pen-stack 5.8.0__tar.gz → 5.10.0__tar.gz

{pen_stack-5.8.0 → pen_stack-5.10.0}/CHANGELOG.md

@@ -3,6 +3,73 @@
 All notable changes to PEN-STACK are documented here. This file follows
 [Keep a Changelog](https://keepachangelog.com/) and the program's phase structure.
 
+## [5.10.0] - 2026-06-11 - v5.10 release: The Experiment Designer (active learning / EIG)
+
+**Closed-Loop arc, Cycle 4 of 7.** The "Learn" brain of a self-driving lab: turn *"I'm uncertain"* into *"run
+this experiment next."* Reads the calibrated v5.9 twin's uncertainty + the v5.6 immune labels, scores each
+candidate experiment by expected information gain, assembles a diverse batch, and proves on held-out data — with
+CIs — that this learns faster than random/greedy (reporting honestly when it does not). Workstreams
+WS-{ACQ,DESIGN,VALIDATE}, SHA-locked.
+
+### Added
+- **WS-ACQ** — `pen_stack/active/acquire.py`: `expected_information_gain` (reducible uncertainty from the twin's
+  predictive distribution; `≥ 0`, monotone in uncertainty), `predictive_entropy` (from the twin's interval width),
+  and **`immune_voi`** — value of information for **validating an immune PROXY axis** (v5.6): an experiment that
+  would measure a still-proxy axis is high-VOI (turns proxy → outcome-validated). `acquisition_score` is fully
+  traceable to twin quantities + v5.6 labels; deterministic; no fabricated values.
+- **WS-DESIGN** — `pen_stack/active/design.py`: `select_batch` greedily maximises acquisition **minus a
+  redundancy penalty** (shared design facets) → a **diverse** batch (not k copies of the most-uncertain point);
+  each experiment carries its expected info gain.
+- **WS-VALIDATE** — `pen_stack/active/validate.py`: `retrospective_active_learning` simulates active vs random vs
+  greedy campaigns on a held-out split, reports mean±CI learning curves and a **bootstrap CI on the curve-area
+  gap**; `active_beats_random` only when the CI excludes zero — else the not-yet-useful negative is reported.
+- **WS-BENCH** — bench **v0.3.6**: new `experiment_design` hard-gate task — the gate is the Learn engine's
+  honesty + falsifiability (twin-sourced EIG monotone in uncertainty + immune-VOI for proxy validation + diverse
+  batch + retrospective active-vs-random with reps+CI); a random selector fails by construction. Active-beats-
+  random is reported informationally.
+- Docs: `docs/experiment_design.md`; prereg `ws_{acq,aldesign,alvalidate}` + SHA locks; deposit `phase_5.10/`.
+
+### Notes
+- The experiment designer is only as good as the v5.9 twin + v5.6 labels it queries; its advantage is validated
+  **retrospectively** with CIs and reported honestly when absent. It chooses informative experiments but **does
+  not run them** — prospective benefit awaits a lab partner (v5.11+). No autonomy claim.
+
+## [5.9.0] - 2026-06-11 - v5.9 release: The Digital Twin (calibrated outcome prediction)
+
+**Closed-Loop arc, Cycle 3 of 7.** The missing layer: *what does the cell do after the write?* — predicted with
+calibrated honesty. The twin computes what mechanism allows, adds an in-distribution virtual-cell estimate
+(OOD-gated), screens immune outcome from the v5.6 profile, and is explicit about its boundary at phenotype.
+Workstreams WS-{VCELL,MECH,OUTCOME,CAL}, SHA-locked.
+
+### Added
+- **WS-VCELL** — `pen_stack/oracles/vcell.py` + scope cards `state`/`scgpt`: `predict_response(cell_state,
+  perturbation, model)` wraps **Arc STATE** / **scGPT** under the v4.0 `OracleResult` contract. A
+  perturbation-response prediction is a **candidate**, OOD-gated (a context outside the documented envelope →
+  `extrapolating`), cached/deferred (value `None` when absent — never fabricated). Encodes the field's own result
+  (Arc Virtual Cell Challenge): perturbation models don't yet consistently beat naive baselines.
+- **WS-MECH** — `pen_stack/twin/mechanistic.py`: `cassette_expression` = `promoter_strength × copy_number ×
+  accessibility` (closed-form steady state); assumptions + scope flags attached; **physics where computable, NOT
+  a phenotype**.
+- **WS-OUTCOME** — `pen_stack/twin/outcome.py`: `predict_outcome(design, cell_state)` fuses mechanism +
+  in-distribution virtual-cell response + the v5.6 immune profile into one prediction with an interval that
+  **widens under OOD**, an immune-outcome dimension, and an explicit **phenotype / in-vivo-magnitude boundary**.
+  In-vivo durability may be **conditioned on the grounded pre-existing-NAb axis** (no invented immune numbers);
+  `output_kind="candidate"`.
+- **WS-CAL** — `pen_stack/twin/calibrate.py`: `calibrate_outcome(...)` reports calibration **two-sided** — interval
+  coverage + a bootstrap CI on the MAE gap vs a naive mean baseline; the twin "beats" naive **only when the CI
+  excludes zero**, else the negative is reported verbatim; abstains at `N < 3`.
+- **WS-BENCH** — bench **v0.3.5**: new `outcome_prediction` hard-gate task (`pen_stack/validate/outcome_prediction.py`)
+  — the gate is the twin's **honesty properties** (two-sided calibration + OOD widening + immune dimension +
+  phenotype out-of-scope), which an overconfident predictor fails by construction; twin-vs-naive skill is reported
+  informationally on a labelled synthetic stream (no public perturbation-outcome calibration set exists).
+- Docs: `docs/digital_twin.md`; prereg `ws_{vcell,mech,outcome,twincal}` + SHA locks; deposit `phase_5.9/`.
+
+### Notes
+- The twin is a **hypothesis engine, not an oracle of truth**: predictions are candidates with intervals;
+  phenotype, in-vivo behaviour, immunogenicity *magnitude*, and durability beyond the computable stay
+  scope-flagged. The interval is a heuristic band, **not** a trained conformal interval (no public outcome
+  calibration set). Immune-outcome is sourced from v5.6, never invented.
+
 ## [5.8.0] - 2026-06-11 - v5.8 release: The Live Agent & Generative Designer
 
