A Civilization-style strategy game about governing Israel from 15 May 1948. Real ministers, real map, every fact sourced. Unity + deterministic C# sim.
M3 release: the browser build
Status: M3 ships as the local browser build. The itch.io upload and the cross-upload save test are deferred to the start of M4, as a private Draft (owner ruling, 2026-10-10); their sections below are kept, marked "Deferred to M4".
What ships
- Folder and zip:
unity/Builds/Web-gzip/uses gzip compression with no decompression fallback (it was Brotli until the owner's decision on #879, 17 October 2026: "gzip for both";docs/RELEASING.md, step 5, says why). It is zipped asunity/Builds/statecraft-m3-web.zip, withindex.htmlat the zip root. Both are build output and are not committed. - File names: the files under
Build/are named by their content hash (issue #437), for exampleBuild/814d5f9032ce4b5847cf75e492700bbb.loader.js, so a browser that cached an earlier build cannot mix its files into this one. The build'sindex.htmlnames them; nothing else may assume a name.index.htmlkeeps its fixed name and a browser may show a cached copy of it (no host this project uses is known to forbid that), so the page asks the server for its current copy (fetchwithcache: "reload", which also stores that copy for the next visit) and loads the files that copy names. The StreamingAssets files (content/,map/,video/and the start screen's recording) keep their names; the player adds the build's id to every URL it fetches them by (?b=<Application.buildGUID>,GameInputs.WithBuildId), so a browser cannot serve one build's cached content to another. - Built by:
"$UNITY" -batchmode -CI false -GITHUB_ACTIONS false -TF_BUILD false -quit -projectPath unity -buildTarget WebGL -executeMethod Statecraft.Game.EditorTools.Build.Web -compression gzip -shareable -logFile "$PWD/unity-build-web-gz.log", thenpython3 tools/package_release.py(it zips the folder,index.htmlat the root, and refuses while any sensitive event is not sensitivity-reviewed, and when the build'sevent-photos.jsonis missing, says the build was not-shareable, or lists an unreviewed sensitive event photograph). Never zip by hand. - After building: drop the build's rewrites of the settings with
git checkout -- unity/Assets/Settings unity/ProjectSettings/ProjectSettings.asset(as indev-setup.md), so no build-cache change gets committed. - Verified locally by:
uv run tools/web_check.py unity/Builds/Web-gzip --query "golden&smoke=5"with the three full--expectvalues pinned indocs/determinism.md(both goldens and the smoke line). This runs both golden hashes in the WebGL IL2CPP player, then plays five turns through the UI's own buttons (on each turn it presses the first option of every event card shown, while Attention allows in rulesets that have it,GameRoot.Smoke).- Result at the merge candidate (2026-10-11, commit a099655, Brotli build):
GOLDEN OK 1950 71e9904d…,GOLDEN OK 1952 73590e32…,SMOKE OK 5 turns 1948-06-19 d3fdd459…(all three matched in full), then PASS. - History: 2026-10-10 at d450566 (Brotli, before the 1948 war-period events):
GOLDEN OK 1950 ee5d0826…,GOLDEN OK 1952 91bb6fb5…,SMOKE OK 5 turns 1948-06-19 7004c29a…, PASS; folder 23 MB in 36 files, zip 20,862,385 bytes. After the 1948 events (playtest 1 E, Disabled build): the current hashes, PASS.
- Result at the merge candidate (2026-10-11, commit a099655, Brotli build):
- Size (merge candidate, a099655):
- The folder is 23 MB (
du -sh) in 51 files. itch.io allows 1,000 files. - The zip is 20 MB (20,941,435 bytes).
StreamingAssets/map/heightmap.r16(8 MB) was served as is, outside Unity's compression. Issue #44 packed it (below).
