[ashtonanderson]

On a real iPhone 12, the analysis page survives repeated visits and reloads that make Safari give up on today's code. After a move, the four lines settle about 0.7 s sooner (about 3.4 → 2.7 s) and the best line reaches depth 18 about 2.4 s sooner. The chart comes 0.1–0.2 s later. A first visit no longer downloads 78 MB of nets, and Stockfish uses about 5× less memory.

The depth label now means something different, so its timings aren't like for like. Today, "d18" appears when four lines have reached depth 18. Here it appears when the best line alone has, with the other lines still shallower. The fair comparison of the whole display is "lines settled".

Draft: awaiting a decision on merging.

What changes

1. Stockfish 19's small net replaces Stockfish 17. It uses one 1.2 MB net served with the site, instead of two nets totalling 78 MB. The net download code and IndexedDB cache are gone, and old nets are removed from returning visitors' storage.
2. The best line goes deep first. The search now runs: every move to depth 6, the best line to the target depth, Maia's moves to depth 14, the other three lines one at a time to the target, every other move to depth 10, then Maia's moves to the target.
3. A position isn't done until the search finishes. Updates carry partial until the last one. Revisits, full-game analysis, Learn From Mistakes and drills all respect it, and unfinished results aren't saved.
4. Less memory reserved. Stockfish reserves 134 MB and Maia 268 MB, and both retry when an iPhone has none to spare.

Results

Main at 3e5ef6d against this branch's code: production builds, warm visits, 6 pairs. The phone is emulated, with every thread at a quarter of full speed; it's far slower than an iPhone 12, so use it to compare builds. Stockfish was sooner in 6 of 6 pairs on every row, and the board and Maia's moves are within the noise.

seconds	phone	desktop
first depth	9.02 → 5.02	1.55 → 0.65
depth 12	18.02 → 7.50	2.80 → 0.80
depth 18	29.54 → 9.46	4.87 → 1.25

• Stepping through a game: steps reaching depth 18 go from 13 to 24 of 24 on the phone, and from 30 to 32 of 32 on a desktop.
• After a step, desktop: the lines stop changing at 3.38 s instead of 5.36 s, and the chart is unchanged at 1.20 s.
• First visit, phone on 4G, before the last search change: Stockfish's first depth goes from about 126 s to about 5.6 s, and Maia's moves from about 103 s to about 53 s.
• Renderer memory, desktop, before the memory fix: peak goes from about 1,045 MB to about 555 MB.

On a real iPhone 12

Today's code against the new code, the whole site in Safari, one or two runs each:

	Today's code	New code
6 return visits	page reloaded, both times	all finished, both times
10 reloads of the whole page	Safari gave up after 4	engines back on 10/10
After a step: depth 12	0.77–0.82 s	0.28–0.29 s
After a step: depth 18	3.27–3.37 s	0.87–1.18 s
After a step: chart	1.19–1.34 s	1.39–1.58 s
After a step: lines settled	3.37–3.42 s	2.59–3.12 s
First visit: Stockfish's first depth	22.0 s	3.6 s
First visit: Maia's moves	19.7 s	10.6 s
First visit: board	0.94 s	3.05 s (one load)

The reload tests are harsher than normal browsing, and this is one phone. On today's code the stepping test itself reloaded the page in 3 of 4 tries, so its step rows come from the runs that finished; the new code finished all 6. No real Android phone has been measured. Maia on WebGPU was tested and not pursued: in the one run that gave numbers, it didn't free the CPU for Stockfish.

Trade-offs to review

• Accuracy: at depth 18 the new engine picks moves about as well as Stockfish 17 against deep native searches, but its final best move on game positions is a little worse (0.42 vs 0.30 win points lost per move against a Stockfish 19 reference, 0.49 vs 0.27 against a Stockfish 17 one).
• Move colours: 87.9% of played moves get the same colour from both engines.
• The depth label: "d18" now means the best line got there, and the other lines fill in afterwards.
• The chart: 0.1–0.2 s later after a step on the iPhone, unchanged within noise on ada. It appears 40–60 ms after Maia's last answer, and rendering it at normal priority didn't help, so the likeliest cause is Maia sharing the CPU with Stockfish. That isn't confirmed, because today's code reloaded before giving the same measurement.
• Saved analysis: it will mix Stockfish 17 and 19 results, and unfinished results are no longer saved.
• human-bar: its loadNnueModelShared and second engine need rewriting.

Before merging

• Run the engines and the whole site on an iPhone (public/phone-lab.html, on claude/phone-lab-base and claude/engine-bench, neither for merging).
• Repeat the reload tests on the final search order: both engines back on 10 of 10, in both tests.
• Try the analysis page by hand on an iPhone: real taps and scrolling aren't covered by those tests.

Checks

npm test (74 tests), npx tsc --noEmit and npm run lint pass. An independent code review of the first three code commits found five defects, which are fixed in d671b88 with tests. Full method and tables are in LOADING.md, round 10. This also carries round 9's notes from the closed #333.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
maia-platform-frontend	Ready	Preview	Nov 5, 2026 3:27am UTC