 **Closed-Loop arc, Cycle 2 of 7.** PEN-STACK turns from a *checker* into a grounded *designer*: it generates

{pen_stack-5.8.0 → pen_stack-5.10.0}/CITATION.cff

@@ -1,7 +1,7 @@
 cff-version: 1.2.0
 message: "If you use PEN-STACK, please cite it as below."
 title: "PEN-STACK: open infrastructure for genome writing"
-version: 5.8.0
+version: 5.10.0
 date-released: 2026-06-11
 authors:
   - family-names: "Mahaboob Ali"

{pen_stack-5.8.0 → pen_stack-5.10.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pen-stack
-Version: 5.8.0
+Version: 5.10.0
 Summary: Open infrastructure for genome writing: the Writable Genome atlas, the Writer Atlas, and the Write Planner.
 Author-email: Anees Ahmed Mahaboob Ali <ahmedaneesm@gmail.com>
 License: MIT
@@ -90,12 +90,12 @@ 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-5.8.0-blue.svg)](CHANGELOG.md)
-[![Tests](https://img.shields.io/badge/tests-317%20passing-success.svg)](tests/)
+[![Version](https://img.shields.io/badge/version-5.10.0-blue.svg)](CHANGELOG.md)
+[![Tests](https://img.shields.io/badge/tests-330%20passing-success.svg)](tests/)
 [![Lint: ruff](https://img.shields.io/badge/lint-ruff-purple.svg)](https://github.com/astral-sh/ruff)
 [![Runtime: Docker](https://img.shields.io/badge/runtime-docker-2496ED.svg)](docker/)
 [![Validation: pre-registered](https://img.shields.io/badge/validation-pre--registered-critical.svg)](prereg/)
-[![Genome-Writing Bench v0.3](https://img.shields.io/badge/benchmark-Genome--Writing%20Bench%20v0.3.4-6f42c1.svg)](benchmarks/genome_writing_bench/)
+[![Genome-Writing Bench v0.3](https://img.shields.io/badge/benchmark-Genome--Writing%20Bench%20v0.3.6-6f42c1.svg)](benchmarks/genome_writing_bench/)
 
 **Built on five prior, separately published repositories:**
 
@@ -134,6 +134,43 @@ Two questions gate every genome-writing project, and before PEN-STACK no resourc
 Everything is built on bulk-downloadable public data, runs on a single GPU, and is validated **blind** against
 a pre-registered, honest baseline before release.
 
+## What is new in v5.10 — The Experiment Designer (active learning / EIG)
+
+v5.10 (**Closed-Loop arc, Cycle 4 of 7**) is the **Learn** brain of a self-driving lab: it turns *"I'm
+uncertain"* into *"run **this** experiment next."* It reads the calibrated v5.9 twin's uncertainty and the v5.6
+immune labels, scores each candidate experiment by the information it is expected to yield, assembles a diverse
+batch, and proves on held-out data — with confidence intervals — that this learns faster than random/greedy.
+
+| Workstream | What it adds | Result |
+|---|---|---|
+| **ACQ** | `active/acquire.py` | EIG from the twin (≥0, monotone in uncertainty); **immune-VOI** rewards experiments that would validate a v5.6 **proxy** axis; deterministic, traceable |
+| **DESIGN** | `active/design.py` | `select_batch` — diverse batch (acquisition − redundancy penalty), not k copies of the most-uncertain point; each carries expected info gain |
+| **VALIDATE** | `active/validate.py` | retrospective active vs random/greedy learning curves with reps + **bootstrap CI** on the curve-area gap; beats random **only if CI excludes 0**, else reports not-yet-useful |
+| **BENCH** | bench **v0.3.6** `experiment_design` hard gate | gate = the Learn engine's honesty + falsifiability; a random selector (no acquisition, no falsifiable curve) fails by construction |
+
+Falsifiable by construction and **lab-optional** — it chooses informative experiments but does not run them
+(prospective benefit awaits a lab; no autonomy claim). See [`docs/experiment_design.md`](docs/experiment_design.md)
+and `prereg/ws_{acq,aldesign,alvalidate}.yaml`.
+
+## What is new in v5.9 — The Digital Twin (calibrated outcome prediction)
+
+v5.9 (**Closed-Loop arc, Cycle 3 of 7**) adds the missing layer — *what does the cell do after the write?* —
+predicted with calibrated honesty. The twin computes what mechanism allows, adds an in-distribution virtual-cell
+estimate (OOD-gated), screens immune outcome from the v5.6 profile, and is explicit about its boundary at
+phenotype. A **hypothesis engine, not an oracle of truth**.
+
+| Workstream | What it adds | Result |
+|---|---|---|
+| **VCELL** | `oracles/vcell.py` + scope cards `state`/`scgpt` | Arc STATE / scGPT under the OracleResult contract; perturbation prediction is a **candidate**, **OOD-gated**, deferred value never fabricated |
+| **MECH** | `twin/mechanistic.py` | `cassette_expression` = promoter × copy × accessibility (closed form); **physics where computable, NOT a phenotype** |
+| **OUTCOME** | `twin/outcome.py` | fuses mechanism + in-dist VC + v5.6 immune; interval **widens under OOD**; in-vivo durability conditioned on the **grounded NAb axis**; phenotype/in-vivo-magnitude scope-flagged |
+| **CAL** | `twin/calibrate.py` | calibration reported **two-sided** (coverage + MAE-gap-vs-naive bootstrap CI); beats naive **only if CI excludes 0**, else the negative is reported |
+| **BENCH** | bench **v0.3.5** `outcome_prediction` hard gate | gate = the twin's honesty properties (two-sided cal + OOD widening + immune dim + phenotype out-of-scope); an overconfident predictor fails by construction |
+
+The interval is a **heuristic band, not a trained conformal interval** (no public perturbation-outcome calibration
+set; Arc's Virtual Cell Challenge shows models don't yet consistently beat naive baselines). See
+[`docs/digital_twin.md`](docs/digital_twin.md) and `prereg/ws_{vcell,mech,outcome,twincal}.yaml`.
+
 ## What is new in v5.8 — The Live Agent & Generative Designer
 
