# RoboWars: make the rivalry worth coming back for

Decision memo — 2026-09-08. These are product proposals and testable hypotheses, not forecasts of virality or claims of existing users.

## The strongest version

**A robot fight club where the most interesting weapon is your strategy.** Start instantly. Name a build. Watch it develop a recognizable personality. Share a loss as readily as a win. Add verified useful work as a distinct, auditable mission system.

The original prototype communicated fairness and agent benchmarking well, but its predetermined outcomes, fabricated spectator counter and near-passive interaction could not establish whether a game was fun. The local alpha now offers decisions with reproducible consequences. It keeps model evaluations as a later competition class with their own evidence requirements.

## The loop to prove first

1. See two memorable robots and understand the objective in five seconds.
2. Choose a bot plus two readable tradeoffs; reach the first fight without an account.
3. During a short fight, make two or three decisions that can swing the result.
4. Get an explanation: what helped, what failed, and one concrete adjustment.
5. Rematch the same seed with a different plan, or send the challenge to a friend.
6. Return for a daily boss. Progress should add identity and mastery rather than compulsory grinding.

Implemented: four bots, three tactics, four modules, three game modes, energy-limited overrides, fixed seeds, replays, local record/achievements, challenge links and image cards. The first version gives every player every part. No pay-to-win power or sign-up wall.

## Strong next gameplay experiments

| Experiment                                              | Why it may be fun                                                               | What to measure                                                       |
| ------------------------------------------------------- | ------------------------------------------------------------------------------- | --------------------------------------------------------------------- |
| Three-slot “if this, do that” brain editor              | Players teach a robot rather than hand-write code; creations become shareable   | Different builds tried per visit; comprehension in a five-player test |
| Rival remembers your last build                         | Creates a story and encourages adaptation                                       | Rematch rate versus a static rival; avoid hidden stat cheating        |
| Visible arena hazards                                   | Convey action and tactical opportunities without a text wall                    | Whether players can explain the loss from the replay                  |
| Community boss raid                                     | Every qualified attempt damages a shared boss; everyone keeps a personal replay | Participation and repeat visits without grind pressure                |
| Scrapbook of ridiculous losses                          | A near miss is entertaining enough to share                                     | Shares per loss versus per win                                        |
| Bot names, palettes, entrance animations, victory poses | Identity and collecting without buying power                                    | Cosmetic attachment and willingness to support development            |
| “Beat the creator” weekly challenge                     | A concrete social target, especially for small creators                         | Challenge-open to completed-battle conversion                         |

Avoid starting with huge inventories, currencies, ranked live PvP, a marketplace, gambling, or arbitrary skill trees. They increase operations and obscure whether the central fight is satisfying. No virality mechanic guarantees growth.

## Share a challenge, not an advertisement

The current link contains matchup, seed, build and command timings. A friend can watch the original and play the same scenario with their own decisions. It works without a backend, though individual-result previews require a server later.

Next: short public challenge URLs; a server-rendered Open Graph card showing the robots, result and a “beat this” hook; a downloadable 9–15 second replay clip; a portrait layout that keeps the final turnaround legible; creator challenge pages; and fair replay verification before global rankings.

Possible invitations:

- “My refrigerator just beat your sniper. Your move.”
- “I won with 2% hull. Please do not improve on this.”
- “Same seed. One EMP. Completely different ending.”

These work only when the actual game can produce the claim. Do not manufacture near misses, results, social proof or research milestones.

## Launch experiments, in order

**First: watch five people play.** Have them explain what their bot is doing and why they won or lost. Time to first fight, whether they use a command, and whether they choose a rematch matter more than aesthetic praise. This needs human feedback; automated tests cannot establish fun.

**Then: a small invitation-only weekend.** Give 20–50 people the same daily boss. Ask for clips and confusing moments. Instrument only the minimum events after deciding a privacy approach: visit, battle-start, first-command, finish, rematch, challenge-share click, incoming challenge, research opt-in. Do not send telemetry by default in this alpha.

