GameNav-CriticalEdits-25K

RegalFire — AI Data Foundry. An original synthetic gaming / procedural navigation computational fixture benchmark.

Problem and differentiator

Predict reachability and minimum landing-cell movement cost before and after toggling one obstacle.

Why download this instead of existing data?

Paired critical one-cell edits include exact optimal route costs and executable path witnesses before and after the edit. This isolates route sensitivity in a small auditable setting, complementing large MovingAI maps and general PCG environments.

Research context: github.com, arxiv.org. These references support the relevance of the task area; they do not establish customer demand or that this release improves any model.

Verified release

25,000 accepted records; 25,000 unique IDs; 25,000 unique semantic task fingerprints; 24,948 source groups. Zero rejected records. Independent computational validation: PASS. Full restarted generation was byte-identical (see reproduction_report.json). No model training/evaluation run is claimed.

Split	Records
train	20020
validation	2527
test	2453

Construction family	Records
boundary_gap	5000
gap_close	5000
gap_open	5000
two_gap_detour	5000
weighted_gap_close	5000

Ground-truth contract

Positions are [row,column], zero based. Four orthogonal moves, no waits or diagonals. # is blocked and . is traversable. Entering a cell costs its supplied positive integer (1–5); the starting cell is free. Toggle exactly the edit cell. Null cost/path means unreachable. Valid optimal paths need not be unique.

Generator: A* with admissible Manhattan bound. Validator: independently written repeated shortest-path relaxation plus path replay.

Record and file structure

Each JSONL row has id, index, seed, family, input, expected, fingerprint, group_id, split and provenance. Input contains the complete task; expected contains computable labels and witnesses where relevant. schema.json specifies exact field/type requirements. There are five full worked examples in examples/sample.json. data/dataset.jsonl is the complete canonical dataset; HF default loading uses only the disjoint train/validation/test files.

For agent evaluation, remove expected before passing inputs to the agent and keep the answer files inaccessible to it. These are public fixed-label fixtures, not contamination-resistant secret evaluations or recorded agent rollouts.

Reproduce and validate

Seed: 20261001. Published source and the generation environment are recorded in SOURCE_LICENSE.json. Exact byte reproduction was verified in that environment; matching seed alone is not a guarantee across Python versions.

pip install -r requirements.txt
python generator/generate.py --out reproduced --count 25000 --seed 20261001
python validator/validate.py --dataset reproduced/dataset.jsonl --count 25000 --out reproduced_qa
python generator/generate.py --out reproduced --count 25000 --seed 20261001 --resume
python -m unittest discover -s tests -p test_oracles.py -v

Writes use committed 500-record chunks, atomic checkpoint updates and fresh final assembly. Resume loads committed shards, verifies binding to seed/count/source, and regenerates any uncommitted work. A missing final file or checkpoint does not destroy committed records. Do not mix different seeds, counts or generator versions in one output folder.

Generator and validator implement different label algorithms; shared helpers only cover serialization, schema and canonical semantic identities. This is independent computational implementation, not independent human scientific review. Nine unit tests cover label corruption, exact small cases, schema types, missing/corrupt files, semantic symmetries and crash-resume reproduction. All passed locally; see the QA and reproduction reports.

Deduplication, splits and quality limits

Family quotas are equal. Split assignment is deterministic by source-group hash (approximately 80/10/10; exact counts above). Identical source groups never cross splits. Grid rotations/reflections are canonicalized; agent-label swaps are canonicalized in coordination tasks. Interval translation equivalents are canonicalized. DNA source grouping canonicalizes reverse complements. Each paired edit or sequence pair remains within its record.

The DNA near-duplicate audit screens 7-mer bottom-four sketches and at most 25 previous candidates per bucket, then checks Jaccard >=0.85 across split groups. It found 0 flags. This is a bounded similarity screen, not exhaustive homology analysis. Grid and interval checks cover exact stated equivalences, not all possible near-isomorphic or related scenarios. Splits are by generated source, not proven biological species or real game distribution. Public fixtures can be memorized.

Provenance, privacy and licensing

Entirely generated by the original supplied procedural source. No external dataset records, genomes, annotations, patient data, PHI, human identity profiles, proprietary game images, dialogue, maps, textures or binaries were used. Cited tools and papers are research context only; their data/code were not copied. MIT applies to this original release and generator/validator code; retain the LICENSE notice. SOURCE_LICENSE.json records permissions, attribution and environment.

Scope and limitations

Small 6–10 by 6–10 constructed barrier maps, not arbitrary game topology, learned agent traces, 3D physics or real telemetry. Most critical edits change connectivity; one family changes a still-solvable detour cost. This is a sensitivity fixture benchmark, not proof of general navigation ability.

Tests establish the specified computational cases, not universal software correctness, biological validity of a simulation, clinical suitability, benchmark novelty over every related work or market value. Generator and validator were authored with Codex and have not been reviewed by a domain expert.

Commercial extension: Private costs, map schemas, movement rules and planner regression fixtures. Domain rules and privacy/access constraints must be agreed and tested before delivery.

Custom / Private Dataset Work

RegalFire builds custom AI datasets, evaluation sets and reproducible data pipelines for research and production systems.

Available adaptations include:

• private schemas
• larger volumes
• custom domains
• custom validators
• private evaluation sets
• continuous generation
• custom environments

Contact: ootiris@gmail.com Hugging Face: RegalFire