 v5.8 (**Closed-Loop arc, Cycle 2 of 7**) turns PEN-STACK from a *checker* into a grounded *designer*: it
@@ -599,11 +636,13 @@ pen-stack/
 │   ├── agent/                        agentic platform: tools / orchestrator / pen_agent / mcp_server / guardrails; v5.0 co_scientist + cite (multi-strategy, self-critique, cited rationale, scope ledger); v5.8 orchestrator_live (live, cache-replayable, generate→oracle→verify→refine)
 │   │                                   + v3.2 epistemic (3-tier status) / scope (known-unknowns matcher)
 │   ├── graph/                        v4.5 living world-model knowledge graph (schema/build/query/ingest/cell_types); typed provenanced edges; gated living loop (propose-only)
-│   ├── oracles/                      v4.0 L1 oracle mesh: OracleResult contract + adapters (genome/structure/protein_design/rna/energetics) over the foundation models; version-pinned cache; v5.2-5.6 delivery-immunology scope cards (delivery_genotoxicity/capsid_epitope/innate_sensing/seroprevalence/antipeg)
+│   ├── oracles/                      v4.0 L1 oracle mesh: OracleResult contract + adapters (genome/structure/protein_design/rna/energetics) over the foundation models; version-pinned cache; v5.2-5.6 delivery-immunology scope cards (delivery_genotoxicity/capsid_epitope/innate_sensing/seroprevalence/antipeg); v5.9 vcell (Arc STATE/scGPT, OOD-gated, output_kind=candidate)
 │   ├── rules/                        v3.3 machine-readable rules engine (schema/evaluators/loader/solver) over configs/rules/*.yaml
 │   ├── verify/                       v3.3 verification service: verify(design) -> Verdict (legal+reasons+confidence+scope; v4.0 writer_critique; v5.1 delivery_profile; v5.6 immune_profile per-axis vector; v5.7 safety SafetyVerdict)
 │   ├── safety/                       v5.7 the Guardian: biosecurity/dual-use gate (registry/screen/policy/gate/audit/redteam); runs first in verify(); refuse short-circuits; tamper-evident audit
 │   ├── design/                       v5.8 generative designer: space (candidate_space) / generate (verifier-as-discriminator; hazardous+illegal discarded) / pareto (frontier w/ grounded v5.6 immune axis)
+│   ├── twin/                         v5.9 digital twin: mechanistic (cassette expression, closed-form) / outcome (fuse mech+vcell+v5.6 immune; OOD widens interval; phenotype-bounded) / calibrate (honest two-sided)
+│   ├── active/                       v5.10 experiment designer: acquire (EIG/immune-VOI over the v5.9 twin) / design (diverse batch) / validate (retrospective active-vs-random, reps+CI, falsifiable)
 │   ├── adapt/                        local recalibration / private-data adaptation behind a gate (v3.1, WS-F)
 │   ├── env/                          v3.4 full Gymnasium environment over router+verifier (genome_writing_env + policies; [env] extra)
 │   ├── monitor/                      PEN-MONITOR living database (Europe PMC)
@@ -615,12 +654,14 @@ pen-stack/
 │   │                                   v3.3 bench_rule_tasks (T12) / v3.4 bench_writetype_tasks + bench_adversarial_tasks (T13-16) + outcome_calibration /
 │   │                                   v5.6 immune_calibration (proxy-vs-observed; labels each axis validated-or-proxy, two-sided) /
 │   │                                   v5.7 safety_screening (the Guardian hard-gate: benign 0-false-refusal · hazards refused/escalated · evasions never clear) /
-│   │                                   v5.8 generative_design (verifier-as-discriminator hard-gate: hazardous+illegal discarded; survivors calibrated+immune; grounded-immune Pareto)
+│   │                                   v5.8 generative_design (verifier-as-discriminator hard-gate: hazardous+illegal discarded; survivors calibrated+immune; grounded-immune Pareto) /
+│   │                                   v5.9 outcome_prediction (digital-twin hard-gate: two-sided calibration + OOD widening + immune dim + phenotype out-of-scope) /
+│   │                                   v5.10 experiment_design (active-learning hard-gate: EIG monotone + immune-VOI + diverse batch + retrospective active-vs-random reps+CI)
 │   ├── data/                         ingestion (genome, chromatin, integration, TRIP, safety annotations)
 │   ├── server/api.py                 FastAPI REST (atlas, crosslink, writable, plan, bridge, ask)
 │   ├── ui/app.py                     Streamlit web app (16 pages; v3.2 PEN-Agent shows confidence + epistemic status)
 │   └── cli.py                        unified CLI
-├── benchmarks/genome_writing_bench/  Genome-Writing Bench v0.3.4 (T1-T16 + co_scientist + safety_screening + generative_design; tasks / harness / solvers / LEADERBOARD / SHAs)
+├── benchmarks/genome_writing_bench/  Genome-Writing Bench v0.3.6 (T1-T16 + co_scientist + safety_screening + generative_design + outcome_prediction + experiment_design; tasks / harness / solvers / LEADERBOARD / SHAs)
 ├── bench/run.py                      one-command bench entrypoint (--agent, --verify)
 ├── scripts/                          reproducible pipeline drivers (p1_*, p2_*, p4_*, p52/p53 delivery-immunology oracle builds, ws_*_report)
 ├── configs/                          pinned datasets + thresholds + curation (YAML); v3.2 known_unknowns /
@@ -629,10 +670,10 @@ pen-stack/
 │                                       seroprevalence / antipeg + oracles/scope_cards (+ v5.6 known_unknowns:
 │                                       cd4_mhcii_help / preexisting_capsid_tcell / complement_carpa);
 │                                       v5.7 safety/{hazard_registry,policy,probes} (Guardian; function/family/taxon-level only)
-├── prereg/                           SHA-locked success criteria (paper1..4 + ws_a..ws_h + v3.2-v5.8 ws_{uq,ep,mc,ba,
+├── prereg/                           SHA-locked success criteria (paper1..4 + ws_a..ws_h + v3.2-v5.9 ws_{uq,ep,mc,ba,
 │                                       r,v,route,env,bench,cal,o,wv,atlas,graph,mon,ct,plan,crit,cite,immune,
 │                                       genotox,epitope,innate,seroprev,peg,calib,profile,screen,policy,redteam,
-│                                       gen,pareto,orch} + SHA256 locks)
+│                                       gen,pareto,orch,vcell,mech,outcome,twincal,acq,aldesign,alvalidate} + SHA256 locks)
 ├── data/curated/                     small committed tables (universe, gene coords, measured bridge profile,
 │                                       v3.2 bridge_offtarget_energetics.json)
 ├── data/llm_bench_cache/             28 cached ungrounded-LLM transcripts (T7, offline/CI replay)

