ShrineTech · Research
Research
ShrineTech takes part in open science challenges, tests its ideas carefully, and publishes honest results and free tools.
Paper · NeurIPS 2026 Smart Buildings workshop
Doing Less Transfers: one heating and cooling controller for buildings it has never seen
A small, readable rule controller for simulated office buildings, tuned by a plan written down before testing and checked on two buildings kept aside until the end.
What's new
- A tuning protocol fixed in advance, with whole buildings held out rather than single scenarios.
- The "doing less" finding: the well-known structural switches each move the score by about 2% at most; the gentle, tuned control loops carry the gain.
- A reproducibility catch: the baseline numbers shipped with the challenge kit can't be reproduced on the current kit.
Result (honest)
On the 2 held-out buildings it scored 0.0204 against 0.0449 for the best kit baseline (lower is better), winning 10 of 10. On the competition's hidden Kaggle test it scored 0.0391, rank 504 of 599, so the transfer has not shown there yet. Simulator only.
Submitted to OpenReview 2026-11-09 · Round One · under review
Community tools · NeurIPS 2026 Virtual Embryo Challenge
Virtual Embryo Challenge: two free tools for participants
Open-source helpers that make a first entry easier and the external-data rule checkable. MIT-licensed, not affiliated with the organisers.
What's new
- vec-t1-zero-data-starter: a first valid Task 1 submission, built and scored on synthetic data, with no registration or 1.2 GB download. It catches the two silent mistakes (wrong gene order, wrong normalisation).
- vec-windows: checks the Rules s.10 external-data "protected windows" cell by cell, writes the drop list and the disclosure text, and extends rpinho's ve-stagecheck one level down.
Result (honest)
Both run on a CPU and come with tests that use synthetic data only. The starter's synthetic scores say nothing about the real leaderboard, and vec-windows' somite and Theiler stage mapping is approximate; both READMEs say so.
Filed as community contributions 2026-11-09 · awaiting review
In progress: Earthquake-fault mapping (DOE GEMS, results after the Dec 3 close) · Vesuvius Challenge contributions · A long-term study of the P vs NP problem (ongoing work)
© ShrineTech LLC ·
Research ·
Privacy