PEN-COMPARE Model Card

Framework Identity

Name: PEN-COMPARE TrueWriterScore v3.2
Version: 0.1.0
Date: 2026-05-26
Authors: Anees Ahmed Mahaboob Ali (ahmedaneesm@gmail.com)
License: MIT
Pre-registration: https://osf.io/4kdvy

Purpose

PEN-COMPARE certifies genome editors as TRUE_WRITER, PROBABLE_WRITER, EMERGING_WRITER, or NOT_WRITER using a 5-gate hierarchical scoring system. The framework operationalizes the Molecular Pen hypothesis: that IS110 bridge recombinases are non-destructive precision writers that avoid double-strand DNA breaks.

Inputs

Field

Type

Description

s_dsb

float [0,1]

DSB-avoidance axis score from pen-score

s_prog

float [0,1]

Programmability axis score

s_cargo

float [0,1]

Native cargo capability axis score

length_aa

int or None

Protein length in amino acids

evidence_sources

list[str]

Subset of {biochemical, structural, computational, cell_based}

intrinsic_cargo_mechanism

bool

Whether native (non-template) cargo delivery is confirmed

split_aav_eligible

bool

Whether split-AAV packaging has been demonstrated

Outputs

Field

Description

tier

TRUE_WRITER / PROBABLE_WRITER / EMERGING_WRITER / NOT_WRITER

qualifying_gates_passed

0-4

has_cell_based_evidence

bool

auto_demoted

bool (True only when G1 fails)

gate_results

Per-gate pass/fail with observed value and threshold

Tier Ladder

G1 FAIL  ->  NOT_WRITER (auto-demote, final)

G1 PASS  +  4 qualifying  +  cell_based  ->  TRUE_WRITER
G1 PASS  +  4 qualifying  +  no cell_based  ->  PROBABLE_WRITER
G1 PASS  +  3 qualifying  +  cell_based  ->  PROBABLE_WRITER
G1 PASS  +  1-2 qualifying  ->  EMERGING_WRITER
G1 PASS  +  0 qualifying  ->  NOT_WRITER

Training Data & Calibration

The gates use no statistical training - all thresholds are mechanistically motivated and pre-registered. Calibration anchors:

Editor

Expected Tier

Verified

ISCro4 (D2TGM5)

TRUE_WRITER

IS621 (A0A2X3M8B0)

PROBABLE_WRITER

Bxb1

PROBABLE_WRITER (G2 fails)

SpCas9 (Q99ZW2)

NOT_WRITER (G1 auto-demote)

Known Limitations

  1. S_Mature > 1.0 pydantic bug: pen-score clamps AxisScores before validation for SpCas9/PE2/etc.

  2. length_aa = None: 13 natural editors have no sequence length in pen-score; G4 falls back to split_aav heuristic.

  3. IS621 PenScore discrepancy: pre-registered predictions_v3.yaml cites 0.929; actual value is 0.9473. Does not affect P2 outcome; must be disclosed.

  4. Universe scope: limited to editors scoreable by pen-score v0.1.3 and pen-assemble v0.5.2 catalog.

Sensitivity Analysis

20,480-combination parameter sweep (16x16x16x5 grid):

  • ISCro4 robustness = 1.000 (stable under all threshold variations)

  • Zero boundary cases (robustness < 0.50)

  • 4 near-boundary editors (50-80%): Bxb1, eePASSIGE, eePASSIGE_v2, phiC31

Pre-Registration Outcomes

All 4 predictions PASS (evaluated 2026-05-26):

ID

Statement

Result

P1

ISCro4 sole TRUE_WRITER

PASS

P2

Zero designs TRUE_WRITER

PASS

P3

>=5 triangulation discrepancies

PASS (30)

P4

RAG LLM >=80% accuracy

PASS (88%)

Citation

@software{pen_compare_2026,
  author = {Mahaboob Ali, Anees Ahmed},
  title = {PEN-COMPARE: Hierarchical Certification Framework for Non-Destructive Genome Editors},
  version = {0.1.0},
  year = {2026},
  url = {https://github.com/ahmedanees-m/pen-compare},
  note = {Pre-registration: https://osf.io/4kdvy}
}