ClarusC64/reasoning-drift-onset-detection-v0.2

A SIOS structured reasoning-state benchmark for detecting when a reasoning trajectory loses a governing constraint, identifying the structural form of that drift, and assessing whether the failure is repaired.

Repository:

ClarusC64/reasoning-drift-onset-detection-v0.2

Version:

0.2.0

Publisher:

Clarus Invariant

Framework:

SIOS

Benchmark identity

Reasoning Drift Onset Detection v0.2 is not a single-label classification benchmark.

It is a structured reasoning-state benchmark.

The task is to reconstruct the state transition of a reasoning trajectory across five primary dimensions:

outcome_mode
first_drift_step
mechanism_family
repair_outcome
residual_constraint

The benchmark asks:

What state did the reasoning trajectory enter, where did the transition begin, what structural failure family governed it, how much repair occurred, and which governing constraint remained unresolved?

The benchmark therefore evaluates both component-level capability and complete-state reconstruction.

Relationship to v0.1

Version 0.1 established the initial drift-onset task but contained a strong positional regularity.

Drift often began at Step 4, and visible failure often appeared at Step 5.

That structure created a positional shortcut.

Version 0.2 removes the fixed-position design and adds:

• Variable trajectory lengths
• Variable onset positions
• Stable no-drift controls
• Latent drift
• Visible failure
• Full repair
• Partial repair
• Failed repair
• Unrepaired drift
• Self-correction
• External correction
• Mixed correction
• Residual constraint annotation
• Coarse and fine mechanism labels
• Gradual constraint loss
• Branching trajectories
• Cosmetic correction
• Oscillating repair
• Blind-input validation
• Two official leaderboards

Version 0.2 supersedes v0.1 for evaluation.

---

Dataset structure

Repository layout

reasoning-drift-onset-detection-v0.2/
├── data/
│   ├── train.csv
│   ├── dev.csv
│   ├── test.csv
│   └── test_blind.csv
├── baseline/
│   ├── baseline.py
│   └── example_predictions.csv
├── scorer/
│   ├── score.py
│   └── validate.py
├── README.md
├── schema.json
├── manifest.json
├── CITATION.cff
├── requirements.txt
└── LICENSE

Split sizes

Split	Records	Labels included
train.csv	16	Yes
dev.csv	8	Yes
test.csv	12	Yes
test_blind.csv	12	No

Total labelled records:

36

The labelled test split and blind test split contain the same scenario identifiers.

The blind file contains only model inputs.

---

Model inputs

The model receives:

scenario_id
domain
task
constraints_json
trajectory_json

scenario_id

A unique scenario identifier.

Example:

RDOD2-025

domain

The reasoning domain represented by the scenario.

Examples include:

clinical_reasoning
scientific_reasoning
legal_reasoning
engineering
machine_learning
policy_analysis
ecology
operations
cybersecurity
historical_analysis

Domain labels support analysis but do not imply that the current dataset is large enough to estimate domain-specific competence reliably.

task

A natural-language description of the reasoning problem.

constraints_json

A JSON array containing the governing reasoning constraints.

Example:

[
  "Preserve multiple plausible causes until discriminating evidence appears.",
  "Separate temporal association from established causation.",
  "Do not treat one recurring factor as sufficient explanation."
]

These constraints define the structure the trajectory is expected to preserve.

trajectory_json

A JSON array of ordered reasoning steps.

Example:

[
  {
    "step": 1,
    "text": "Headaches often occur after poor sleep."
  },
  {
    "step": 2,
    "text": "They also occur after missed meals and during periods of high stress."
  },
  {
    "step": 3,
    "text": "Poor sleep is selected as the primary cause because it appears most consistently in the history."
  }
]

Trajectory steps must:

• Begin at 1
• Use contiguous numbering
• Contain non-empty text
• Preserve their original order

---

Primary prediction targets

The benchmark defines five primary targets.

1. outcome_mode

The final structural state of the trajectory.

Valid values:

no_drift
latent_drift
recovered_drift
visible_failure

no_drift

The trajectory preserves its governing constraints.

A hypothesis may become more plausible without drift occurring, provided alternatives and evidential limits remain active.

latent_drift

A governing constraint is lost, but the trajectory does not yet produce an explicit terminal error, decision, or action.

The failure exists in the reasoning structure even though no visible final failure appears.

recovered_drift

Drift occurs and is subsequently fully or partially repaired.

visible_failure

Drift propagates into an explicit unsupported conclusion, recommendation, classification, decision, or action.

2. first_drift_step

The earliest step at which the reasoning trajectory no longer preserves its governing constraints.

