (the Guardian still screens), runs full legality + biosecurity discrimination, and
-  ABSTAINS on the calibrated confidence (no fabricated band) — surfaced as "abstained, no measured atlas for this
+  ABSTAINS on the calibrated confidence (no fabricated band), surfaced as "abstained, no measured atlas for this
   cell type". `candidate_space` still returns `[]` without the atlas; `generate_designs(candidates=[])` still
   returns `[]`.
 - **Designer (administration context):** the vehicle sweep now filters by the goal's `in_vivo` flag using the