- The folder is 23 MB (
- Smaller download: the packed heightmap (issue #44, 2026-10-10). StreamingAssets files are served as they are, so
the raw 16-bit grid cost 8,396,802 bytes on every load. It is now
heightmap.r16z(4,062,705 bytes): the grid's row deltas split into a low-byte and a high-byte plane, in one zlib stream (tools/heightmap_codec.pywrites it,HeightmapCodec.csdecodes it withSystem.IO.Compression.DeflateStream, which IL2CPP backs with its bundled zlib on WebGL too, then checks the size and the Adler-32). It decodes to the old file byte for byte:MapTestsandtest_heightmap_codec.pypin the SHA-256 of the raw file (b6882a77…), and the terrain tests, goldens and smoke are unchanged.- Options weighed (same grid, bytes): raw 8,396,802; zlib of the raw grid 6,284,586; Brotli 11 of the raw grid
(what moving it into the
.datafile would give: Unity's Brotli build matches quality 11, 10,988,872 against 10,953,925 for the same.data) 4,982,506; row deltas in byte planes + zlib (chosen) 4,062,705; the same planes under Brotli 11 3,822,573. Moving it into.datasaves less (3.4 MB on the wire against 4.3 MB; 1.3 MB of zip against 2.2 MB), puts the bytes in the file the engine must load before it starts (the StreamingAssets fetch runs beside the content fetches), grows the uncompressed build used for local and CI checks by 8 MB, and moves the file every map tool reads. The Brotli-planes variant would save 0.24 MB more but needs the same decoder plus all of that. - Size, fb31ce1 -> issue #44 (Brotli build, measured on the folders; the zip is
tools/package_release.py's deflate zip of the same files, made without the release gate): folder 28,310,337 -> 23,954,695 bytes (-4.36 MB, -15%; this is what a browser downloads), zip 25,782,074 -> 23,535,021 bytes (-2.25 MB, -9%; the old zip already deflated the raw grid to about 6.3 MB), 51 files both. The wasm grows by 14 KB (.wasm.br8,622,716 -> 8,636,749: DeflateStream). The zip has grown since a099655 (20.9 MB) with visual v1's baked textures. - Boot (
uv run tools/headed.py boot --runs 3, Disabled build, fb31ce1 -> #44; load average 6-12 from other sessions):--cpu 1:BOOT map_build_ms median=432->480,BOOT first_frame_ms median=3580->3646;--cpu 4:map_build_ms median=1897->2035,first_frame_ms median=7320->7401. The decode is the "heightmap decoded" phase: 7 -> 41 ms at--cpu 1, 34 -> 184 ms at--cpu 4. Every span stays within the visual v1 boot budget (Task 1 baselines 443 / 3547 ms at--cpu 1and 1813 / 6963 ms at--cpu 4, plus max(300 ms, 50%)); the tool exits 0 on each. Served from localhost the download is free, so these runs show only the decode's cost; on a real connection 4.3 MB less to fetch is worth far more (3.5 s at 10 Mbit/s). - Goldens:
uv run tools/web_check.py unity/Builds/Web-disabled --query "golden&smoke=5"with the three--expectvalues fromdocs/determinism.md: PASS on both builds (the Brotli build too).
- Options weighed (same grid, bytes): raw 8,396,802; zlib of the raw grid 6,284,586; Brotli 11 of the raw grid
(what moving it into the
- Play it locally: serve the folder with web_check's handler, which sends
Content-Encoding: gzipfor*.gz(andbrfor*.br). Plainpython3 -m http.serverdoes not, so a compressed build would not load. Then open http://localhost:8000/.
The uncompressed build also works withuv run --python 3.12 python -c "import sys,functools,http.server; sys.path.insert(0,'tools'); import web_check as w; http.server.ThreadingHTTPServer(('127.0.0.1',8000), functools.partial(w.Handler, directory='unity/Builds/Web-gzip')).serve_forever()"python3 -m http.server 8000 --bind 127.0.0.1 --directory unity/Builds/Web-disabled. - Console lines a browser check can rely on:
GAME OK <date> turn <n> hash <state hash>at start (a new game, or the autosave continued).TURN <n> <date> hash <state hash>after every End Turn.CAMERA distance=<m> yaw=<deg>with?camlog.PROFILE ...with?profile.QUALITY level=<n> scale=<s> tiltshift=<on|off> towns=<close|far> p95_ms=<ms>whenever the runtime quality fallback steps (WebGL only; off with?nofallback; docs/visual-v1.md "Runtime quality fallback").BOOT phase=<name> at=<ms> ms dt=<ms> msfor each boot phase with?bootlog(playtest 2, P2-1).NAV f=<frame> ...with?navlog: every raw mouse event, UI pointer and submit event, screen toggle and map click, with its frame (playtest 2, P2-4).
- Player-facing wording: the Budget screen gives each ministry's multiplier as "spending effect 150%" (the share's effect on that ministry's effects, on the 0-100 scale the rest of the game uses), never "x1.50" (playtest 2, P2-6).
WebGL terrain profile
Results from ?profile: 300 frames per setting after 60 warm-up frames, from the diorama camera over central Israel. All runs were in Chrome for Testing 1228 with --disable-frame-rate-limit --disable-gpu-vsync.
heightmapPixelError |
headless (web_check.py) avg / p95 ms |
headed, Metal (Apple M5 Pro, 1728x858) avg / p95 ms |
|---|---|---|
| 1 (shipped; keeps the hex overlay draped) | 4.92 / 7.00 | 5.38 / 7.00 |
| 5 (Unity's default) | 1.51 / 2.00 | 2.19 / 4.00 |
Decision: keep 1. The p95 is 7 ms, far under the 33 ms of 30 fps (M2 ruling K2: above 1, the hex overlay sinks into valleys).
Checks in a real browser (headed Chrome for Testing, ANGLE Metal), 2026-10-10
These were run with a small Chrome DevTools Protocol driver: uv run --with websockets==15.0.1, with a temporary profile and a fixed origin, http://localhost:8765.
Rendering
- The terrain, hex overlay, settlement markers and labels, all UI Toolkit panels and the minimap render. Nothing is magenta.
