{"challengeChanged":"The challenge changed. Test again before keeping an upgrade. Previous comparisons remain in your history.","upgradeUses":"The upgrade uses today’s terrain costs, or counts steps for a distance goal.","candidateAlways":"This candidate always assumes rough ground is dry. Try changing the weather.","rulePrices":"This rule prices rough ground at five. If it is already active, this is a no-op.","noComplete":"No complete journey. Paint grass over some water to reopen a route.","startAnd":"The start and home stay in place. Paint a different square.","terrainUpdated":"Terrain updated. Both routes have been recalculated.","experimentUpdated":"Experiment updated. The previous result no longer applies.","playbackPaused":"Playback paused. Use the journey slider or continue.","bothRoutes":"Both routes shown. Reduced motion is on.","bothJourneys":"Both journeys finished. Compare the measured costs, then try another condition.","newReproducible":"A new reproducible terrain layout is ready. Paint it to make it yours.","workshopReset":"Workshop reset. Any separately saved challenge is still on this device.","candidateEarns":"Improved on these 13 development cases. You can keep this version in the workshop.","noImprovement":"Valid and non-regressing on these 13 development cases, but unchanged. No improvement to keep.","keepThe":"Keep the current version. This candidate makes at least one challenge worse.","noDecision":"No decision yet. A journey is blocked; restore a route and test again.","upgradeKept":"Version kept for this workshop. Development-case results do not establish an improvement on unseen tasks. You can restore the previous version.","previousVersion":"Previous version restored. The experiment history is preserved.","workingPlayground":"Your working playground is ready to save. Open the HTML file in a browser, even offline.","challengeSaved":"Challenge saved on this device. No comparison result is treated as saved evidence.","savedChallenge":"Saved challenge restored. Run a fresh comparison before promotion.","savedChallengeMore":"Saved challenge removed. Your open experiment is unchanged.","choose":"Choose a challenge JSON file smaller than 100 KB.","challengeImported":"Challenge imported. Previous evaluation claims were not imported.","importFailed":"Import failed; your current challenge is unchanged. ","message26":"(prefers-reduced-motion: reduce)","message27":"The proposed route saves ","message28":"Both routes use the same ","message29":"Column,row coordinates. Current: ","message30":"Both journeys shown at the finish.","message31":"Last paint change undone.","message32":"Preparing your offline playground…","message33":"my-upgrade-playground.html","message34":"my-upgrade-challenge.json","message35":"Your challenge and decision record are ready to save. Imported results will need a fresh comparison.","message36":"No saved challenge on this device yet.","message37":"Could not complete that action. ","roughCostPrefix":"Rough ground costs ","roughCostSuffix":" energy per step.","moreCost":"The proposed route uses ","noComparisons":"No comparisons yet.","originalSteady":"Original steady rule","historyCurrent":" against "}

Compare an AI-assisted routing change

Help a rover
search less.

Change the ground. Can the upgrade examine fewer squares without a costlier route? Compare both on your map.

The proving ground

Paint the landscape

Dry ground

Paint:

Paint with larger controls

Choose a row and column, then apply the selected terrain brush. The highlighted square shows your selection.

Paint rowPaint column

Row 1, column 1: grass.

Tap a square to paint. Water blocks the route. The start and home stay put.

Current versionProposed upgradeRough ground costs 2 energy per step.

Show squares examined

Journey

Ready. The route previews are computed locally; press Compare routes to follow them.

Jonathan directed this workshop; Codex implemented it using established Dijkstra and A* algorithms. Your edits run real searches locally. AI helped build the tool; it does not invent a new program during your visit.

This interactive workshop needs JavaScript. The idea: compare an original rule with a proposed change on the same tasks. The comparison checks route cost first, then squares examined. Both planners use established algorithms. Read about empirical self-improvement research.

Keep what earns its place

Does the upgrade hold up?

Try your map, then compare both programs on the same development cases. Keeping a version changes this workshop; it does not establish performance on unseen tasks.

Every tested journey must finish and none may get worse. Improvement needs at least one better result. Compare route cost first, then squares examined when costs tie. This counts search work, not seconds.

Inspect every comparison

Each cell shows route cost / squares examined. Lower cost takes priority.
Challenge	Current	Proposed	Outcome

Try to break the improvement

Does it survive a different world?

Run 128 scenarios across dry or wet ground and energy or distance goals. The batch includes repeated open-ground controls. Open ground, obstacles, rough terrain and sealed barriers expose different failures. A separate Bellman-Ford implementation checks both answers. It was written by the same authoring team; the data remains development data.

Seeds make runs repeatable. These are development tests, not an untouched holdout or a speed benchmark. Unreachable maps must be identified correctly. Cases sharing one map are related observations.

Reproduce the test outside this website

Download the evaluator and routing model to reproduce a report with Node: import('./stress.mjs').then(m => console.log(m.runStress(42))). Save both files in the same folder; replace 42 with the seed in your report.

Your creation, beyond this page

Take the playground home.

Keep your terrain, rules and a working route explorer in one HTML file. Open it offline, change the weather, and try another idea.

Import a challenge

Optional browser storage. No account or upload.

What did AI actually change?

The first version compared a fixed terrain estimate with the true cost. This revision starts with Dijkstra's exact-cost search. The proposed A* search adds Manhattan distance to its priority and prefers positions closer to home when priorities tie. Each move costs at least one, so this distance never overestimates the remaining cost.

These are established algorithms, not an AI discovery. Codex implemented the comparison, stress generator, reference check and exports under Jonathan's direction. The opening map is a selected example. No human performance benchmark or autonomous self-improvement is claimed.

Development record and exact sources

Your experiment history

Rejected and unchanged ideas stay in the record even when they do not become the active version. The most recent 20 events stay in this visit and in your evidence download.

No comparisons yet.

Follow the routes as text

Run a comparison to inspect the coordinates.

What can these comparisons establish?

This workshop lets you compare and retain a routing change on stated development cases. It does not include an untouched evaluation set. Freeze the candidate, the baseline, the task and the goal. Check the actual result. Once feedback guides another change, those examples are development data, not an untouched final test. Editing this workshop invalidates its previous comparison.

Real self-improvement research can use a model to propose code and an evaluator to check it. Improving that task program is different from improving the process that invents future changes.

File details on this page

Format, size and a short SHA-256 identifier, measured when this page was built.

provenance.jsonJSON, 4.0 KB, sha256 b282cb0516d7