{pen_stack-5.8.0 → pen_stack-5.10.0}/README.md

@@ -15,12 +15,12 @@ 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-5.8.0-blue.svg)](CHANGELOG.md)
-[![Tests](https://img.shields.io/badge/tests-317%20passing-success.svg)](tests/)
+[![Version](https://img.shields.io/badge/version-5.10.0-blue.svg)](CHANGELOG.md)
+[![Tests](https://img.shields.io/badge/tests-330%20passing-success.svg)](tests/)
 [![Lint: ruff](https://img.shields.io/badge/lint-ruff-purple.svg)](https://github.com/astral-sh/ruff)
 [![Runtime: Docker](https://img.shields.io/badge/runtime-docker-2496ED.svg)](docker/)
 [![Validation: pre-registered](https://img.shields.io/badge/validation-pre--registered-critical.svg)](prereg/)
-[![Genome-Writing Bench v0.3](https://img.shields.io/badge/benchmark-Genome--Writing%20Bench%20v0.3.4-6f42c1.svg)](benchmarks/genome_writing_bench/)
+[![Genome-Writing Bench v0.3](https://img.shields.io/badge/benchmark-Genome--Writing%20Bench%20v0.3.6-6f42c1.svg)](benchmarks/genome_writing_bench/)
 
 **Built on five prior, separately published repositories:**
 
@@ -59,6 +59,43 @@ Two questions gate every genome-writing project, and before PEN-STACK no resourc
 Everything is built on bulk-downloadable public data, runs on a single GPU, and is validated **blind** against
 a pre-registered, honest baseline before release.
 
+## What is new in v5.10 — The Experiment Designer (active learning / EIG)
+
+v5.10 (**Closed-Loop arc, Cycle 4 of 7**) is the **Learn** brain of a self-driving lab: it turns *"I'm
+uncertain"* into *"run **this** experiment next."* It reads the calibrated v5.9 twin's uncertainty and the v5.6
+immune labels, scores each candidate experiment by the information it is expected to yield, assembles a diverse
+batch, and proves on held-out data — with confidence intervals — that this learns faster than random/greedy.
+
+| Workstream | What it adds | Result |
+|---|---|---|
+| **ACQ** | `active/acquire.py` | EIG from the twin (≥0, monotone in uncertainty); **immune-VOI** rewards experiments that would validate a v5.6 **proxy** axis; deterministic, traceable |
+| **DESIGN** | `active/design.py` | `select_batch` — diverse batch (acquisition − redundancy penalty), not k copies of the most-uncertain point; each carries expected info gain |
+| **VALIDATE** | `active/validate.py` | retrospective active vs random/greedy learning curves with reps + **bootstrap CI** on the curve-area gap; beats random **only if CI excludes 0**, else reports not-yet-useful |
+| **BENCH** | bench **v0.3.6** `experiment_design` hard gate | gate = the Learn engine's honesty + falsifiability; a random selector (no acquisition, no falsifiable curve) fails by construction |
+
+Falsifiable by construction and **lab-optional** — it chooses informative experiments but does not run them
+(prospective benefit awaits a lab; no autonomy claim). See [`docs/experiment_design.md`](docs/experiment_design.md)
+and `prereg/ws_{acq,aldesign,alvalidate}.yaml`.
+
+## What is new in v5.9 — The Digital Twin (calibrated outcome prediction)
+
+v5.9 (**Closed-Loop arc, Cycle 3 of 7**) adds the missing layer — *what does the cell do after the write?* —
+predicted with calibrated honesty. The twin computes what mechanism allows, adds an in-distribution virtual-cell
+estimate (OOD-gated), screens immune outcome from the v5.6 profile, and is explicit about its boundary at
+phenotype. A **hypothesis engine, not an oracle of truth**.
+
+| Workstream | What it adds | Result |
+|---|---|---|
+| **VCELL** | `oracles/vcell.py` + scope cards `state`/`scgpt` | Arc STATE / scGPT under the OracleResult contract; perturbation prediction is a **candidate**, **OOD-gated**, deferred value never fabricated |
+| **MECH** | `twin/mechanistic.py` | `cassette_expression` = promoter × copy × accessibility (closed form); **physics where computable, NOT a phenotype** |
+| **OUTCOME** | `twin/outcome.py` | fuses mechanism + in-dist VC + v5.6 immune; interval **widens under OOD**; in-vivo durability conditioned on the **grounded NAb axis**; phenotype/in-vivo-magnitude scope-flagged |
+| **CAL** | `twin/calibrate.py` | calibration reported **two-sided** (coverage + MAE-gap-vs-naive bootstrap CI); beats naive **only if CI excludes 0**, else the negative is reported |
+| **BENCH** | bench **v0.3.5** `outcome_prediction` hard gate | gate = the twin's honesty properties (two-sided cal + OOD widening + immune dim + phenotype out-of-scope); an overconfident predictor fails by construction |
+
+The interval is a **heuristic band, not a trained conformal interval** (no public perturbation-outcome calibration
+set; Arc's Virtual Cell Challenge shows models don't yet consistently beat naive baselines). See
+[`docs/digital_twin.md`](docs/digital_twin.md) and `prereg/ws_{vcell,mech,outcome,twincal}.yaml`.
+
 ## What is new in v5.8 — The Live Agent & Generative Designer
 
