README.md

---
license: apache-2.0
language:
  - en
pretty_name: VLR-Recurrence-Verified Paper-Scale
task_categories:
  - text-generation
tags:
  - code-repair
  - preference-optimization
  - external-verification
  - protected-recurrence
  - semantic-family-ood
  - regression-aware
configs:
  - config_name: primary-tuples
    data_files:
      - split: train
        path: data/primary_tuples_train.jsonl
      - split: validation
        path: data/primary_tuples_validation.jsonl
      - split: test
        path: data/primary_tuples_test.jsonl
  - config_name: preference-balanced
    data_files:
      - split: train
        path: data/preference_balanced_train.jsonl
      - split: validation
        path: data/preference_balanced_validation.jsonl
      - split: test
        path: data/preference_balanced_test.jsonl
  - config_name: preference-exhaustive
    data_files:
      - split: train
        path: data/preference_exhaustive_train.jsonl
      - split: validation
        path: data/preference_exhaustive_validation.jsonl
      - split: test
        path: data/preference_exhaustive_test.jsonl
  - config_name: accepted-action
    data_files:
      - split: train
        path: data/accepted_action_train.jsonl
      - split: validation
        path: data/accepted_action_validation.jsonl
      - split: test
        path: data/accepted_action_test.jsonl
  - config_name: certified-generation
    data_files:
      - split: train
        path: data/certified_generation_train.jsonl
      - split: validation
        path: data/certified_generation_validation.jsonl
      - split: test
        path: data/certified_generation_test.jsonl
  - config_name: recursive-trajectory
    data_files:
      - split: train
        path: data/recursive_trajectory_train.jsonl
      - split: validation
        path: data/recursive_trajectory_validation.jsonl
      - split: test
        path: data/recursive_trajectory_test.jsonl
  - config_name: terminal-stop-auxiliary
    data_files:
      - split: train
        path: data/terminal_stop_auxiliary_train.jsonl
      - split: validation
        path: data/terminal_stop_auxiliary_validation.jsonl
      - split: test
        path: data/terminal_stop_auxiliary_test.jsonl
---

# VLR-Recurrence-Verified

VLR-Recurrence-Verified is a synthetic-data construction release for studying
evidence-convergent program repair. It operationalizes a protected partial order:
a candidate is positive only when it preserves every already-satisfied
obligation and strictly improves at least one unresolved obligation.

## Scale

| Split | Tasks | Families | Transitions | Balanced pairs | Certified finals |
| --- | ---: | ---: | ---: | ---: | ---: |
| Train | 3,500 | 28 | 12,250 | 49,000 | 3,500 |
| Validation | 750 | 10 | 2,623 | 10,492 | 750 |
| Test | 750 | 12 | 2,623 | 10,492 | 750 |
| Total | 5,000 | 50 | 17,496 | 69,984 | 5,000 |

The release contains 224,952 records across the same seven views and 21 files
as the earlier version. These records derive from 5,000 independent tasks; view
rows that share a task or transition are not independent observations.

## Construction and labels

Each task defines 16–31 protected executable obligations and a deterministic
2–5 step canonical repair path. The external verifier executes the incumbent,
accepted candidate and every negative against the same immutable obligations.
Only componentwise non-regression with at least one strict improvement receives
a positive label, and only an all-zero EvidenceRank receives certification.
The 50 semantic families span additive corrections, multiplicative gains and
bit-mask repairs over integer, text, sequence and structured-record programs.

Balanced preference materialization includes exactly one example from each
negative class per transition:

- `equal_rank`: the artifact changes but protected evidence does not;
- `protected_regression`: at least one passing obligation breaks and none improves;
- `pareto_incomparable`: improvements and regressions coexist without aggregate gain;
- `aggregate_trap`: aggregate failures decrease while a protected obligation regresses.

All v4 hard negatives are structural parameter regressions over protected input
regions. The audit rejects candidates containing the legacy single-hidden-input
exception pattern.

## Leakage controls

Splits are assigned before view materialization at task and semantic-family
level. Validation and test semantic families never appear in optimization.
Cross-split overlap is zero for task IDs, family IDs, exact accepted artifacts
and identifier/literal-normalized accepted Python ASTs. Model-visible prompts
contain sanitized failure counts and opaque obligation IDs, never protected
inputs, expected hidden outputs or full audit receipts.

## Recommended post-training

Use `accepted_action_train`, `certified_generation_train` and
`preference_balanced_train`. Keep the synchronized primary, exhaustive,
trajectory and STOP views out of ordinary loss unless an explicitly reported
ablation changes that contract. The supplied trainer uses the shared Hugging
Face engine with a VLR-specific length-normalized protected-order objective,
inverse-frequency route balancing and stronger regression-boundary weights.

Do not infer deployment correctness from model likelihood or a generated STOP
token. The trained model proposes candidates; an independent verifier retains
exclusive commit, rollback, certification and stopping authority.

## Evaluation protocol

Checkpoint selection uses the family-OOD validation split. Primary outcomes are
externally executed strict-progress rate, protected-regression rate and
zero-rank completion. The release gate evaluates one terminal transition per
independent task, reports Wilson 95% intervals, compares the frozen base and
adapter with an exact paired McNemar test, and requires no degradation on the
independent code, mathematics and quantum retention panel. The sealed test
split may be evaluated once only after configuration freeze.

For a data-construction-method paper, report at least the following ablations
under identical models and training budgets: accepted-action SFT only;
preference without protected hard negatives; unbalanced joint optimization;
the complete v4 objective; and data-scale checkpoints at 10%, 25%, 50% and
100%. Report every seed and confidence interval rather than selecting the best
seed or test epoch.

## Limitations

The task domain is deterministic Python function repair with synthetic,
machine-verifiable specifications. Semantic-family OOD evaluation is stronger
than random row splitting but does not establish repository-scale or
cross-language generalization. The evidence labels certify the generated data;
they do not make an unverified model generation correct at runtime.