- WebGL reports
ANGLE (Apple, ANGLE Metal Renderer: Apple M5 Pro). - Three hidden URP utility shaders (
Hidden/CoreSRP/CoreCopy,Hidden/Universal Render Pipeline/StencilDitherMaskSeed,Hidden/Universal/HDRDebugView) need shader target 4.5, which WebGL 2 (ES 3.0) lacks, so the browser reported them "not supported on this GPU". URP loads them up front (Blitter, UniversalRenderer, DebugHandler), but this build never draws with them: the render-graph copy needs framebuffer-fetch emulation, which Core RP disables on WebGL; stencil LOD cross-fade is off (Blue Noise, no LOD groups); the HDR debug view belongs to the Rendering Debugger.Assets/Editor/WebShaderStripping.csstrips them from WebGL builds only (playtest 1, item 12). The console no longer shows the three errors, and the goldens and the 5-turn smoke give the same hashes as before. - The only 404 is
/favicon.ico.
Wheel zoom
- One wheel notch (Chrome sends 100 px) zooms by 10%: 120000 m to 108000 m to 97200 m.
- A 300 px burst zooms by three steps.
- The wheel over a panel does not zoom.
- Before this release a notch moved the camera only 0.08%. The Input System's default scroll behaviour already normalises a notch to 1, and the controller divided by 120 again; that division is now removed.
Saves
Each step below passed:
| Step | Result |
|---|---|
| (a) Two turns, then a hard reload | Continued on the same turn and hash (turn 2, 1948-05-29). |
(b) Tab closed immediately after End Turn's TURN line, then a new tab |
Continued on turn 3. |
| (b) Whole browser quit immediately after a turn, then relaunched on the same profile | Continued on turn 4. |
(c, local) Same origin, a second upload directory (/upload-b/, as an itch.io re-upload does) |
Continued on turn 4. |
| (d) IndexedDB | Database /idbfs, store FILE_DATA, keys /idbfs/statecraft-saves and /idbfs/statecraft-saves/autosave.txt. No JS_FileSystem_Sync deprecation message. |
(e) A private browser context (Target.createBrowserContext) |
The game starts (a new game) and plays a turn. A second private context starts a new game again; no persistence is expected there. |
Firefox is not installed on the development Mac. A Firefox result from a contributor, and the WebKit check, are in the next section.
Firefox cached-build return (#437), 19 October 2026
Local pass, with limits: installed Firefox 158.0 on macOS 27.0.1 arm64, headless through
WebDriver BiDi,
1280x720 at DPR 1. This is Firefox itself, rather than Playwright's patched Firefox in the older report below.
Both consecutive builds used upstream main 9e57134201f74412bf3d7513e26b600a9c760f5c, Unity 6000.3.25f1,
Build.Web -compression disabled, separate output directories, and no source or template change.
The server sent Cache-Control: public, max-age=3600; web_check.py's no-store server was not used for this
cache test. One Firefox process, temporary profile, origin and URL were kept for A, a normal return to A, and a
normal return after replacing the served directory with B. Navigation went through about:blank, without a hard
reload or cache clearing. After the switch the server exposed B only, not A's removed files.
| Observation | Fresh A | Normal return to A | Normal return to B |
|---|---|---|---|
Navigation transferSize (bytes) |
4701 | 0 | 0 |
BiDi responses marked fromCache |
0 | 14 | 7 |
| Inline document's original build names | A | A | A |
| Runtime configuration and StreamingAssets build id | A | A | B |
| Golden manifest values / UI smoke | 8/8 and 5 turns | 8/8 and 5 turns | 8/8 and 5 turns |
A's data file was 20648dff2e2641ced75ed42ef260fd23.data; B's was
ab61a67d11f2a3077b1cb2b5f4eec503.data. The loader, framework and wasm names were unchanged between the
builds. StreamingAssets' b value changed from bcbd32b8194e467bb1906085e78bb3f2 to
020ffe9f466f45e7ad95f62deb9359d0. On the B return, the cached document still contained A's original inline
file names, while useCurrentPage selected B's data URL; B's data request returned 200, and no A-only build
resource was requested from the B server. This exercises stale HTML, not just a fresh page load.
All three visits matched every value in the golden manifest
and logged SMOKE OK 5 turns 1948-10-15 3005405d64abc6e7418fa1f0d35c36ce5f64ed3af7c5a4c24a73d5dfdb16121f.
The harness exited 0; across 33 server requests there were no HTTP errors, no BiDi console error entries and no
GOLDEN FAIL, SMOKE FAIL, BOOT FAIL, exception or recursion failure. Network traces, build-file SHA256
manifests, console lines and 1280x720 captures were retained locally; the exact standalone harness and commands
are included in the contribution's PR description. The B-return capture SHA256 is
daf943e3b3286c109e745648f086b3d5a475329a4f182b32a34dce7076723331.
This checks the existing #501 fix on a local, uncompressed EN build. It does not check Firefox on Windows, wheel or touch input, compressed hosting, save migration or two itch.io uploads. The itch.io check remains the owner's step in #43; #437 stays open for that step. No gameplay, source-review status or release gate changed.