 v5.8 (**Closed-Loop arc, Cycle 2 of 7**) turns PEN-STACK from a *checker* into a grounded *designer*: it
@@ -524,11 +561,13 @@ pen-stack/
 │   ├── agent/                        agentic platform: tools / orchestrator / pen_agent / mcp_server / guardrails; v5.0 co_scientist + cite (multi-strategy, self-critique, cited rationale, scope ledger); v5.8 orchestrator_live (live, cache-replayable, generate→oracle→verify→refine)
 │   │                                   + v3.2 epistemic (3-tier status) / scope (known-unknowns matcher)
 │   ├── graph/                        v4.5 living world-model knowledge graph (schema/build/query/ingest/cell_types); typed provenanced edges; gated living loop (propose-only)
-│   ├── oracles/                      v4.0 L1 oracle mesh: OracleResult contract + adapters (genome/structure/protein_design/rna/energetics) over the foundation models; version-pinned cache; v5.2-5.6 delivery-immunology scope cards (delivery_genotoxicity/capsid_epitope/innate_sensing/seroprevalence/antipeg)
+│   ├── oracles/                      v4.0 L1 oracle mesh: OracleResult contract + adapters (genome/structure/protein_design/rna/energetics) over the foundation models; version-pinned cache; v5.2-5.6 delivery-immunology scope cards (delivery_genotoxicity/capsid_epitope/innate_sensing/seroprevalence/antipeg); v5.9 vcell (Arc STATE/scGPT, OOD-gated, output_kind=candidate)
 │   ├── rules/                        v3.3 machine-readable rules engine (schema/evaluators/loader/solver) over configs/rules/*.yaml
 │   ├── verify/                       v3.3 verification service: verify(design) -> Verdict (legal+reasons+confidence+scope; v4.0 writer_critique; v5.1 delivery_profile; v5.6 immune_profile per-axis vector; v5.7 safety SafetyVerdict)
 │   ├── safety/                       v5.7 the Guardian: biosecurity/dual-use gate (registry/screen/policy/gate/audit/redteam); runs first in verify(); refuse short-circuits; tamper-evident audit
 │   ├── design/                       v5.8 generative designer: space (candidate_space) / generate (verifier-as-discriminator; hazardous+illegal discarded) / pareto (frontier w/ grounded v5.6 immune axis)
+│   ├── twin/                         v5.9 digital twin: mechanistic (cassette expression, closed-form) / outcome (fuse mech+vcell+v5.6 immune; OOD widens interval; phenotype-bounded) / calibrate (honest two-sided)
+│   ├── active/                       v5.10 experiment designer: acquire (EIG/immune-VOI over the v5.9 twin) / design (diverse batch) / validate (retrospective active-vs-random, reps+CI, falsifiable)
 │   ├── adapt/                        local recalibration / private-data adaptation behind a gate (v3.1, WS-F)
 │   ├── env/                          v3.4 full Gymnasium environment over router+verifier (genome_writing_env + policies; [env] extra)
 │   ├── monitor/                      PEN-MONITOR living database (Europe PMC)
@@ -540,12 +579,14 @@ pen-stack/
 │   │                                   v3.3 bench_rule_tasks (T12) / v3.4 bench_writetype_tasks + bench_adversarial_tasks (T13-16) + outcome_calibration /
 │   │                                   v5.6 immune_calibration (proxy-vs-observed; labels each axis validated-or-proxy, two-sided) /
 │   │                                   v5.7 safety_screening (the Guardian hard-gate: benign 0-false-refusal · hazards refused/escalated · evasions never clear) /
-│   │                                   v5.8 generative_design (verifier-as-discriminator hard-gate: hazardous+illegal discarded; survivors calibrated+immune; grounded-immune Pareto)
+│   │                                   v5.8 generative_design (verifier-as-discriminator hard-gate: hazardous+illegal discarded; survivors calibrated+immune; grounded-immune Pareto) /
+│   │                                   v5.9 outcome_prediction (digital-twin hard-gate: two-sided calibration + OOD widening + immune dim + phenotype out-of-scope) /
+│   │                                   v5.10 experiment_design (active-learning hard-gate: EIG monotone + immune-VOI + diverse batch + retrospective active-vs-random reps+CI)
 │   ├── data/                         ingestion (genome, chromatin, integration, TRIP, safety annotations)
 │   ├── server/api.py                 FastAPI REST (atlas, crosslink, writable, plan, bridge, ask)
 │   ├── ui/app.py                     Streamlit web app (16 pages; v3.2 PEN-Agent shows confidence + epistemic status)
 │   └── cli.py                        unified CLI
-├── benchmarks/genome_writing_bench/  Genome-Writing Bench v0.3.4 (T1-T16 + co_scientist + safety_screening + generative_design; tasks / harness / solvers / LEADERBOARD / SHAs)
+├── benchmarks/genome_writing_bench/  Genome-Writing Bench v0.3.6 (T1-T16 + co_scientist + safety_screening + generative_design + outcome_prediction + experiment_design; tasks / harness / solvers / LEADERBOARD / SHAs)
 ├── bench/run.py                      one-command bench entrypoint (--agent, --verify)
 ├── scripts/                          reproducible pipeline drivers (p1_*, p2_*, p4_*, p52/p53 delivery-immunology oracle builds, ws_*_report)
 ├── configs/                          pinned datasets + thresholds + curation (YAML); v3.2 known_unknowns /
@@ -554,10 +595,10 @@ pen-stack/
 │                                       seroprevalence / antipeg + oracles/scope_cards (+ v5.6 known_unknowns:
 │                                       cd4_mhcii_help / preexisting_capsid_tcell / complement_carpa);
 │                                       v5.7 safety/{hazard_registry,policy,probes} (Guardian; function/family/taxon-level only)
-├── prereg/                           SHA-locked success criteria (paper1..4 + ws_a..ws_h + v3.2-v5.8 ws_{uq,ep,mc,ba,
+├── prereg/                           SHA-locked success criteria (paper1..4 + ws_a..ws_h + v3.2-v5.9 ws_{uq,ep,mc,ba,
 │                                       r,v,route,env,bench,cal,o,wv,atlas,graph,mon,ct,plan,crit,cite,immune,
 │                                       genotox,epitope,innate,seroprev,peg,calib,profile,screen,policy,redteam,
-│                                       gen,pareto,orch} + SHA256 locks)
+│                                       gen,pareto,orch,vcell,mech,outcome,twincal,acq,aldesign,alvalidate} + SHA256 locks)
 ├── data/curated/                     small committed tables (universe, gene coords, measured bridge profile,
 │                                       v3.2 bridge_offtarget_energetics.json)
 ├── data/llm_bench_cache/             28 cached ungrounded-LLM transcripts (T7, offline/CI replay)

{pen_stack-5.8.0 → pen_stack-5.10.0}/benchmarks/genome_writing_bench/SHA256SUMS

@@ -1,4 +1,4 @@
-d5d306cdd81cb6f0c67e88de4741610da1898cdb7a2a097e6acfa9226b0c7024  benchmarks/genome_writing_bench/tasks.yaml
+e00a8a901015eac185ac75366cf1ce5b0783451c9bc862d27d745b7d038887a8  benchmarks/genome_writing_bench/tasks.yaml
 5ca511c6763b4e3703e8009031ff8dc385c2a99540e2731d875a9fbbb16f5ecd  configs/gsh_validated_heldout.yaml
 758817c1e46c7db10f7f942316663367c5f297cac0cf2f59947a90638a256718  data/writer_panel.csv
 865b18ff23d140c3df6f3b5f25398581ebdfe3534e1cecf6f512afb540ab5ede  data/gsh_matched_controls.parquet

