PureOne/YEORIN-X-v4-Research

YEORIN-X / 여린엑스

v4 — Deadline-Leased Causal Transport

An executable research kit for uncertain camera-response correction, bounded image transport, and causal presentation experiments. This dataset repository includes the Python/C++ source, synthetic observations, separately stored scoring truth, recorded results, manuscript, shaders, browser laboratory, Windows testbed source, and practical support for scientists, engineers, and AI coding agents.

Research version: 4.0.0. Hugging Face packaging version: 4.0.0-hf.1, prepared 10 October 2026. Packaging adds workflows and documentation; it does not introduce a new validated camera algorithm.

Start with START_HERE.md. AI agents should read AGENTS.md, then ai/ARCHITECTURE.md. Engineers can go directly to the implementation guide; researchers to the experiment guide.

What you can do immediately

• Open the self-contained browser laboratory after downloading it. Compare source, identity, conditional, and naive correction on its controlled planar scene.
• Run the scalar correction-lease example and examine expiry, native freshness, and uncertainty charges.
• Build the portable native libraries and run the supplied numerical and causal-contract tests.
• Replay the primary RGB sequence with an input-only directory, then score in a separate process.
• Generate new seeded camera-lock, occlusion, zoom, gain-change, or other stress cases without overwriting published evidence.
• Give an AI coding agent the documented subsystem contracts, file map, tasks, and acceptance criteria.
• Build and investigate the own-window Windows/D3D11 testbed on a suitable Windows development machine. Its source is supplied; a validated Windows binary is not.

This release is a controlled research implementation. Real-game latency reductions, optical measurements, broad game compatibility, and superiority to Reflex 2 have not been established. An external native-preserving game presenter is still missing. These boundaries are preserved in both the human documentation and machine-readable claims.

Quick start

Use Python 3.11 or newer for the pinned replay profile below; the packaging verification environment is recorded in release/HF_RELEASE_VERIFICATION.json. The source package's unpinned dependency ranges also permit Python 3.10, which requires a compatible dependency selection. CMake 3.16 or newer and a C++17 toolchain are needed for both native builds and all parity tests.

From this repository's root, create and activate a virtual environment:

python -m venv .venv
# Linux/macOS:
source .venv/bin/activate
# Windows PowerShell instead:
# .\.venv\Scripts\Activate.ps1
python -m pip install -r requirements-replay.txt
python -m pip install -e . --no-deps
python tools/research.py doctor
python examples/lease_example.py

The replay profile uses headless OpenCV because no OpenCV GUI is needed. The --no-deps installation deliberately avoids installing a second OpenCV distribution. Use one OpenCV package per environment; see troubleshooting. Third-party packages are fetched during installation and are not bundled.

Build and verify:

python tools/research.py build
python tools/research.py test --out runs/tests-first
python tools/research.py replay --recorded --backend cpp --out runs/recorded-primary

The replay writes input_only/, replay.json, summary.json, summary.csv, logs, a provenance manifest, and a recorded-decision comparison. Its observer directory contains only config.json and observable.npz; scoring truth is supplied only to the subsequent scoring process. Recorded timing reproduces the publication schedule. Fresh timing measures this machine and can change admission decisions.

Run a new stress experiment:

python tools/research.py experiment --scenario camera_lock --seed 9101 --out runs/camera-lock-9101
python tools/research.py experiment --scenario zoom_change --seed 9102 --out runs/zoom-9102

Every output directory must be empty or new. The wrapper rejects writes to the original evidence trees. --backend python works without the compiled selector but is slower and does not constitute native parity verification. auto uses a locally built selector when available. The Windows native testbed and browser laboratory are separate from these offline RGB experiments.

Research mechanism

An image-based observer estimates camera-response hypotheses from source frames and causal input history. The replay publishes estimates only after their processing completion; it accounts for observation age and measured CPU work in an event schedule. Correction fields taper toward frame boundaries and explicitly pinned interface regions, preserving the source-image domain by allowing bounded nonrigid deformation.