Other browsers (#43), 2026-10-14
Build: main 04bf7cb, WebGL, -compression disabled, served with web_check's handler on http://localhost. Mac: Apple
M5 Pro, macOS 27.0.1. Chromium ran through tools/web_check.py as it is. WebKit and Firefox are Playwright's builds
(Playwright 1.64.0, from ~/Library/Caches/ms-playwright): Playwright's WebKit is the engine Safari uses, not Safari
itself, and its Firefox is a patched Firefox. Chromium in the headed run is the same Chrome for Testing, driven by
Playwright. The headed run is scripted: it starts a new game, zooms with the wheel, orbits with Q, plays four
turns answering each event with its first option, opens Cabinet, Budget, Defence and Why? and closes each with Escape,
turns on the First Truce card's Illustration, then reloads the page and restarts the browser on the same profile.
| Chrome for Testing 149.0.7827.55 | WebKit 27.2 (Playwright) | Firefox 157.0 (Playwright) | |
|---|---|---|---|
golden&smoke=5 |
PASS: both goldens and the smoke value from docs/determinism.md. 11 s for the whole web_check run; in a Playwright run of the same check the goldens arrive 5.5 s after the page opens, the smoke line at 9.3 s |
PASS, headless. Goldens at 10.3 s, smoke at 19.3 s | Not run: the browser does not start on macOS 27 (below) |
| WebGL renderer | ANGLE Metal, Apple M5 Pro | "Apple GPU" | - |
| Start screen, map, HUD, panels, fonts | Correct | Same as Chromium; the screenshots match apart from the small difference in camera angle left by the Q orbit | - |
| Hebrew lines on the start screen | Right to left, correct | Same (Unity draws the text, so the browser does not lay it out) | - |
| Wheel, one notch (100 px) | 120000 m to 108000 m (10%) | Same | - |
| Wheel over a panel | No zoom | No zoom | - |
| Q held 0.6 s | Orbits to yaw 323.5 | Orbits to yaw 323.7 | - |
| Mouse clicks: New Game, End Turn, options, the four screens, Continue | All work | All work | - |
Event card video (first_truce.mp4) |
VIDEO playing, the clip shows |
VIDEO playing, the clip shows |
- |
| Audio | The game has no sound (the card clips are silent), so there is nothing to play. Chrome warns "The AudioContext was not allowed to start" until the first click | No warning | - |
| Save: four turns, then reload, then Continue | Same date, turn and hash (turn 4, 1948-06-12) | Same | - |
| Save: browser closed and restarted on the same profile | Same date, turn and hash | Same | - |
| IndexedDB databases | /idbfs, UnityCache |
/idbfs, UnityCache |
- |
| Console warnings | Six "WebGL: INVALID_ENUM: getInternalformatParameter" per load; the AudioContext warning above | One "WebGL: non-portable extension enabled: WEBGL_polygon_mode" per load | - |
| Console errors | None | None | - |
Contributor result (Chaiky3, in #43): Firefox 157.0, Windows 11 at 125% scale, main fa10633: pass. It loads and
renders, the start screen's Hebrew shows, a touchpad zoom step is 10%, and Continue after a reload returns to the same
turn. Without a hard refresh Firefox mixed cached files from an older build ("too much recursion"); that is #437.
Differences and notes:
Firefox does not start here. Playwright's Firefox 157.0 build exits with "Could not find profile folder" headless, and fails to start headed, on macOS 27.0.1, so this Mac has no Firefox result. The Windows result above is the Firefox evidence so far.
Wheel in lines: no bug (T41). The first run had three lines (
deltaMode1, sent to the canvas by script) zoom 0.16% in Chrome and 0.32% in WebKit instead of 10%. That came from the order of the test, not from the line mode. Unity's web player hands the browser's rawdeltaYanddeltaModeto its engine code (the wheel handler in*.framework.jscopies both), and the engine counts the smallest wheel delta it has seen since the page loaded as one notch, whatever the mode. C# sees only the result inMouse.scroll; the Input System package only turns the engine's scaling on or off (scrollDeltaBehavior, left atUniformAcrossAllPlatforms). Measured on the same build, each case from a fresh page, events sent to the canvas by script, the same in Chrome and WebKit:Wheel events, in order Zoom from each, in notches (one notch is 10%) 3 lines 1 1 line, or 1 page, or 3 px 1 3 lines, 3 lines (also 10 s apart) 1, 1 100 px, 3 lines 1, 0.03 3 lines, 100 px 1, 3 (33, clamped) 100 px, 50 px, 100 px 1, 0.5, 2 100 px, 200 px, 100 px 1, 2, 1 0.5 px, 4 px, 100 px 1, 3, 3 (8 and 200, clamped) So a mouse that always sends three lines a notch zooms 10% a notch, as one that sends 100 px does. The earlier run sent a 100 px notch first (in Chrome, Playwright's 100 px notch counts as 200: a 100 px event after it zooms half a notch), so the three lines after it counted as 3/100 (3/200 in Chrome). Nothing in the game's code was changed for this.