{pen_stack-5.8.0 → pen_stack-5.10.0}/benchmarks/genome_writing_bench/tasks.yaml

@@ -8,7 +8,7 @@
 # A task names a `scorer` (module.function in pen_stack.validate / pen_stack.bridge) and a `metric` key to
 # read from its report. Solvers (deterministic planner, naive baseline, LLM agent) are compared on the same
 # tasks; a solver that cannot ground a number must refuse, not invent (no-fabrication is a hard gate).
-version: "0.3.4"
+version: "0.3.6"
 prepared: "2026-06-11"
 
 taxonomy:
@@ -41,6 +41,10 @@ taxonomy:
   SF_safety_screening:  "screen a design for select-agent / pandemic-pathogen / controlled-toxin signatures: benign therapeutics pass (0 false refusals), hazards refused/escalated, AI-homolog + chimera + reframing evasions never `clear` (vs a no-safety baseline that clears everything)"
   # v0.3.4 (v5.8): the generative designer - verifier-as-discriminator over a mixed candidate pool.
   GD_generative_design: "generate candidate writing systems and keep ONLY legal+safe+calibrated+immune-profiled survivors (hazardous toxin payload + illegal oversize/form-incompatible designs DISCARDED), returning a Pareto frontier with a grounded immune-risk axis (vs an ungrounded generator that ships hazardous/illegal designs)"
+  # v0.3.5 (v5.9): the digital twin - calibrated, OOD-aware, immune-aware, phenotype-bounded outcome prediction.
+  OP_outcome_prediction: "predict a write's computable outcome with HONEST calibration (two-sided MAE-vs-naive with bootstrap CI), an interval that WIDENS under OOD, an immune-outcome dimension from v5.6, and phenotype out of scope (vs an overconfident predictor with a fixed narrow interval and no scope awareness)"
+  # v0.3.6 (v5.10): the experiment designer - active learning / EIG, falsifiable vs random/greedy.
+  ED_experiment_design: "choose the experiment(s) that most reduce model uncertainty: EIG from the calibrated twin (monotone in uncertainty), immune-VOI rewarding proxy-validating experiments, a diverse batch, and a RETROSPECTIVE active-vs-random validation with reps + bootstrap CI (vs a random selector with no acquisition signal and no falsifiable curve)"
 
 tasks:
   - id: site_selection_blind_gsh
@@ -261,3 +265,36 @@ tasks:
     note: "v5.8 verifier-as-discriminator: the ungrounded generator returns the raw pool (ships hazardous +
       illegal designs) and fails by construction; the grounded designer discards them and returns calibrated,
       immune-profiled candidates on a Pareto frontier with a grounded immune-risk axis."
+
+  - id: outcome_prediction
+    family: OP_outcome_prediction
+    scorer: "pen_stack.validate.outcome_prediction:run"
+    metric: "twin_honest_and_calibrated"
+    baseline_metric: "overconfident_predictor_honest"
+    higher_is_better: true
+    hard_gate: true
+    gate_rule: "twin_honest_and_calibrated == True (calibration two-sided w/ CI AND OOD widens interval AND immune-outcome present AND phenotype out-of-scope)"
+    ground_truth: "structural honesty properties of the outcome prediction (two-sided calibration + OOD widening +
+      immune dimension from v5.6 + phenotype out-of-scope), NOT a beat-the-world claim - non-circular; twin-vs-naive
+      skill is reported informationally on a labelled synthetic stream because no public perturbation-outcome
+      calibration set exists (Arc Virtual Cell Challenge: models do not yet consistently beat naive baselines)"
+    circular: false
+    note: "v5.9 the digital twin: a trustworthy predictor is calibrated + OOD-aware + immune-aware + phenotype-
+      bounded; an overconfident predictor (fixed narrow interval, no OOD/scope awareness) fails the gate by
+      construction. Honest about what the field cannot yet do."
+
+  - id: experiment_design
+    family: ED_experiment_design
+    scorer: "pen_stack.validate.experiment_design:run"
+    metric: "experiment_designer_honest"
+    baseline_metric: "random_selector_honest"
+    higher_is_better: true
+    hard_gate: true
+    gate_rule: "experiment_designer_honest == True (EIG monotone in uncertainty AND immune-VOI rewards proxy validation AND batch diverse AND retrospective active-vs-random reported with reps+CI)"
+    ground_truth: "structural honesty + falsifiability properties of the Learn engine (twin-sourced EIG, immune-VOI,
+      diverse batch, retrospective active-vs-random with reps + bootstrap CI), NOT a beat-the-world claim -
+      non-circular; the active-beats-random outcome is reported either way (not-yet-useful is a valid result)"
+    circular: false
+    note: "v5.10 the experiment designer: a trustworthy Learn engine computes acquisition from the calibrated twin,
+      rewards proxy-validating experiments, selects a diverse batch, and validates active-vs-random retrospectively
+      with CIs; a random selector has no acquisition signal and no falsifiable curve and fails by construction."

{pen_stack-5.8.0 → pen_stack-5.10.0}/configs/oracles/scope_cards.yaml

@@ -179,3 +179,26 @@ oracles:
     generalizes_to_unseen_loci: false
     license: "open (this work; Chen 10.1021/acs.analchem.6b03109, Yang&Lai 10.1002/wnan.1339,
       Armstrong 10.1002/cncr.22739, Kozma 10.1016/j.addr.2020.07.024)"
+
+  state:                          # v5.9 WS-VCELL: Arc STATE virtual-cell perturbation-response model
+    family: vcell
+    version: "state-2025"
+    output_kind: candidate        # a perturbation-response PREDICTION is a candidate/hypothesis, never a claim
+    valid_for: "IN-DISTRIBUTION single-gene / chemical / cytokine perturbation transcriptional-response estimates
+      in TRAINED human cell contexts (Arc STATE; ~70 contexts, 167M+100M cells)"
+    not_valid_for: "OUT-OF-DISTRIBUTION cell types / perturbations (Arc's own Virtual Cell Challenge result:
+      perturbation models do NOT yet consistently beat naive baselines -> OOD sets extrapolating=True); phenotype;
+      in-vivo behaviour; clinical endpoints"
+    generalizes_to_unseen_loci: false
+    license: "Arc Institute (github.com/ArcInstitute/state)"
+
+  scgpt:                          # v5.9 WS-VCELL: scGPT single-cell foundation model (perturbation prediction)
+    family: vcell
+    version: "scgpt-2024"
+    output_kind: candidate
+    valid_for: "IN-DISTRIBUTION genetic-perturbation transcriptional-response estimates within trained single-cell
+      contexts (scGPT; >33M cells)"
+    not_valid_for: "OOD cell types / perturbations (does not consistently beat naive baselines); phenotype; in-vivo;
+      absolute expression"
+    generalizes_to_unseen_loci: false
+    license: "MIT (github.com/bowang-lab/scGPT; Nature Methods 10.1038/s41592-024-02201-0)"