The separate v4 scalar lease API ties an already prepared correction to immutable source identity, resource generation, profile epoch, publication time, and expiry. Its native-relative bound charges both uncertainty growth and the advantage that a fresher untouched image might have. For norm caps R and N, the sharp native-advantage charge is

[ \Phi(R,N)=R^2-(R-N)_+^2. ]

Lease admission is conditional on valid camera, realization, and native-path bounds in the same weighted displacement metric. It does not discover those bounds automatically. The scalar arithmetic is inexpensive; acquiring images, estimating the camera, submitting graphics work, and presenting photons remain separate costs.

Integration boundary: the delivered RGB replay uses the inherited cone/rational selector and modeled frame contract. It does not invoke the v4 scalar lease as an additional late stage. The real threaded worker, lease API, browser lab, and Windows source are complementary components rather than a completed arbitrary-game runtime. Architecture and readiness explain the seams.

Recorded evidence

The values below are imported from the verified v4 source release. Current packaging checks are separately recorded in HF_RELEASE_VERIFICATION.json.

Evidence	Recorded result	Interpretation
Supplied Python suite	253 passed, 0 skipped	Numerical and software contracts
C++/Python lease parity	4,096 cases	Scalar implementation agreement
Direct-grid distance checks	6,235 cases	Discrete moment implementation
Constructed lease stress	50,000 draws; 0 in-model violations	Conditional vector-set test
Frozen synthetic evaluation	10 cases; 1,440 frames	480 × 270 controlled observations
Primary native → corrected RMSE	3.740679 → 2.295348 px	38.638% geometric-error reduction
Primary naive point-cone RMSE	1.066917 px	Better on this benign trace; retained comparator
Hidden camera lock	1 incorrect intervention	Remaining semantic failure
Zoom transition	3 second-half actions before latch publication	Geometry unscored
Real games / optical trials / Reflex 2 comparisons	0 / 0 / 0	Practical superiority unmeasured

The primary error reduction is not a latency-reduction percentage. Camera correspondence is scored at the causal input cutoff using a same-source geometric counterfactual. The image-pass duration remains modeled. CPU arithmetic timings and GPU timestamp intervals must not be reported as input-to-photon latency.

results/summary.json is authoritative for final recorded cases. results/pre_latch/ retains a failed development version. results/inherited_v3/ contains archival fixtures, not fresh v4 results. The current release preserves the full unfavorable baseline and known failures.

Dataset contents and Viewer

This is a dataset repository, containing scientific artifacts and evaluation records rather than trained model weights. Explicit Viewer configurations select homogeneous JSONL tables; the binary arrays and source code are downloaded as repository files.

Subset	Rows	Contents
cases	10	Frozen case summaries, metric names, timing statistics, failures, evidence paths
decisions	1,440	Per-frame decisions joined with scoring records; source/model times, amplitudes, bounds, outcomes
claims	12	Conditional, tested, retained-failure, unimplemented, and unvalidated statements
documents	See hf_data/manifest.jsonl	Full Markdown references for local retrieval and AI assistance

cases and decisions use the test split to identify evaluation records. claims and documents use train as an indexing convention; they are reference text, not a validated training benchmark. Do not train on final scoring truth and then describe those same seeds as held out.

After publishing, replace the placeholder with your actual repository ID:

from datasets import load_dataset
cases = load_dataset("YOUR_USERNAME/YEORIN-X-v4-Research", "cases", split="test")
decisions = load_dataset("YOUR_USERNAME/YEORIN-X-v4-Research", "decisions", split="test")

Offline table loading:

from datasets import load_dataset
cases = load_dataset("json", data_files="hf_data/cases.jsonl", split="train")

For the complete code and NPZ observations, use a repository snapshot rather than only load_dataset:

from huggingface_hub import snapshot_download
snapshot_download(repo_id="YOUR_USERNAME/YEORIN-X-v4-Research",
                  repo_type="dataset", local_dir="yeorin-x")

Schemas, types, units, null conventions, and provenance are documented in docs/DATASET_SCHEMA.md. tools/research.py export regenerates the Viewer tables from the stored evidence; it does not rerun or improve the scientific results.