What the measurements did show is a different risk: one small delta sets the notch for the rest of the session. Fixed (T45). A scripted swipe of 31 events, 0.5 to 8 px about 16 ms apart (108 px in all, a trackpad's ramp and momentum), zoomed 95% (28 notches, three a frame, stopped by the 6 km limit) in Chrome and WebKit, on its own or after a mouse notch; after it, one 100 px notch out zoomed three notches. The camera now counts the browser's wheel events itself (
WheelInput.jslibqueues each event'sdeltaYanddeltaMode, andCameraController.WheelEventNotchesscales them):- a pixel event of 50 px or more is one notch, whatever its size. A mouse sends one event a notch, but its size
varies: 100 px in Chrome and Edge on Windows, more with display scaling (125 or 150 px at 125% or 150%, if Chrome
scales it; not measured), less at a page zoom above 100% (80 px at 125%), and 200 px from Chrome's synthetic
wheel (Playwright and CDP send 100 and the page sees 200, at
devicePixelRatio1). The threshold is half of Chrome's 100 px notch, so a page zoom up to 200% still gives a notch; - a whole number of macOS mouse ticks (4.000244140625 px, what Chrome and Safari send on a Mac; an accelerated tick is several ticks' worth) is one notch;
- three lines, or one page, is one notch;
- smaller pixel events, as touchpads send, count 100 px to the notch. The measured swipe's events are 0.5 to 8 px. There is no documented ceiling for a real trackpad's events, and a fast flick can send events of tens of pixels or more; one of 50 px or more counts as one notch, so an event between 50 and 100 px zooms up to twice what its length would, and a bigger one less. Each frame is still clamped to three notches.
The same swipe now zooms 1.08 notches (10.8%) over 30 or 31 frames, never more than 0.15 in a frame, and a notch of 100, 150 or 200 px or three lines, before or after it, zooms 10%.
uv run tools/headed.py wheelrepeats these cases in Chrome, andpytest tools/tests/test_headed.py --headedchecks them, including Chrome's own 100 px notch.Firefox 88 and later report a mouse notch in lines only to a script that reads
deltaModebeforedeltaY, and in pixels (three line heights) otherwise; Unity's own handler readsdeltaYfirst. So the listener sits onwindowin the capture phase, ahead of Unity's handler, and reads the mode first; Firefox then keeps the notch in lines for both. That listener is what handles Firefox; the 50 px rule is only a backstop in case a notch still arrives in pixels. Not measured: Playwright's Firefox does not start on this Mac (headless: "Failed to launch the browser process"; headed and with a persistent profile: no window within 60 s). Were the notch to arrive in pixels after all, three line heights of 50 px or more would still be one notch. With a default font of 14 px or less, three line heights come to about 48 px, so such a notch would zoom about 0.48 of a step: too little, never too much.Known limits: a mouse notch below 50 px that is not a whole macOS tick counts in proportion, for instance a Mac mouse tick at 110% page zoom (3.64 px, 3.6%), a notch at a page zoom of 200% or more, or a Firefox notch in pixels with a default font of 14 px or less (about 48 px, 0.48 of a step). Still to be tried by hand: a Mac trackpad (including fast flicks) and a Mac mouse in Chrome and Safari, a mouse in Firefox (Windows, Linux, Mac), Chrome or Edge on Windows at 125% and 150% display scale (one notch), and a mouse on Linux.
- a pixel event of 50 px or more is one notch, whatever its size. A mouse sends one event a notch, but its size
varies: 100 px in Chrome and Edge on Windows, more with display scaling (125 or 150 px at 125% or 150%, if Chrome
scales it; not measured), less at a page zoom above 100% (80 px at 125%), and 200 px from Chrome's synthetic
wheel (Playwright and CDP send 100 and the page sees 200, at
Aborted requests in Chrome. At each start Chrome lists about 30 of the 51 content files (and the map files) as
net::ERR_ABORTEDin its network log. The files arrive whole: both browsers log the sameBOOT OK contenthash and the goldens pass. WebKit lists none. Not a console error, so it is left as it is.Known warnings, nothing to fix. On every load Chrome logs six "WebGL: INVALID_ENUM: getInternalformatParameter" warnings and, until the first click, "The AudioContext was not allowed to start"; WebKit logs one "WebGL: non-portable extension enabled: WEBGL_polygon_mode". With the aborted requests above, these are the expected console noise of this build: none is an error, and the game runs the same with them.
Input in these runs is synthetic (Playwright's mouse, wheel and keys), not a physical mouse or touchpad.
Still open: Safari itself (only its engine was run here), Firefox on a Mac, Chrome and Edge on Windows, a real mouse wheel in Firefox, a real touchpad and a Mac mouse (above), and the two-upload itch.io save test below. Each needs the owner's machines or the project's itch.io account.
OWNER STEP: sensitivity review before any shared build
Spec section 5.5 requires a recorded sensitivity review for events with sensitive tags (holocaust, terror, massacre,
war-casualties; and displacement, R-GATE-001). The owner's sign-off for all seven events below is recorded in commit
25d2ef6
(2026-10-10; decision D11 in docs/oss/DECISIONS.md): each event's YAML status: is now
sensitivity-reviewed. For any new sensitive event, the owner records the sign-off only by moving its YAML
status: to sensitivity-reviewed and committing it (the loader accepts that value). Nothing else counts as a sign-off.
Re-review, 2026-10-15. The event re-timing for monthly turns (#577) changed the text of altalena, ten_days,
operation_hiram and bernadotte_assassination; they went back to fact-checked until the owner re-reviewed them
("Approve all 4", reviewer: the owner, recorded on #600 and in D11), and they are sensitivity-reviewed again.