For no_drift, the value is blank.

The onset should identify the first structural transition, not merely the later step where the error becomes obvious.

3. mechanism_family

The coarse structural family governing the drift.

Valid values:

none
premature_constraint_closure
scope_collapse
causal_scope_collapse
multi_causal_collapse
multi_objective_collapse
proxy_substitution
evidence_suppression
evaluation_scope_collapse

none is used only for no_drift.

premature_constraint_closure

The trajectory closes uncertainty, a differential, or a hypothesis set before sufficient discriminating evidence is available.

scope_collapse

A claim expands beyond the scope supported by the evidence or procedure.

causal_scope_collapse

An association, temporal sequence, adjusted result, or intervention is treated as supporting more causal certainty than the evidence permits.

multi_causal_collapse

An interacting causal structure is reduced to one dominant cause or event.

multi_objective_collapse

Several active decision objectives are retained descriptively but one objective captures the operative decision process.

proxy_substitution

A proxy, surface measure, visible artifact, or partial indicator replaces the broader state it was meant to represent.

evidence_suppression

Missing, weak, conflicting, or unresolved evidence is converted into support for closure or is removed from active reasoning.

evaluation_scope_collapse

Performance under a limited evaluation setting is treated as evidence of broader generalisation, robustness, safety, or readiness.

4. repair_outcome

The degree to which the drift was structurally repaired.

Valid values:

not_applicable
full_repair
partial_repair
failed_repair
unrepaired

not_applicable

No drift occurred, so repair was unnecessary.

full_repair

The lost governing constraint is restored and no residual failure remains.

partial_repair

Some correction occurs, but the governing constraint remains incompletely restored.

failed_repair

A correction is attempted but does not repair the relevant structural failure.

unrepaired

No meaningful repair occurs.

5. Residual constraint

The residual constraint is the governing constraint that remains unresolved at the end of the trajectory.

In labelled files, it is stored as:

residual_constraint_failure

The value must exactly match one item in constraints_json, unless no residual failure remains.

For no_drift and full_repair, the value is:

none

In prediction files, the model does not reproduce the full constraint text.

It selects the unresolved constraint using:

predicted_residual_constraint_id

This is a 1-based index into constraints_json.

Example:

[
  "Preserve multiple plausible causes until discriminating evidence appears.",
  "Separate temporal association from established causation.",
  "Do not treat one recurring factor as sufficient explanation."
]

A prediction of:

1

selects:

Preserve multiple plausible causes until discriminating evidence appears.

For no residual constraint, use:

none

or leave the field blank.

---

Secondary annotations

The labelled dataset contains several secondary annotations.

These support diagnosis and analysis but are not all official prediction targets in v0.2.

drift_mechanism

A fine-grained diagnostic subtype within the broader mechanism family.

Example:

{
  "mechanism_family": "causal_scope_collapse",
  "drift_mechanism": "adjustment_to_causation"
}

The mechanism family is the official prediction target.

The fine-grained mechanism remains exploratory because many subtypes currently have sparse support.

correction_source

Identifies where the corrective pressure came from.

Valid values:

self_correction
external_intervention
mixed
not_applicable

self_correction

The reasoning process detects and repairs its own drift.

external_intervention

A reviewer, evaluator, new measurement, external observation, or other outside input introduces the correction.

mixed

Correction develops through both internal revision and external intervention.

not_applicable

No correction occurred or no drift was present.

lost_constraint

The first governing constraint lost when drift begins.

For drift cases, it must exactly match one item in constraints_json.

For no-drift cases:

none

visible_failure_step

The step at which an explicit unsupported conclusion, action, or decision first appears.

This is distinct from first_drift_step.

A trajectory may begin drifting several steps before the failure becomes visible.

Only visible_failure cases define this field.

severity

A coarse estimate of the consequence of the drift within the designed scenario.

Valid values:

none
low
medium
high

Severity is a secondary annotation and should not be interpreted as a universally calibrated risk scale.

---

Repair guidance

recommended_repair

The name of the structural operation required to restore coherence.

Examples include:

evidence_reset
scope_restoration
causal_reset
objective_reintegration
branch_reopening
timeline_reconstruction
system_scope_restoration
evaluation_scope_restoration
readiness_decomposition

repair_instruction

A natural-language description of the required repair.

Example:

Restore the differential and obtain discriminating evidence before confirmation.

Repair guidance is explanatory metadata.

It is not part of either official leaderboard in v0.2.

