# The 10% Project: useful work needs a receipt and a recipient

Design and implementation notes — 2026-09-08.

## What is actually running today

The browser can perform a bounded local search on a mathematical graph. It starts only when the player explicitly clicks **Try a search**, or opts into one job after every tenth completed battle. The automatic option is off by default. One worker runs at a time, targets a 350 ms search budget with at most 1,200 candidate changes, and is terminated after 2 seconds if it has not returned. Starting a fight or hiding the page terminates the worker. No work continues after closing the tab.

This is **local CPU computation, not LLM inference**. One job per ten battles is not a measurement of ten percent of resources. No model API, paid account, donation, remote queue or receiving institution is connected. Results stay in local browser storage until the user downloads them. Local record counters are editable and are not certificates.

## Mission 001: Pattern Hunter

The task colors each edge of the complete graph on 43 vertices red or blue and tries to eliminate monochromatic complete subgraphs on five vertices. The search starts randomly or resumes the best local graph. It proposes edge flips, tracks the change in the number of forbidden groups, and independently recounts all groups before returning its best candidate.

The April 2026 survey records bounds of 43 ≤ R(5,5) ≤ 46 and describes evidence for the conjecture that the exact value is 43. A valid coloring on 43 vertices with **zero** monochromatic K5s would imply a lower bound of 44 and would require careful independent examination. [Radziszowski, Small Ramsey Numbers, revision 18, sections 2.1 and note (e)](https://www.cs.rit.edu/~spr/ElJC/ejcram18.pdf)

This deliberately small heuristic is an exploratory prototype, not a competitive research solver. It is very unlikely to find a frontier counterexample. Fewer conflicts measure improvement on this local optimization objective, not a published theorem or independently established advance. Failure to find a graph does not prove nonexistence. A million unsuccessful random trials do not resolve the open problem.

Counters show completed local jobs, the best graph's conflict count, and worker-measured elapsed time. They do not show “discoveries,” “cures,” global compute donations or researcher-accepted progress. There is currently no deduplication across players, campaign coverage guarantee or recipient. A hash identifies a graph; it does not authenticate who computed it or whether it is scientifically novel.

## Check the evidence

Download the latest JSON receipt, then run:

```bash
python3 scripts/verify-research.py robowars-research-receipt.json
```

The standalone Python verifier reconstructs the graph and enumerates all 5-vertex subsets. It checks the claimed conflict count using an implementation separate from the browser optimizer. It returns a graph SHA256 and an explicit scientific-status label. It does not verify the claimed elapsed time, seed history, number of moves or novelty. Node tests also compare optimized counts and incremental deltas against a separate exhaustive enumerator on small graphs.

A zero-conflict receipt would need a second implementation, a reproducible construction, qualified mathematical review, a current literature check and publication of the complete artifact. The website must not automatically publish a discovery claim from a client-submitted result.

## A meaningful future ten percent

Define the fraction in money, with a clear denominator. Example **proposal** for a fixed $100 monthly model/compute envelope: authorize up to $90 of battle inference and $10 of research work. Verification and rejected/failed research jobs are charged to the research allocation and reported separately from accepted outputs. Hosting, payment fees and ordinary site costs are outside this narrowly defined envelope unless the published policy says otherwise. There is no such funded budget configured today.

A production allocator needs atomic reservations: reserve the full maximum job cost before dispatch, prevent concurrent jobs from overbooking the cap, reconcile actual provider billing, return unused reservations, and stop dispatch on exhausted budget or unavailable verifier. Report reserved, actually spent, failed, verified and accepted figures separately. Unused research funds should remain visibly unspent, never silently relabeled as impact.

Publish the mission definition, version, dataset provenance, fixed evaluation suite, compute cap, provider invoice reconciliation, candidate hashes, checker version and recipient acceptance status. Make opt-out straightforward. Treat a “10% allocated” promise and a “10% actually spent” claim as different things.

## Better first missions than an open-ended miracle hunt

| Mission                       | Model/computation role                                                                        | Independent acceptance                                                                                                 |
| ----------------------------- | --------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------- |
| Open-source repair clinic     | Agents propose fixes to maintainer-selected reproducible bugs                                 | Hidden regression tests, human review and maintainer acceptance; passing tests alone does not guarantee a useful patch |
| Formal-math subgoals          | Propose lemmas/proofs in a researcher-chosen formalization project                            | Kernel-checked proof plus confirmation the statement matches the intended task; exclude sorry/axiom shortcuts          |
| Bounded combinatorial search  | Explore registered seeds or verify a candidate construction                                   | Reproducible full artifacts, independent verifier, task-owner-defined usefulness                                       |
| Materials data quality        | Extract measurements with provenance, units, pressure and uncertainty from an approved corpus | Independent source check and curator acceptance; no model-consensus substitute for the experiment                      |
| Materials candidate screening | Propose candidates within a scientist's fixed simulation workflow                             | Qualified physics validation, uncertainty/calibration, stability checks and eventual experimental follow-up            |

Citizen-science games can support research when the scientific problem, player interaction and verification are deliberately connected. Quantum Moves 2 studied this for particular quantum-control tasks. That precedent does not establish that arbitrary robot combat or unused animation time produces valuable science. [Primary research](https://arxiv.org/abs/2004.03296)

## About room-temperature levitation

Make it a long-term scientific theme, not the launch promise. Superconductivity at room temperature and ordinary atmospheric pressure involves more than predicting a high transition temperature: a proposed material must also be stable, synthesizable and experimentally characterized. A materials workflow could assist candidate screening or evidence curation, but an LLM suggestion or an image of levitation does not establish a superconductor. [Computational hydride study](https://www.nature.com/articles/s41524-024-01214-9)

The Materials Project provides public calculated materials properties that could inform a carefully scoped mission, subject to dataset terms and scientific review. It is an existing research resource, not a RoboWars partner. [Materials Project documentation](https://docs.materialsproject.org/)

Before spending serious compute: recruit a task owner, agree on what constitutes useful output, run a small capped pilot, and compare it against a specialist solver or existing workflow. If the game adds no research value, direct the reserved money to an effective external project instead of generating impressive-looking activity.