Alice Project
中文
May 11, 2026Historical

V1.0.3 — Robustness & v1.0.2 Close-Out

Pins down the world replay's URL and fixes the issues that kept it from opening, then makes the whole simulation repeatably testable without spending real AI calls — including checkpoint resume for long runs.

Highlights

  • Fixed: the world replay page had several issues that made it fail outright (black screen, crash, 404 links) — all fixed, and its address is now pinned to /alice-project-world-replay/.
  • Innovation: a full in-game day (144 steps) can now be run without a single real AI call — fast, cheap, and stable to test.
  • New: the simulation supports checkpoint resume — a long run can stop midway and restore exactly from a saved point.

Fix: the replay page wouldn't open

After v1.0.2 shipped, opening the world replay in an actual browser surfaced three problems that broke the page outright: a missing endpoint caused a black screen, the replay data was missing fields and crashed the page, and several links 404'd. This release fixes all of them and pins the replay's address to /alice-project-world-replay/. Entering the world replay is now a clean path end to end.

Innovation: tests that cost nothing

Every step of Alice's world simulation calls a large language model, so a real full-day run is slow and expensive. The approach here: record the AI's answer at each step of one real run, then have later tests replay those "recordings" instead of calling the AI for real.

That lets us run a whole day (144 steps) — or much longer — over and over at zero cost, to confirm the system holds up over long runs. It also flushed out and fixed two hidden bugs that only surface under fast, continuous execution.

New: checkpoint resume

A long simulation can now save midway and restore exactly from that point, with results identical to running it in one go.

Technical Notes

This release is mostly engineering hardening; only the system axis advances among the five version axes:

Axisv1.0.2v1.0.3
world (world_id)uva_village/v1.0.2uva_village/v1.0.2
systemv1.0.2v1.0.3
cognitionv1.0.2v1.0.2
replay runtimev1.0.2v1.0.2
replay bundlev1.0.2v1.0.2

So v1.0.2 bundles play in a v1.0.3 runtime unchanged, with no regeneration needed. The "recording" test setup has three parts: a toggleable recorder for AI calls, a mock provider that replays those recordings, and checkpoint resume so long runs can restore from a saved point.

What's Next

V1.1 turns to content polish and product fit-and-finish: lifting more characters from rough sketches into full personas, improving long-run simulation stability, and adding search, a timeline scrubber, and share cards to the replay.

About