---
language:
- en
tags:
- masked-diffusion
- learning-theory
- reproducibility
- numerical-verification
pretty_name: MDM Coverage Experiments
---

# MDM Coverage Experiments

Code, configurations, numerical data, finite certificates, and figures for
**Sharp Accuracy Thresholds for Parallel Masked Sampling of Noisy Pairs**.
This backup combines the original experiment server with the local audits and
the compact reproducibility package. It preserves exploratory candidates,
failed and retried solver cases, and superseded plots alongside the final results.
No neural-network training, external dataset, or GPU is needed.

## Where to start

| Folder | Purpose |
| --- | --- |
| `reproducibility/` | Curated sources, frozen inputs, results, and runnable verification instructions for all four phases |
| `server/` | Original server experiment tree, including full candidate searches, logs, retries, historical archives, and the September 25 revision |
| `local_archive/` | Local experiment copies, independent audit scripts/results, and corrected presentation artifacts |
| `environment/` | Archived original server virtual-environment wrapper and portability notes |
| `MANIFEST.sha256` | SHA-256 checksums for every other uploaded file |
| `BACKUP_CONTENTS.json` | Backup scope and source-to-copy checksum verification |

For final paper-facing plots, start with
`local_archive/revision_20260925/results_final/figures/`, the corrected legacy
plots in `local_archive/revision_20260925/figures/`, and the phase-3
`presentation/` directory. Historical figures remain in their original paths;
they are not all current manuscript figures.

## Quick exact verification

From the repository root:

```sh
cd reproducibility
python3 -B verify_frozen.py --phase all --output-dir verification_output
python3 -B checks/verify_interface_selection.py
```

The first command requires a **new** output directory and uses only the Python
standard library. It recomputes 27 fair-model and 15 biased-model certificates
and compares every stored non-timing field. Detailed commands and the numerical
interpretation are in [the reproducibility README](reproducibility/README.md).
The stored validation report is `reproducibility/VALIDATION.json`.

## Numerical environment

The reported numerical experiments used Python 3.13.13, NumPy 2.2.6, SciPy
1.17.1, and Matplotlib 3.10.9. Create a fresh environment outside the frozen
result directories:

```sh
python3.13 -m venv .venv-local
source .venv-local/bin/activate
python -m pip install -r reproducibility/requirements-numerical.txt
export OPENBLAS_NUM_THREADS=1
export OMP_NUM_THREADS=1
export MKL_NUM_THREADS=1
export MPLBACKEND=Agg
```

The original `.venv` is also retained as an archive, but it shares its parent
Conda installation and is **not a self-contained or portable environment**.
Use the pinned requirements for a fresh machine.

## Full experiment commands

Run phases 1--3 from `server/`. Use unique run names beginning with `run-`;
the runners protect existing result directories.

```sh
cd server
python src/run_study.py --config configs/phase1.json --run-id run-local-phase1
python phase2/src/run_phase2.py --config phase2/configs/phase2.json --run-id run-local-phase2
python phase3/run_phase3.py --run-id run-local-phase3
```

Phase 3 uses the committed phase-2 witness inputs; its results are therefore
independent of the name of a newly run phase-2 search. The original recovery
protocol for the two phase-2 LP timeouts, individual-stage commands, and plotting
commands are documented in [the server README](server/README.md).

For phase 4, from the repository root:

```sh
cd server/revision_20260925
python run_revision.py --output results_local --max-seconds 480
```

This writes new results without replacing `results_final/`. The full protocol,
matched-KL cases, infeasible setting, and floating-point quadrature diagnostics
are in `results_final/REPORT.md` and `results_final/quadrature_validation.json`.
The curated README also explains how to reproduce plots in a copied directory.

## Interpreting the data

- Phase 1 contains floating-point model/execution diagnostics and small policy LPs.
- Phase 2 contains rational finite-oracle certificates for selected fair models,
  as well as the exploratory candidate search and its unsuccessful cases.
- Phase 3 contains fixed-schedule comparisons and rational biased-model certificates.
- Phase 4 contains deterministic floating-point count-event lower witnesses
  for tested policies and evaluations of an analytic selection upper bound.

The exact finite certificates and the floating-point diagnostics have different
guarantees. These experiments do not replace the analytic proofs, establish
asymptotic exponents by fitting curves, or measure performance on language data.
Source/configuration hashes and original status records are preserved.

## Backup integrity and scope

Verify a downloaded copy with either command:

```sh
sha256sum -c MANIFEST.sha256       # Linux
shasum -a 256 -c MANIFEST.sha256  # macOS
```

The source snapshot is dated October 7, 2026 (America/Los_Angeles). Server and
local copies are kept separately where provenance or presentation differs.
Git metadata, unpacked Python bytecode caches, SSH credentials, and authentication
tokens are excluded. Historical run archives are preserved as received after
inspection; they may contain redundant source and bytecode copies. The manuscript,
review correspondence, and unrelated server projects are outside this backup.

No new software or data license is granted by this backup. Existing vendored
dependency license notices remain in the archived environment. An explicit
research-code/data license can be added separately by the owner.