---

Audit metadata

The labelled files include:

trajectory_length
relative_onset_position
trajectory_structure
annotation_confidence

These fields exist for benchmark auditing and stratified analysis.

They are not model inputs.

trajectory_length

The number of steps in trajectory_json.

It supports analysis by short, medium, and long trajectory length.

relative_onset_position

Calculated as:

first_drift_step / trajectory_length

Example:

first_drift_step = 4
trajectory_length = 8
relative_onset_position = 0.5

This field supports analysis of early, middle, and late drift.

It must never appear in blind model input because it is derived from the gold onset.

trajectory_structure

A descriptive label for the designed trajectory geometry.

Examples include:

stable_multi_source_integration
gradual_objective_weight_drift
branch_suppression_without_terminal_failure
external_correction_with_residual_overclaim
mixed_repair_with_oscillation

This is audit metadata, not a prediction target.

annotation_confidence

The dataset author’s confidence in the annotation.

This field does not replace:

• Independent annotation
• Expert adjudication
• Inter-annotator agreement

---

Prediction format

Prediction files must contain:

scenario_id,predicted_outcome_mode,predicted_first_drift_step,predicted_mechanism_family,predicted_repair_outcome,predicted_residual_constraint_id,confidence

confidence is optional.

A prediction with confidence:

scenario_id,predicted_outcome_mode,predicted_first_drift_step,predicted_mechanism_family,predicted_repair_outcome,predicted_residual_constraint_id,confidence
RDOD2-025,visible_failure,3,premature_constraint_closure,unrepaired,1,0.91

A no-drift prediction without confidence:

scenario_id,predicted_outcome_mode,predicted_first_drift_step,predicted_mechanism_family,predicted_repair_outcome,predicted_residual_constraint_id
RDOD2-026,no_drift,,none,not_applicable,none

Predictions must contain exactly one row for every gold scenario.

The scorer rejects:

• Missing predictions
• Extra predictions
• Duplicate scenario identifiers
• Unknown scenario identifiers
• Invalid mechanism families
• Invalid repair outcomes
• Invalid constraint indices
• State-inconsistent combinations

---

Official evaluation structure

The benchmark exposes two official leaderboards.

They answer different questions and should not be collapsed into one ranking.

---

Leaderboard A — Component Capability

Leaderboard A evaluates performance on four principal reasoning-state components.

The official metrics are:

outcome_mode_macro_f1
exact_onset_accuracy
mechanism_family_macro_f1
repair_outcome_macro_f1

These measure whether the model can independently identify:

• The state reached by the trajectory
• The point at which drift began
• The structural family governing the drift
• The degree to which the drift was repaired

Primary ranking metric

Models are ranked by:

component_mean

Formula:

Component Mean =
(
    Outcome-Mode Macro F1
    + Exact Onset Accuracy
    + Mechanism-Family Macro F1
    + Repair-Outcome Macro F1
) / 4

Each component receives equal weight.

The Component Mean is a transparent convenience aggregate.

It does not imply that:

• The four tasks are theoretically equivalent
• Their weights have been empirically optimised
• The aggregate should replace the individual metrics

All four component metrics must be displayed beside the Component Mean.

What Leaderboard A asks

How well can the model identify each component of the reasoning state?

Leaderboard A is primarily diagnostic.

It can reveal whether a model’s weakness lies in:

• Outcome recognition
• Transition localisation
• Mechanism attribution
• Repair assessment

---

Leaderboard B — Complete Reasoning State

Leaderboard B evaluates whether the model reconstructs the full primary reasoning state for each case.

A case is counted as correct only when all five primary targets match exactly:

outcome_mode
first_drift_step
mechanism_family
repair_outcome
residual_constraint_id

Primary ranking metric

Models are ranked by:

complete_primary_accuracy

Formula:

Complete Primary Accuracy =
Number of cases with all five primary targets correct
/
Total number of evaluated cases

Partial agreement receives no credit.

A prediction is not a complete-state match when:

• The outcome is correct but the onset is wrong
• The onset is correct but the mechanism family is wrong
• The categorical fields are correct but the residual constraint is wrong
• A repair is detected but its repair outcome is misclassified

What Leaderboard B asks

How often can the model reconstruct the complete reasoning state as a coherent whole?

This is a stricter task than component-level evaluation.

A model may perform strongly on Leaderboard A while performing substantially worse on Leaderboard B.

That gap distinguishes:

component competence
from
integrated state reconstruction

---

Secondary diagnostic metrics

The scorer reports additional metrics that do not determine the primary ranking on either leaderboard.

Within-one onset accuracy

Reports whether the predicted onset is within one step of the gold onset.