pen_stack-5.10.0/docs/digital_twin.md

@@ -0,0 +1,55 @@
+# The digital twin (v5.9)
+
+From v5.9, PEN-STACK can predict *what the cell does after the write* — calibrated, OOD-gated, and scope-bounded.
+It computes what mechanism allows, adds an in-distribution virtual-cell estimate, screens immune outcome from the
+v5.6 profile, and is honest about its boundary at phenotype. The twin is a **hypothesis engine, not an oracle of
+truth**.
+
+```python
+from pen_stack.twin import predict_outcome
+o = predict_outcome(design, cell_state="k562")
+o["predicted_outcome"]   # {relative_expression, vcell_response, units}
+o["interval"]            # heuristic band; WIDENS under OOD
+o["immune_outcome"]      # the v5.6 per-axis profile (sourced, not invented)
+o["scope_flags"]         # phenotype_not_modeled, in_vivo_magnitude_unknown, (vcell_OOD if extrapolating)
+```
+
+## Mechanism where computable (`pen_stack/twin/mechanistic.py`)
+
+`cassette_expression(design, chromatin_ctx)` = `promoter_strength × copy_number × accessibility` — a closed-form
+steady-state estimate. Assumptions (steady-state, no silencing, linear copy scaling) and scope flags
+(`episomal_durability_unknown`, `phenotype_not_modeled`) travel with the output. It is **physics where
+computable, never a phenotype**.
+
+## Virtual-cell oracle, OOD-gated (`pen_stack/oracles/vcell.py`)
+
+`predict_response(cell_state, perturbation, model="state")` wraps Arc **STATE** / **scGPT** under the v4.0
+`OracleResult` contract. A perturbation-response prediction is a **candidate**, never a claim; a cell context or
+perturbation outside the documented validity envelope sets `extrapolating=True` / `in_scope=False`. The backend is
+deferred/cache-replayed (value `None` when absent — never fabricated). This encodes the field's own result (Arc's
+Virtual Cell Challenge): **perturbation models do not yet consistently beat naive baselines** and do not
+generalize to unseen contexts.
+
+## Fused outcome (`pen_stack/twin/outcome.py`)
+
+`predict_outcome(design, cell_state)` fuses the computable mechanistic estimate (backbone) + an in-distribution
+virtual-cell response (when available) + the v5.6 immune profile. The interval **widens under OOD** rather than
+over-trusting an extrapolating model. For in-vivo vehicles, durability **may be conditioned on the grounded
+pre-existing-NAb axis** (no invented immune numbers). Phenotype and in-vivo magnitude stay scope-flagged.
+
+> The interval is a heuristic band, **not** a trained conformal interval — there is no public
+> perturbation-outcome calibration set. The twin says so.
+
+## Honest calibration (`pen_stack/twin/calibrate.py`)
+
+`calibrate_outcome(predictions, observations, intervals=…)` reports calibration **two-sided, whatever the shape**:
+interval coverage vs nominal, and a skill comparison against a naive mean baseline with a **bootstrap CI on the
+MAE gap**. The twin "beats" the baseline **only when the CI excludes zero** — otherwise the negative is reported
+verbatim. At `N < 3` it abstains honestly.
+
+## Honest scope
+
+The twin predicts what mechanism computes, what an in-distribution virtual-cell model supports, and what the v5.6
+immune profile screens — with calibrated intervals. It does **not** predict phenotype, in-vivo behaviour,
+immunogenicity *magnitude*, or durability beyond the computable; these stay scope-flagged. Perturbation prediction
+is an open problem; the twin is a calibrated hypothesis engine, honest where it is weak.

pen_stack-5.10.0/docs/experiment_design.md

@@ -0,0 +1,44 @@
+# The experiment designer (v5.10)
+
+From v5.10, PEN-STACK turns *"I'm uncertain"* into *"run **this** experiment next."* It reads the calibrated v5.9
+twin's uncertainty and the v5.6 immune-risk labels, scores each candidate experiment by the information it is
+expected to yield, assembles a diverse batch, and proves on held-out data — with confidence intervals — that this
+learns faster than random or greedy, reporting honestly when it does not. The **Learn** brain of a self-driving
+lab: lab-optional and falsifiable by construction.
+
+## Acquisition (`pen_stack/active/acquire.py`)
+
+```python
+from pen_stack.active import acquisition_score, expected_information_gain, immune_voi
+```
+
+- **`predictive_entropy(outcome)`** — the twin's current uncertainty, from its interval width (Gaussian
+  differential entropy).
+- **`expected_information_gain(candidate, cell_state)`** — reducible uncertainty: `entropy now − expected
+  posterior entropy` (a measurement collapses the predictive sd toward a noise floor); `≥ 0`. Monotone in the
+  twin's uncertainty (an OOD candidate yields more EIG).
+- **`immune_voi(candidate)`** — value of information for **validating an immune PROXY axis** (v5.6): an axis still
+  labelled a proxy that this experiment would measure is high-VOI (it would turn proxy → outcome-validated).
+- **`acquisition_score`** = `w_eig·EIG + w_unc·entropy + w_imm·immune_voi`. Fully traceable to twin quantities +
+  v5.6 labels; deterministic given inputs (no fabricated values).
+
+## Diverse batch (`pen_stack/active/design.py`)
+
+`select_batch(candidates, cell_state, k)` greedily maximises summed acquisition **minus a redundancy penalty**
+(shared design facets) against the already-chosen set — so a batch is a *diverse* set of informative experiments,
+not k copies of the single most-uncertain point. Each chosen experiment carries its `expected_info_gain`.
+
+## Retrospective falsifiability (`pen_stack/active/validate.py`)
+
+`retrospective_active_learning(dataset, strategies=("active","random","greedy"))` simulates campaigns per
+strategy on a held-out split, records the held-out-MAE learning curve per round, and over repetitions reports
+mean±CI curves and a **bootstrap CI on the curve-area gap** (`random_area − active_area`). The active learner
+"beats" random **only when the CI excludes zero**; otherwise the not-yet-useful negative is reported verbatim — a
+valid, published outcome.
+
+## Honest scope
+
+The experiment designer is only as good as the v5.9 twin and the v5.6 labels it queries. Its advantage is
+validated **retrospectively** on existing data with confidence intervals, and reported honestly when absent. It
+chooses informative experiments — including ones that would validate an immune proxy — but it **does not run
+them**; prospective benefit awaits a lab partner (v5.11+).