File map

Path	Purpose
src/yeorinx4/	Lease, worker/gates, guarded observer, replay, evidence and event analyzers
src/yeorinx2/, src/yeorinx3/	Inherited geometry, spectrum, observer and rational-selector dependencies
native/, native4/	Portable C++ selector, scalar lease, distance compiler and selftest
tools/research.py	Safe build/test/replay/experiment/export entry point
data/primary/	Observable arrays plus separately stored scoring truth
data/scoring_records/	Other cases' configs and scoring truth; images regenerated from seeds
hf_data/	Viewer-ready tables and document corpus
ai/, AGENTS.md, llms.txt	Agent reading order, contracts, tasks, prompts and architecture
examples/	Small lease and causal-worker examples
docs/	Implementation, experiments, platform guide, measurement protocol and limits
manuscript/	Research PDF, Markdown and editable LaTeX sources
windows/, web/, shaders/	Own-source D3D11 lab, offline browser lab and shader references
release/	Preserved upstream README/inventory, claims and packaging verification
SHA256SUMS.txt	Integrity inventory for the Hugging Face payload

AI-assisted implementation

The release contains a short agent entry point, a detailed architecture map, machine-readable descriptive contracts, prioritized engineering tasks with acceptance criteria, and ready-to-use audit and implementation prompts. The document subset can be used in a local retrieval system; no model service or remote agent is automatically invoked.

Recommended first tasks are observer profiling, hidden-transition diagnostics, controlled Windows compilation, timestamp-contract tests, and identity-pass versus source-copy measurement. Agents should add new runs under runs/, retain failed cases, distinguish measured and assumed quantities, and report validation with exact evidence paths. No component should be called integrated merely because its separate tests pass.

Limitations and research directions

Known camera calibration is still required. Unknown FOV, translucent HUDs, camera translation, depth-dependent effects, roll, arbitrary orbit pivots, and incoherent generated frames are outside the scored rotation model. Hidden lock and response changes cannot always be known before observation; strict abstention trades benefit for uncertainty.

The timing quantiles are observations, not execution-time guarantees. Refusing a late correction cannot retract submitted GPU work. Game integration must preserve the actual native alternative and be validated at equal image quality, with optical instrumentation. The physical protocol is a plan, not evidence of completed trials.

The source title's historical ~10^43 Hz label is symbolic naming, not a clock, bandwidth, energy, or physical performance specification. The contribution is a candidate research synthesis; foundational world-first novelty is not certified.

Provenance, license, and citation

The source was imported from YEORIN_X_v4_Engineering_Package.zip; its 205 original payload hashes were checked before repackaging. Original research code, data and documents are retained under their supplied MIT license. Named dependencies and third-party products retain their own licenses and trademarks.

Project curator: Maciej Nowicki. Development and release preparation are AI assisted, using Eve as the research persona. These labels do not imply human peer review, endorsement by NVIDIA or Microsoft, or a completed physical invention.

Use CITATION.cff or cite: YEORIN-X v4: Deadline-Leased Causal Transport Research Kit, research v4.0.0, packaging v4.0.0-hf.1, 10 October 2026. Add the actual dataset URL and immutable Hub revision after publication. No DOI or hosted URL is invented in this package.

For integrity checking after download:

python tools/verify_inventory.py .

For a useful issue report, include the OS, Python/OpenCV versions, backend, command, seed, run_manifest.json, logs, and whether the observation is synthetic geometry, software timing, or calibrated optical evidence. See CONTRIBUTING.md.