hi, stelios here.|

What I implemented

The inverse-problem pipeline, measurement operators, conditioning and data consistency, four solver modes, checkpoint conversion and experiment analysis. The original model architecture and pretrained prior are credited to the reference work.

What the experiments showed

Sample averaging improved reconstruction at additional sampling cost. Short adaptation was evaluated under domain shift, including a subject-disjoint OASIS-1 study. Uncertainty-guided acquisition did not outperform random selection under an equal extra-column budget.

What this does not establish

These are retrospective, single-coil reconstruction experiments on public imaging datasets. Uncertainty ranking is not a calibrated probability of failure. There was no prospective scanner study or clinical validation.

Bounded execution

Recorded replay is free to browse. An optional curated trial makes one GPT-OSS 20B request for a fixed task, with isolated verification and persistent spending limits. Live trials are separate from benchmark results; visitors cannot submit code or arbitrary prompts. Provider failures remain visible.

Policy comparison

Strong-only, cheap-only, escalation-on-failure, static and adaptive policies share a six-call, ten-decision cap. The controlled research profile routes GPT-OSS 20B and 120B through a fixed OpenRouter provider. Hidden checks grade the final candidate after all routing decisions.

Transfer pilot and four-seed follow-up

A separately frozen transfer pilot completed 162 episodes across 27 new miniature tasks and six policies. The simpler learned baseline solved 17/18 primary test tasks; fitted-Q, strong-only and escalation solved 16/18. One seed and six test families do not establish general superiority. All policies shared the same additional verification rule.

A later four-seed replication revisited the same, now-exposed catalog in 1,296 attempted episodes. On its original primary slice, strong-only solved 131/144 at lower accounted model cost than the fitted-Q router's 128/144. VERIFY-on and VERIFY-off each solved 372/432 attempts across policies. Four provider failures remain in the counts; this is not new-task generalization.

Limits of the evidence

These are authored miniature repositories, not evidence of general repository repair or long-horizon autonomy. Sparse controller states use a declared fallback. The dashboard distinguishes actual coverage, provider charges and uncertain reservations; no general policy advantage is claimed.

Implementation

React, Express and PostgreSQL, with nginx and systemd on a VPS. The work includes application flows and the operational checks needed to maintain them.

Boundaries

A single-VPS deployment is not a highly available infrastructure. Encryption of selected records and backups is not whole-disk or end-to-end encryption. Membership verification does not certify a medical licence. Private source and member data are not published here.

What the live system establishes

Public-feed integration, bounded historical storage and automatic temporal evaluation. Forecasts require at least 14 days plus coverage and freshness checks. The target is a later arrival-estimate proxy, not actual train passages, passenger waiting time or validated incident prediction.

What the replay does not establish

The small replay contains three constructed scenarios, not official MTA incident ground truth or an independent benchmark. Higher recall in this fixture comes with more false alarms than the baselines. Neither the replay nor a working live map establishes validated incident prediction.

Observed throughput

At concurrency 32, the later study measured cached/uncached throughput ratios of 1.23–1.47 with shared prefixes across its GPU/model combinations; unique prefixes were near 1.00. Four-repeat means and ranges, latency and exact-output hash counts are in the report.

Scope of the comparison

Each hardware class used one allocated GPU and synthetic fixed-length prompts. Output text often differed. The $3.45 allocation cost is an estimate; exact provider billing is unavailable. Neither study measures answer quality, quantization or general production performance.