{pen_stack-5.8.0 → pen_stack-5.10.0}/pen_stack/__init__.py

@@ -1,2 +1,2 @@
 """PEN-STACK v3.0 - open infrastructure for genome writing."""
-__version__ = "5.8.0"
+__version__ = "5.10.0"

pen_stack-5.10.0/pen_stack/active/__init__.py

@@ -0,0 +1,20 @@
+"""pen_stack.active — the experiment designer / the "Learn" brain of a self-driving lab (v5.10).
+
+Turn "I'm uncertain" into "run THIS experiment next": score each candidate experiment by the information it is
+expected to yield (from the calibrated v5.9 twin), reward experiments that would validate an immune PROXY axis
+(v5.6), assemble a diverse batch, and prove on held-out data — with confidence intervals — that this learns
+faster than random or greedy, reporting honestly when it does not. Lab-optional, falsifiable by construction.
+"""
+from __future__ import annotations
+
+from pen_stack.active.acquire import (
+    acquisition_score,
+    expected_information_gain,
+    immune_voi,
+    predictive_entropy,
+)
+from pen_stack.active.design import batch_diversity, select_batch
+from pen_stack.active.validate import retrospective_active_learning
+
+__all__ = ["expected_information_gain", "immune_voi", "predictive_entropy", "acquisition_score",
+           "select_batch", "batch_diversity", "retrospective_active_learning"]

pen_stack-5.10.0/pen_stack/active/acquire.py

@@ -0,0 +1,73 @@
+"""Acquisition functions for the experiment designer (v5.10, WS-ACQ).
+
+Score each candidate experiment by the information it is expected to yield — computed from the calibrated v5.9
+twin's predictive uncertainty (never fabricated). Three signals:
+  * expected_information_gain — reducible predictive uncertainty (entropy now - expected posterior entropy),
+  * predictive_entropy        — the twin's current uncertainty (from its interval width),
+  * immune_voi                — value of information for VALIDATING an immune PROXY axis (turns proxy -> validated).
+The acquisition is only as good as the v5.9 twin and the v5.6 labels it queries; it chooses informative
+experiments, it does not run them.
+"""
+from __future__ import annotations
+
+import math
+
+# a measurement does not resolve uncertainty perfectly: a noise floor on the post-experiment entropy.
+_MEASUREMENT_NOISE_SD = 0.05
+_TWO_PI_E = 2.0 * math.pi * math.e
+
+
+def _interval_sd(outcome: dict) -> float:
+    """Std-dev implied by the twin's (approx 95%) interval: sd ~ width / (2 * 1.96)."""
+    lo, hi = outcome.get("interval", [0.0, 0.0])
+    return max(1e-6, (float(hi) - float(lo)) / (2.0 * 1.96))
+
+
+def _gaussian_entropy(sd: float) -> float:
+    return 0.5 * math.log(_TWO_PI_E * sd * sd)
+
+
+def predictive_entropy(outcome: dict) -> float:
+    """Differential entropy of the twin's predictive distribution, from its interval width."""
+    return _gaussian_entropy(_interval_sd(outcome))
+
+
+def _expected_posterior_entropy(outcome: dict) -> float:
+    """Entropy expected AFTER running the experiment: the measurement collapses predictive sd toward the
+    measurement noise floor (cannot go below it)."""
+    post_sd = max(_MEASUREMENT_NOISE_SD, min(_interval_sd(outcome), _MEASUREMENT_NOISE_SD * 2))
+    return _gaussian_entropy(post_sd)
+
+
+def expected_information_gain(candidate: dict, cell_state: str, model_ctx: dict | None = None) -> float:
+    """EIG ~ reducible uncertainty = predictive entropy now - expected posterior entropy. Computed from the
+    calibrated twin's predictive distribution; >= 0 (a measurement never increases expected uncertainty)."""
+    from pen_stack.twin.outcome import predict_outcome
+    o = predict_outcome(candidate, cell_state or candidate.get("cell_state", ""))
+    return max(0.0, predictive_entropy(o) - _expected_posterior_entropy(o))
+
+
+def immune_voi(candidate: dict, cell_state: str = "") -> float:
+    """Value of information for validating an immune PROXY axis (v5.6): an axis still labelled a proxy that this
+    experiment would MEASURE is high-VOI (turns proxy -> outcome-validated). Reads the v5.6 validation labels."""
+    from pen_stack.twin.outcome import predict_outcome
+    prof = predict_outcome(candidate, cell_state or candidate.get("cell_state", "")).get("immune_outcome") or {}
+    measures = {str(a).strip().lower() for a in (candidate.get("measures_immune_axes") or [])}
+    voi = 0.0
+    for axis, rec in prof.get("axes", {}).items():
+        label = (rec.get("validation") or "").lower()
+        is_proxy = "proxy" in label and "not outcome-validated" in label
+        if is_proxy and (not measures or axis.lower() in measures):
+            voi += 1.0
+    return voi
+
+
+def acquisition_score(candidate: dict, cell_state: str, model_ctx: dict | None = None,
+                      *, w_eig: float = 1.0, w_unc: float = 0.3, w_imm: float = 0.4) -> float:
+    """Weighted acquisition: information gain + raw uncertainty + immune value-of-information. Fully traceable
+    to twin quantities + v5.6 labels (no fabricated values); deterministic given the inputs."""
+    from pen_stack.twin.outcome import predict_outcome
+    o = predict_outcome(candidate, cell_state or candidate.get("cell_state", ""))
+    eig = expected_information_gain(candidate, cell_state, model_ctx)
+    unc = predictive_entropy(o)
+    return w_eig * eig + w_unc * unc + w_imm * immune_voi(candidate, cell_state)