**Working targets, not industry benchmarks:** first match within 30 seconds; most starters finish; at least one in three finishers voluntarily rematches; some shares actually become completed battles. Adjust targets after observing real players.

**Then: creator rivalries.** Give small robotics, coding, math and strategy creators a custom boss they can genuinely beat. Build a weekly highlight ritual around one hilarious loss and one clever tactic. A “world first” headline must wait for verified evidence.

## How real AI agents enter

Keep two named categories:

- **Garage League:** this cheap deterministic tactics game. Built-in rule-based bots now; later agents can generate valid, bounded strategies or choose a small action set.
- **Proof League:** actual repository-repair, tool-planning or evidence tasks. The runner enforces matched budgets and independent graders. Objective test successes generate corresponding combat effects. The visualization must not imply that an invented punch measures intelligence.

Do not call an expensive LLM on every animation frame. A model can select a tactical policy once, or act at explicit decision windows. The server validates its JSON action schema and executes it in an isolated runner. Render physics locally at 30–60 fps. Never put a provider API key in browser JavaScript, a WordPress page, a replay or a share link.

Before public ranking: authoritative server simulation, version-pinned policies, signed manifests, hidden test seeds, abuse controls, replay/event storage, repeated evaluation and uncertainty. The existing [battle protocol](BATTLE_PROTOCOL.md) remains the starting point. Challenge URLs in this alpha are unsigned local game records and must never be treated as proof of model ranking.

## The research story

“We reserve a defined fraction of our compute spending for verified open research” can be meaningful. “Every punch helps discover a room-temperature superconductor” would be unsupported.

Start with one tightly specified task, an independent verifier, researcher-defined usefulness, public receipts, and a fixed spending limit. Research work should earn an optional contribution badge, never combat power. A failed search is evidence of work, not automatically evidence of scientific progress.

[Detailed proposal and implemented local experiment](RESEARCH.md).

## Keep the infrastructure small

- Keep WordPress for news, community pages and search-friendly material. Embed this static game with the included plugin, or serve it under `/play/`.
- Use a separate service for future matchmaking, replay signing, model API calls and compute queues. Shared WordPress hosting is not the place to run arbitrary agents or long scientific jobs.
- Start with one region, one match runner, capped queues, a daily dollar ceiling and an emergency stop. Cache deterministic replays. Rate-limit by both participant and workload.
- Measure server spend per completed match before opening unrestricted model access.
- Optional cosmetics, memberships and transparent sponsored research missions are plausible revenue experiments. Do not sell competitive advantage, research outcomes or speculative tokens.

## Delivery gates

1. **Local alpha — implemented:** real playable loop, bounded engine, local replay/sharing, research proof-of-concept, static and WordPress packaging.
2. **Hosted playtest:** install on a private/staging WordPress page; verify MIME types, theme width, iframe resizing, touch controls, dialogs, downloads and Cloudflare behavior. No domain migration required.
3. **Public social loop:** short challenge links, verified replay service, privacy-conscious analytics, feedback collection, community moderation.
4. **Proof League:** one objective agent task, one audited runner, two adapters, reproducible grading and spending limits.
5. **Research pool:** a task owner agrees that the outputs are useful; independently checked results reach their destination; budget allocation is reconciled against invoices before any “10% funded” claim.

## Sources behind the practical constraints

- [HostGator File Manager](https://www.hostgator.com/help/article/file-manager-getting-started) documents static-file upload locations.
- [WordPress plugin installation](https://wordpress.org/documentation/article/manage-plugins/) supports ZIP installation alongside existing themes/pages.
- [Quantum Moves 2 research](https://arxiv.org/abs/2004.03296) offers a precedent for a game designed around a specific scientific optimization task; it does not establish that arbitrary gameplay computation is useful.