Helix StudioGenome IDE
← Docs
Helix CLI docs
Browse Helix CLI docs

Outcome Digital Twin (Outcome Space + Repair Mixtures)

Helix is distribution-first: it models the outcome space (a discrete set of named terminal edits) and predicts a probability distribution over that space, rather than implicitly assuming a single “winning” edit.

That choice is the foundation for auditability, reproducibility, and downstream optimization.

What’s Unusual (Best-in-Class Signals)

1) Explicit outcome enumeration (named outcomes)

Outcomes are referential objects (e.g. del3@-2, del5@0), not anonymous histogram bins. That makes them:

  • Discrete (finite outcome set)
  • Indexable (stable keys for caches / exports)
  • Diffable (regression tests can compare distributions by outcome name)
  • Auditable (exports are deterministic and attributable)

2) Escape states are first-class

Helix represents “no cut / no edit” as probability mass, rather than treating repair as guaranteed. This is required for any credible digital-twin narrative (failure modes must be explicit).

3) Distribution-level reasoning

Helix computes and exposes distribution-level properties (dominance, entropy, tail risk), because the right question is not “what wins?” but “what is the full posterior over outcomes?”

The 4.5 → 5.0 Upgrade: Representation, Not New Math

Reviewers want to see mechanistic structure emerge from probabilities, not sit next to them as labels.

The minimal representation lift is a repair-mixture view:

Mixture identity: P(outcome) = Σ_pathway P(pathway) × P(outcome | pathway)

This does not change the underlying simulation math; it makes the mechanistic story inspectable.

Repair pathway objects (derived)

Helix derives an explicit pathway decomposition from the outcome distribution and stable heuristics, exposed as objects in:

  • src/helix/studio/repair_pathways.py
    • RepairPathway(pathway, probability_mass, outcome_distribution)
    • build_repair_pathway_mixture(...)
    • build_stochastic_narrative_v1(...)

This is intentionally conservative: unless the simulator explicitly emits pathway labels, Helix treats the pathway assignment as a representation layer derived from outcome labels/categories.

Stochastic narrative (without sampling)

Even if the implementation is deterministic, the conceptual model is stochastic and is now surfaced explicitly:

  1. Cleavage occurs with probability Pc (vs P(no_cut)).
  2. Repair pathway choice is a categorical distribution conditional on cleavage.
  3. Outcome selection is the conditional distribution P(outcome | pathway).

This is what reviewers mean by a “digital twin”: a visible branching model with explicit mixture weights.

Where You See It

Studio UI (Outcome Explorer)

  • Outcome rows/bars carry a pathway tag.
  • A “Repair pathway mixture (derived)” readout is shown in Analysis/Debug detail levels.
  • The Outcome Explorer filter includes a pathway selector to view distributions per pathway.

Exports

  • Outcome CSV/HTML exports include a pathway column.
  • Session report HTML renders a “Repair pathway mixture (derived)” section, including the mixture identity and a compact stochastic narrative when decision-grade probabilities are available.

Future-Proofing (No Rewrites Required)

When the simulator begins emitting explicit pathway objects, Helix can replace the heuristic classifier with ground-truth pathway provenance while keeping:

  • the same mixture interface,
  • the same export schema shape,
  • the same UI controls,
  • and the same deterministic audit story.