For numerical onsets:

absolute error <= 1

For no-drift cases, only a blank onset receives credit.

This metric is not included in the Component Mean because exact onset accuracy is already represented there.

Mean absolute onset error

Measures the average numerical distance between predicted and gold onset steps where both are numerical.

It describes localisation error but does not replace exact onset accuracy.

Residual-constraint exact accuracy

Measures whether the model selected the correct unresolved constraint ID.

Residual constraints are scored by exact constraint selection, not free-text token overlap.

Expected calibration error

Confidence is calibrated against complete-primary correctness.

A confidence value therefore estimates:

How likely is it that all five primary targets are exactly correct?

Confidence is optional.

When a model supplies no confidence values, expected calibration error is omitted.

Confidence coverage

Reports the proportion of prediction rows that include a valid confidence value.

Per-class metrics

The scorer reports precision, recall, F1, and support for:

• Outcome mode
• Mechanism family
• Repair outcome

Class support

Raw support counts are always reported.

This is especially important because the current test set is small and some classes have limited representation.

Prediction table

The scorer returns a per-scenario audit table containing:

• Gold targets
• Predicted targets
• Field-level correctness
• Complete-primary correctness
• Onset absolute error
• Confidence

---

Scorer output

The scorer returns a JSON report structured as:

{
  "benchmark": {
    "name": "Reasoning Drift Onset Detection v0.2",
    "evaluation_type": "structured_reasoning_state",
    "n": 12
  },
  "leaderboard_a_component_capability": {
    "primary_ranking_metric": "component_mean",
    "component_mean": 0.7425,
    "outcome_mode_macro_f1": 0.81,
    "exact_onset_accuracy": 0.75,
    "mechanism_family_macro_f1": 0.68,
    "repair_outcome_macro_f1": 0.73
  },
  "leaderboard_b_complete_reasoning_state": {
    "primary_ranking_metric": "complete_primary_accuracy",
    "complete_primary_accuracy": 0.42
  },
  "secondary_diagnostics": {
    "within_one_onset_accuracy": 0.88,
    "mean_absolute_onset_error": 0.64,
    "residual_constraint_exact_accuracy": 0.79,
    "expected_calibration_error": 0.11,
    "confidence_coverage": 1.0
  },
  "class_support": {},
  "per_class_metrics": {},
  "prediction_table": []
}

---

Validation

The repository includes a validator for:

• Labelled dataset files
• Blind model-input files
• Prediction files
• Full repository structure

Validate the repository

python scorer/validate.py repository

This validates:

data/train.csv
data/dev.csv
data/test.csv
data/test_blind.csv

It checks:

• Required files
• Exact header order
• Split overlap
• Matching test and blind identifiers
• JSON validity
• Constraint uniqueness
• Trajectory structure
• Contiguous step numbering
• Outcome-state consistency
• Mechanism-family validity
• Repair-outcome validity
• Correction-source validity
• Constraint-text matching
• Onset bounds
• Visible-failure bounds
• Trajectory-length accuracy
• Relative-onset accuracy
• Annotation-confidence range
• Blind-split label exclusion

Validate labelled files

python scorer/validate.py labelled \
  data/train.csv \
  data/dev.csv \
  data/test.csv

Validate the blind split

python scorer/validate.py blind \
  data/test_blind.csv

Validate predictions

python scorer/validate.py predictions \
  predictions.csv \
  --gold data/test.csv

The prediction validator confirms:

• Exact scenario-set matching
• Valid output labels
• Valid residual-constraint IDs
• Optional confidence range
• State-consistent prediction combinations

---

Baseline

The repository contains a deliberately weak heuristic baseline.

Run it with:

python baseline/baseline.py \
  data/test_blind.csv \
  predictions.csv

Then validate the predictions:

python scorer/validate.py predictions \
  predictions.csv \
  --gold data/test.csv

Then score them:

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

Save the score report:

python scorer/score.py \
  data/test.csv \
  predictions.csv \
  --output results/score_report.json

The baseline is intended to verify the input-output pipeline.

It is not intended as a competitive model.

---

Interpretation guidance

Leaderboard A and Leaderboard B must be interpreted separately.

A high Component Mean indicates

Broad component-level competence across:

• Outcome classification
• Onset localisation
• Mechanism-family identification
• Repair-outcome classification

A high Complete Primary Accuracy indicates

Reliable integrated reconstruction of the complete reasoning state.

Neither result should be presented without:

the full component breakdown
complete-primary accuracy
class-support counts
test-set size
per-class metrics

Macro F1 can be unstable when individual labels have limited support.