- Gate:
Build.Web ... -shareableandBuild.Mac ... -shareablerefuse to build while any sensitive event is notsensitivity-reviewed, and name them (Statecraft.Sim.ReleaseGate.SensitiveUnreviewed).tools/package_release.pymakes the shareable zip and runs the same check (Statecraft.Cli --release-gate) on the content inside the build, so it printsPACKAGE REFUSED ...and exits 1 while any sensitive event is not sensitivity-reviewed. It also refuses a build folder whoseevent-photos.json(written by every build) is missing, not-shareable, or lists an unreviewed sensitive event photograph (docs/event-images.md). Use-shareableand the script for every build that leaves this machine, including the M4 private Draft once it is shared with anyone. Local dev builds (no flag) are not gated, because web_check and the golden checks need them. - Alternate-timeline content (R-ALT-011, Q-ALT-9): Unity player builds made with
-shareableandStatecraft.Cli --release-gate, and thereforetools/package_release.pyandscripts/oss/release_gate.sh, refuse while any consequence card (kind: consequence), any event withvariantsor any armistice card with anearly_text(R-WAR-008, #1709) has noreviewed_by, naming them (Statecraft.Sim.ReleaseGate.AlternateUnsigned): Unity on itsBUILD REFUSEDline, the CLI on aRELEASE GATE REFUSEDline, and the CLI also prints each asunsigned alternate card: <id>. The sign-off is a person'sreviewed_by, written only by that person. - Pinned set:
FinalWaveTests.TheSensitiveEventsArePinnedlists the twelve ids, andTheOwnerTableListsEverySensitiveEventfails if a sensitive event is missing from this table, so a new one cannot skip the review.
| Event | Tags | What to review |
|---|---|---|
burma_road_complete |
war-casualties |
Objective 3 completion card and selected Road of Valour photograph (#791). The owner approved the photograph (#1226) and the card's text as written on 21 October 2026 (#791); it is sensitivity-reviewed. |
altalena |
security, war-casualties | Jews killed by Jews: 16 Irgun and 3 IDF dead (in the YAML comment, not the card) |
ten_days |
security, war-casualties | Lydda and Ramle are shown as a military operation only (Lydda airport, the Jerusalem road); leaving out the population's flight and expulsion is an editorial choice the notes flag for explicit approval |
operation_yoav |
security, war-casualties | the Hebron-hills decision; military text only |
operation_hiram |
security, war-casualties | military text only; the contested population figures and the 26 Sep cabinet quote are left out |
operation_horev |
security, war-casualties | military text only |
bernadotte_assassination |
security, terror | the terror tag follows the Israeli government's own Prevention of Terrorism Ordinance |
reparations_claim |
holocaust | the claim's framing (the cost of absorbing survivors: 500,000 at $3,000 each); the "refuse" option |
deport_acre |
displacement | Deport's follow-up card for Acre (#801): the in-game consequence and one line of history adapted from he.wiki (the residents kept leaving after the 18 May 1948 surrender, but no expulsion orders were issued and no pressure was put on them to leave), whose footnote is Morris p. 153. The line departs from he.wiki in one place: "after the conquest" becomes "after Acre surrendered on 18 May 1948". The owner approved the card as written, English and Hebrew, on 2 November 2026 ("Approve both cards", #2197); it is sensitivity-reviewed. Deport stays out of play (mvp.deport_in_play: false). |
deport_nazareth |
displacement | Deport's follow-up card for Nazareth (#801): the in-game consequence and one line of history adapted from he.wiki and the IDF account (the residents stayed when the city surrendered in July 1948, in Operation Dekel, despite fighting near the town), whose he.wiki footnote is Morris p. 271. The line departs from he.wiki's "in and around the city" by saying "near the town", because the cited "מבצע דקל" has the IDF enter without a battle. The owner approved the card as written, English and Hebrew, on 2 November 2026 ("Approve both cards", #2197); it is sensitivity-reviewed. Deport stays out of play (mvp.deport_in_play: false). |
degania_defence |
security, war-casualties | the defence of Degania Alef and Degania Bet, 20 May 1948: military text only, no casualty figures, no gun count (the four-or-two record is unresolved, [C] in the YAML); the stopped-vehicle photograph's identification (Syrian per the NLI record, a Universal Carrier per Commons) is flagged for review (#1565) |
reparations_vote |
holocaust | the Knesset vote of 9 January 1952 on direct talks with West Germany (#1564); the opposition named by party, not by leader; the riot of 7 January kept to the YAML comment; the three design costs. The owner approved the card text and options, the Benno Rothenberg photograph as a sensitive image and the three costs on 25 October 2026 ("Approve all three", #1575); it is sensitivity-reviewed. |
degania_defence (#1565) was signed off separately: the owner approved its card text, effects and photograph alt text on 24 October 2026 ("approved", #1568), and it is sensitivity-reviewed.
