ClarusC64/oncology-signal-latency-boundary-v0.5

What this dataset does

This dataset tests whether a model can detect timing failure in a synthetic tissue ecology.

The task is not cancer diagnosis.

The task is to classify whether a tissue-state scenario can detect and act before the repair opportunity closes.

Core Stability Idea

A tissue may detect the correct signal, interpret it correctly, and coordinate a response, but still fail because action arrives too late.

This dataset tests the timing layer of biological sensing.

The positive class does not indicate cancer.

It indicates that the tissue ecology has crossed into latency-driven instability risk.

Prediction Target

label = 1

The tissue-state scenario has crossed into latency-driven instability risk.

label = 0

The tissue-state scenario remains inside the repair window and can preserve self-correction.

Row Structure

Each row represents a synthetic tissue-state scenario.

Columns:

• scenario_id
• damage_burden
• signal_detection_latency
• repair_activation_latency
• immune_activation_latency
• metabolic_adaptation_latency
• repair_window_width
• signal_decay_rate
• compensatory_capacity_score
• case_type
• label

Case Types

fast_detection

Signals arrive early and repair can begin before the repair window narrows.

These rows represent stable timing dynamics.

delayed_detection

Signal detection is delayed, but the repair window is still wide enough for recovery in some scenarios.

These rows prevent detection latency from becoming a deterministic failure marker.

delayed_response

The signal is detected, but repair, immune, or metabolic action begins too late.

These rows test the difference between sensing and acting.

narrow_repair_window

The repair opportunity is unusually short.

Even moderate delay can become dangerous.

This case type includes paired examples where timing, repair window width, and signal decay remain identical while compensatory capacity changes.

This makes the intended interaction visible:

A narrow window with fast decay can remain stable if compensation is strong enough.

The same timing geometry can become unstable when compensation is insufficient.

missed_window

Repair activation occurs after the useful repair window has effectively closed.

These rows represent strong latency-driven failure.

The missed_window case type is a core contribution of this dataset.

It separates signal detection from repair activation.

In these rows, the system may detect the signal in time but still fail because repair activation occurs after the useful repair window has closed.

compensation_rescue

Timing is poor, but strong compensatory capacity and a wide repair window preserve stability.

These rows break the shortcut that high latency always implies failure.

latency_recovery

The system begins delayed but catches up before collapse.

These rows represent timing recovery near the boundary.

latency_collapse

Moderate delays accumulate across channels.

No single delay is catastrophic, but the combined timing geometry drives instability.

Anti-Shortcut Design

This dataset is designed to prevent simple threshold solutions.

High latency does not always imply failure.

Low latency does not always imply stability.

A wide repair window can rescue delayed response.

A narrow repair window can turn moderate delay into failure.

High compensatory capacity can rescue narrow-window timing pressure.

High compensatory capacity does not always rescue a missed window.

Moderate damage burden can still fail when timing closes the repair opportunity.

High damage burden can remain stable when timing, decay, and compensation remain favorable.

The model must learn the interaction between latency, repair window width, signal decay, burden, and compensation.

Evaluation

Submit predictions in the format:

scenario_id,prediction

Run:

python scorer.py predictions.csv data/test.csv

The scorer returns:

• accuracy
• precision
• recall
• f1
• confusion_matrix
• accuracy_
• count_
• case_type_accuracy_macro

Scorer Design

The scorer automatically discovers case types from the test set.

It does not rely on a hardcoded list of pathway categories.

Any case type present in the evaluation set automatically receives:

• accuracy_
• count_

This allows the benchmark to evolve without requiring scorer modifications.

Macro Case-Type Accuracy

case_type_accuracy_macro is the unweighted mean of accuracy across case types.

Each pathway receives equal weight regardless of how many rows belong to that pathway.

This prevents dominant case types from masking poor performance on rarer but theoretically important timing pathways.

Structural Contribution

Most oncology datasets attempt to predict disease presence.

This dataset attempts to predict whether biological self-correction still has time to act.

The benchmark represents multiple timing pathways:

1. Fast detection
2. Delayed detection with recovery
3. Delayed response despite detection
4. Narrow repair window
5. Missed repair window
6. Compensation rescue
7. Latency recovery
8. Latency collapse through accumulated delay

A successful model cannot rely on latency, burden, repair window width, decay, or compensation alone.

It must infer whether the tissue ecology can respond before the repair opportunity closes.

Structural Note

This dataset is synthetic.

It is designed to evaluate structural reasoning rather than provide medical diagnosis.

The generator logic is intentionally withheld.

The purpose is to test whether a model can infer latency-boundary risk from interacting timing constraints rather than relying on simple correlations.

License

MIT