A class represented once can only produce extreme recall values.

The benchmark therefore reports raw support alongside aggregate metrics.

---

Current limitations

Small dataset

The dataset contains only 36 labelled trajectories.

This is sufficient for task definition, schema validation, scorer development, and initial model probing.

It is not sufficient for mature model ranking.

Small test split

The blind test contains 12 cases.

The test exercises the supported outcome modes, repair outcomes, and correction-source categories, but several labels have sparse support.

Label-space coverage is not the same as statistical stability.

Sparse mechanism support

Some mechanism families have only a small number of examples.

Fine-grained drift_mechanism labels are even more sparse.

Mechanism-family performance should therefore be interpreted cautiously.

Fine-grained mechanism prediction is not an official leaderboard task in v0.2.

Synthetic trajectories

The trajectories are designed benchmark examples.

They are not naturally generated model traces.

Synthetic design makes structural control possible but may also produce regularities not present in real reasoning.

Explicit constraints

The governing constraints are supplied to the model.

The benchmark does not currently test whether a model can infer hidden governing constraints.

Domain breadth without domain depth

The dataset spans several reasoning domains, but most domains contain only a small number of examples.

The benchmark demonstrates cross-domain instantiation.

It does not support reliable domain-specific performance estimates.

Annotation subjectivity

Outcome mode is expected to be more reproducible than fine-grained mechanism subtype.

The present release does not include formal inter-annotator agreement.

No human baseline

The current release does not include a human performance estimate.

No hidden evaluation server

The labelled test file is public.

Evaluation remains reproducible, but test-set adaptation cannot be prevented without a hidden server or private test set.

---

Inter-annotator agreement roadmap

A future annotation study should use multiple independent annotators and report agreement separately for each field.

Recommended measures:

Field	Suggested measure
Drift presence	Fleiss’ kappa
Outcome mode	Fleiss’ kappa
First drift step	Exact and within-one agreement
Mechanism family	Fleiss’ kappa
Repair outcome	Fleiss’ kappa
Residual constraint	Exact constraint-selection agreement
Fine drift mechanism	Exploratory agreement
Severity	Weighted agreement or adjudicated comparison

Expected agreement is likely to be highest for:

drift presence
outcome mode
repair outcome

Moderate for:

first drift step
mechanism family

Lowest for:

fine mechanism subtype
severity
repair instruction

---

Recommended status

Version 0.2 should be treated as:

A structured task-definition, annotation-schema, scorer, validator, and evaluation-pipeline seed.

It should not yet be treated as:

• A definitive model-ranking benchmark
• A production safety evaluation
• Evidence of domain-specific professional competence
• A statistically mature leaderboard
• A substitute for expert review

---

Roadmap

A stronger future release should contain at least:

train: 240
dev: 60
test: 60

A more robust target would be:

500–1000 total trajectories

Future development should include:

• Several examples per mechanism family
• Balanced outcome support
• Balanced repair-outcome support
• Wider onset-position coverage
• More 3–4 step trajectories
• More 8–12 step trajectories
• Multiple drift events
• Recurrent drift
• Partial repair
• Failed repair
• Overcorrection
• Recovery followed by relapse
• Branching reasoning
• Lexically adversarial controls
• Stable near-neighbour controls
• Natural model-generated trajectories
• Hidden-constraint evaluation
• Domain-held-out testing
• Human annotation
• Expert adjudication
• Inter-annotator agreement
• Human baselines
• Public model baselines
• Hidden-label evaluation

---

Licence

MIT.

The dataset is synthetic and contains no third-party source data.

---

Citation

Caplan, M. (2026).
Reasoning Drift Onset Detection v0.2.
Clarus Invariant / SIOS.
Dataset.

See:

CITATION.cff

for machine-readable citation metadata.

---

Status

status: structured_reasoning_state_seed
version: 0.2.0
publication_ready: true
schema_validated: true
evaluation_pipeline_validated: true
model_comparison_validated: false
blind_test_label_space_complete: true
blind_test_class_support_sufficient: false
leaderboard_statistically_mature: false
known_fixed_position_shortcut: false
human_baseline_available: false
inter_annotator_agreement_available: false
hidden_test_server_available: false

Official benchmark framing

Leaderboard A — Component Capability
Primary metric: Component Mean

Leaderboard B — Complete Reasoning State
Primary metric: Complete Primary Accuracy

The two leaderboards answer complementary questions:

Leaderboard A:
Can the model identify the parts?

Leaderboard B:
Can the model reconstruct the whole?