Map lines (display only; not gated by the release gate). The 1949 armistice lines and zones
(content/map/armistice_lines_1949.geojson, decision D-A) need the same review before a shared build shows them. The owner
records the sign-off by moving each feature's status to sensitivity-reviewed (after fact-checked). The 1948
sub-districts (D-B) are reviewed with them. On 23 October 2026 the owner signed off 11 of the 16 features (#196, #1444;
the owner's review also stands as their fact check); egypt_adl_gaza, egypt_mv1950_demarcation, egypt_mv1950_zone_lines,
jordan_adl_a_latrun and syria_adl were flagged and stay draft.
| Line or zone | Front, from | What to review |
|---|---|---|
egypt_adl_gaza |
Egypt, 24 Feb 1949 | the Gaza-Rafah line, Wadi Hasi to Gaza only (the stretch to the frontier is a gap, shown as nothing); its name |
egypt_el_auja_dmz |
Egypt, 24 Feb 1949 | the El Auja Demilitarized Zone and its name |
lebanon_adl |
Lebanon, 23 Mar 1949 | the line shown as the Mandate frontier it follows; "the international boundary" in its name |
syria_adl |
Syria, 20 Jul 1949 | the line and its name |
syria_dmz_north, syria_dmz_central, syria_dmz_south |
Syria, 20 Jul 1949 | the three parts of the Demilitarized Zone, their names (place names chosen by us: Banias, Tell el Qadi, Dardara, Mishmar HaYarden, Ein Gev, El Hamma; a mix of Arabic forms, Tell el Qadi rather than Dan and El Hamma rather than Hamat Gader, and Hebrew ones) |
egypt_mv1950_demarcation, egypt_mv1950_zone_lines |
Egypt, 22 Feb 1950 | the line as drawn on the signed modus vivendi map and the zone A and C lines; that the label and panel name the modus vivendi apart from the 1949 line |
jordan_adl_iraqi_* (four stages), jordan_adl_a_latrun, jordan_adl_bc_south, jordan_adl_d_arava |
Jordan, 3 Apr 1949 (line A by its deadlines: 8 May, 22 May, 17 Jul 1949) | low accuracy, traced from a low-resolution scan, and labelled so; the Art. VI.3 stage areas as read; Latrun traced along the outer loop; Jerusalem, the no-man's-lands and Mount Scopus left out (gaps) |
| On-screen strings | from 1949 | the map label ("Mandate sub-districts, 1947 · 1949 armistice lines: Egypt (and the 1950 modus vivendi), Lebanon, Jordan (low accuracy), Syria"), the info panel lines ("On the ... · low accuracy, from a low-resolution scan"), the agreement titles in the "Egypt-Israel" style (the agreements' own: "Egyptian-Israeli", "Lebanese-Israeli", "Hashemite Jordan Kingdom-Israel", "Israeli-Syrian"), and the seven Hebrew strings in strings.he.yaml (Hebrew is off until reviewed) |
Upload to itch.io
The steps below are the project's first setup (OWNER STEP: they need the owner's itch.io account). Nightly builds are now
pushed with butler by tools/nightly.py, enabled only after nightly.py savetest passes (owner, 27 October 2026,
"Public: website + itch"). The build must pass the sensitivity gate above (-shareable).
- Sign in at https://itch.io. Go to Dashboard, then Create new project.
- Set the title to "Statecraft: Israel (M3 preview)", Kind of project to HTML, and Classification to Games.
- Under Uploads, upload the zip made by
bash scripts/oss/package_web.sh unity/Builds/Web-gzip unity/Builds/statecraft-m3-web.zip(it puts the licence and attribution files inside, then runstools/package_release.py; seedocs/RELEASING.md) and tick This file will be played in the browser. - Embed options:
- Viewport 1280 × 720. It fits a 1366×768 or 1440×900 laptop page without scrolling, and under the UI's
ScaleWithScreenSize (reference 1920×1080) it lays out exactly like 1920×1080;
UiLayoutTestscovers 1280×720. - Tick Fullscreen button. ("Click to launch in fullscreen" is the alternative launch mode if the embed feels small.)
- Leave SharedArrayBuffer support off; the build uses no threads.
- Viewport 1280 × 720. It fits a 1366×768 or 1440×900 laptop page without scrolling, and under the UI's
ScaleWithScreenSize (reference 1920×1080) it lays out exactly like 1920×1080;
- Visibility & access:
- Choose Draft first.
- Open the page and play a turn. The browser console (DevTools) must show
GAME OKand thenTURN 1. - Then choose Restricted (password) or Public.
- Record the page URL below under "Published at".
No server settings are needed. itch.io detects gzip from the bytes and sends it with Content-Encoding: gzip; it also
detects .br, but its CDN decodes Brotli and re-encodes by content type, so a Brotli .data reached players
uncompressed (checked with curl on two public Unity games, 17 October 2026). That is why the release build is gzip
(#879); see https://itch.io/docs/creators/html5, "Compression".
Published at: (owner fills in)
Saves on itch.io: the fix, and its test
- Problem: on WebGL,
Application.persistentDataPathis/idbfs/<MD5 of the page's directory URL>. itch.io serves every upload from a new URL, so a new upload would hide the saves made with the previous one. - Fix (Task 21): the autosave lives at the fixed path
/idbfs/statecraft-saves/autosave.txt.- The template sets
autoSyncPersistentDataPath: true, so closing the file persists it to IndexedDB. - IndexedDB is per origin, so saves survive re-uploads to the same project and stay separate per browser profile.
- The local checks above passed: reload, tab close, browser restart, and a second directory on the same origin.
- The template sets
- Automated test:
uv run tools/nightly.py savetest --tag <published nightly>runs this plan against a separate itch test page (savetest_itch_targetin the job'sconfig.json) and records its result in the job's evidence file. It pushes to the test page twice, never to the public page, and refuses to run without that key or if it names the public project. It loads each upload's game frame from the upload and build ids ofbutler status(https://html-classic.itch.zone/html/<upload>-<build>/index.html), never from the test page, so that page can stay Restricted (#2056). The owner or an agent may start it, since it pushes only to the Restricted test page (owner, 2 Nov 2026, "2 (a)"). - Test plan (after the first two uploads):
- Use one browser profile throughout, without clearing site data or using private mode. Record browser/version, OS,
project page URL and the actual game-frame URL/origin (not only the outer itch.io page). Upload A as above. Play two
turns and copy the last
TURNline, including date, turn and full state hash. Record the upload/build ID, the git commit the build was made from and the zip's SHA-256. The game does not print its content hash; an identical zip guarantees identical content. - Reload A. Expected:
GAME OKgives the same date, turn and state hash as that lastTURNline. Wait for the game to finish loading before reading it. Do not press New Game or End Turn after taking the baseline. - Upload B to the same project, replacing A, with unchanged game content and save schema for this persistence control. Re-uploading the same build is enough to isolate the changed upload URL. Record both upload/build IDs. Confirm B has a different upload URL but the same actual game-frame origin. If its origin differs, record that separately: this fixed-folder test cannot prove persistence between origins, because IndexedDB is origin-scoped.
- First open B with a normal load, not a hard reload (issue #437: a returning player's browser has A cached). It must
start and print
GAME OK, and DevTools' Network panel must show theBuild/file names that B'sindex.htmlnames, not A's. Record the result, and anyBOOT FAILline, as a separate row. Then open B with a hard reload. Pass only if date, turn and full state hash match the baseline. Copy the completeGAME OKline, the oldTURNline and any load warning/error. A date/turn match alone is not a pass. Reload B once more and check the same tuple, so a one-time replay is not mistaken for a durable save. - If B starts again on 15 May 1948, or the state hash differs with unchanged content, record fail with both build IDs, browser/OS and frame origins. Do not reset or overwrite the save before capturing the evidence. Follow #43's fallback: if the save does not survive re-upload, save export/import becomes the next issue. No unrelated account uploads.
- Record the result below. This plan is not a completed itch.io test or a Firefox/Safari/Windows runtime check.
- Use one browser profile throughout, without clearing site data or using private mode. Record browser/version, OS,
project page URL and the actual game-frame URL/origin (not only the outer itch.io page). Upload A as above. Play two
turns and copy the last
Separate changed-content replay check: after the unchanged-content control, a build with intentional content changes may show "This save is from an earlier version of the game, so events may unfold a little differently." Record that warning and both builds' git commits. Matching date/turn in this case is not evidence of exact state preservation, and it does not replace the unchanged-content persistence control above. A changed state hash may be a replay effect rather than a storage failure.
Result: not run on itch.io yet.
| Evidence | Result |
|---|---|
| Browser/version | pending |
| OS | pending |
| Project page URL | pending |
| Upload A/build ID and game-frame URL | pending |
| Upload B/build ID and game-frame URL | pending |
| Game-frame origin A / B | pending |
| Git commit and zip SHA-256 A / B | pending |
| Date / turn before and after | pending |
| State hash before | pending |
| State hash after | pending |
| Full TURN / GAME OK lines and warnings | pending |
| Reload B result | pending |
B opened without a hard reload: GAME OK, and B's Build/ names (issue #437) |
pending |
| Pass / fail and fallback issue if needed | pending |
Review basis: unity/Assets/Scripts/SaveStore.cs fixes the WebGL directory; sim/Statecraft.Sim/SaveGame.cs records the
content hash and command log; unity/Assets/Scripts/UI/GameRoot.cs logs the current state hash; SaveUiTests.cs covers
changed page-directory paths and changed-content warnings. Those code/tests support the plan, not an itch.io runtime pass.
Known limits of this build
- Placeholder settlement markers; per-era art is M4.
- No Diplomacy screen or legacy report (M4).
- English only.
- The headless smoke plays five turns; a full-slice smoke is M4.
- The wheel-zoom step (one notch = 10%) was measured in Chrome on macOS only. Firefox reports wheel deltas in lines and Windows at 125% scaling reports 125 px per notch, so re-check the step in the owner's M4 itch.io check.
- Nightlies upload to itch.io once the cross-upload save test passes (owner, 27 October 2026; until then this build is
not uploaded there). The 12 sensitive events (pinned in
FinalWaveTests) aresensitivity-reviewed; the owner's sign-off of 10 October 2026 for seven of them is recorded in commit 25d2ef6 (see above).