Release v1.2.0a · Pr1nted/Open-Doctrines

game 1.2.0a

OpenDoctrines plays inside Discord. The game is an Activity now: launch it
from the rocket button in a voice channel and it runs in the client, with no
download and no install. Whatever is deployed is what everybody in that channel
is playing, the moment it goes up.

The work was mostly in one place, and it is not the obvious one. Inside an
Activity a Content Security Policy allows exactly the hosts named in the app's
URL mappings, so every request the game makes has to be rewritten to go through
Discord's proxy. What is deliberately not rewritten is the account service's
identity: a host compares the issuer named in a join ticket against its own, so
repointing it would have made an Activity player disagree with every host in
existence. The transport is diverted; the identity is not.
You can join a game with the invite code alone. Leave the address blank and
the game goes through the account service's relay: no port to forward, no tunnel
to keep alive, and the host never learns your IP address — so the panel warning
you about that is shown only when you actually type an address.

This is the only way a player in a browser or in Discord can join at all, since
a sandboxed Activity cannot open a socket to somebody's home connection whatever
they paste in. Hosts opt in with Host through the account service on the host
screen, or relay: true in a dedicated server's config.

The trade is stated plainly because it is real: a relayed game needs the account
service for the whole of its life, where a listening host only needs it to start.
A dedicated server could never open a session, and now can. It skipped
Game::init() — correctly, because that opens an OpenGL window — and that is
also where the game's config is read and the account client is set up. So it ran
with no issuer, no token and no credential, and failed at the last step with
"Sign in and register this server before hosting": advice that could not be
followed, because the machine was signed in and this process had simply never
looked.

Two more things were hiding behind that. The failure check ran once,
synchronously, before the asynchronous open could possibly have failed — so a
later refusal was never looked at again and the server sat silent for ever with
no join code and no error. And a server credential that had stopped being valid
was unrecoverable, because registration only ran when the field was empty. All
three are fixed; a failed open now says why, clears a dead credential, and exits
non-zero instead of telling a supervisor it shut down cleanly.
Telling us how long you played, if you want to. A new Advanced setting,
off by default and off after every update. When it is on, the game sends one
message at the end of a session: how long it lasted as one of five ranges, and
whether you were on web, desktop or Android. That is the entire message — no
account, no nickname, no installation id, no device, no address.

Because nothing links two reports, there is no "delete mine" and we do not
pretend to offer one; every report expires by itself after ninety days, and the
whole set can be erased on request. It also means returning players cannot be
counted, which is a real thing given up on purpose: counting them needs an
identifier, and an identifier is the thing being refused. PRIVACY.md and TERMS.md
now say all of this, including the sentence that used to promise no usage
reporting at all.
Paste works in text fields. Ctrl+V and Cmd+V were handled, but the web build
called a desktop clipboard function that returns nothing in a browser — so in
Discord every field silently ignored a paste, including the one you put an invite
code into. The page now listens for the browser's own paste event, which needs no
permission because the keypress is the consent.

Four screens had never called the shared editing code at all, so paste had never
worked in them on any platform: the admin fields, the bug report, the moderator's
note and the account lookup. They keep their own typing rules — the duration
field still refuses anything but a duration — and take the paste through one
shared path.
The caret sits at the end of what you typed. It was worked out after the
wrapping loop, from a line buffer the loop's last pass had just cleared and a y
position it had just advanced — so it sat at the left margin, one line below the
text. Fixed once in the mail composer months ago; the fix did not travel, because
a copied loop does not inherit a later one. Three more copies had it. All four
fields now draw through one function, and the screenshot that photographs that
screen focuses a field so the caret is actually in the picture.
The Android build reports its own version. AndroidManifest.xml carried
versionCode="1" and versionName="1.0.6a" as literals while every other
platform stamped itself from the VERSION file — so the APK announced 1.0.6a
and every build ever made declared version code 1. That is the value every app
store refuses an update on if it does not increase, so the next upload to any
of them would have been rejected. Both are stamped at packaging time now, and the
build fails if a placeholder survives.
The hosted web build ships the game, not just the menu. The deploy script
copied four files and left data/ behind — 61 MB of maps, music, fonts and the
AI model that are served next to the page rather than inside the preload. Every
deploy since hosting was set up put a playable menu online with nothing behind
it. It was invisible because the host answers a missing path with index.html
and status 200, so the game asked for a 1.2 MB map, was handed 10 KB of HTML, and
reported that the download had failed. The deploy now stages data/, ships a
404 page so a missing file fails as one, and checks the live URL afterwards
rather than trusting the upload.
Mail: letters between countries, delivered when the turn resolves. A new
Mail button opens a correspondence with any country you may write to — one
thread each, read like a messenger. Write now, it arrives next turn, and until
the turn resolves it is still on your desk: you can rewrite it or tear it
up. Once delivered it is history and cannot be touched.

That delay is the whole design rather than a limitation. It is the one thing a
chat box cannot offer — the chance to change your mind — and it is what makes
an AI correspondent fair, because a bot that answered instantly would out-talk
every human at the table.

The button appears only when somebody could answer: on a server, or with
the language-model module loaded. In a single-player game with neither, it is
absent rather than greyed out — a greyed control invites you to work out how to
enable it, and there is nothing to enable.

Hosts choose who may write: nobody, players only, advisors only, or everyone.
Players choose who may write to them, and that always wins — a host can
narrow who speaks on their server, but nothing a host permits forces a letter
on somebody who has shut their door. Hosts can also keep a list of words their
server will not carry, which is a blunt instrument and is documented as one.

Map scripts can read game.llm_diplomacy and game.mail_enabled, so a
scenario written around bargaining with a machine can tell.
AI advisors: countries that answer their own mail. An optional module
points the game at a language-model runner — one it installs for you, one you
already run, or a remote API with your key. Countries then write back, in your
own language, and they talk to each other as well as to you.

They are allowed to lie. A diplomat who cannot is not a diplomat, so they
bluff, flatter and promise things their country will not do. What they may
never do is pretend to be human: every letter carries a bot tag, in every
language, on every screen that shows it. The lying is a game mechanic;
concealing the speaker would be a lie about the game itself.

Each correspondence is sealed. The model writing as Britain to France sees
the Britain–France thread and nothing else — not Britain's other letters, not
the world state, not your orders. It is given a coarse word for relative
strength rather than a number, because a correspondent who can quote your army
size is reading the save file and every player can tell.

It writes letters and nothing else. The existing AI still runs the war, the
budget and the diplomacy — that AI is measured and deterministic, and handing a
language model the levers would trade all of that away for something nobody
can bench.

Installing a runner downloads and executes third-party software, so: the
version is pinned in the source, the archive is checked against a
SHA-256 compiled into the game before anything is unpacked, it comes over
HTTPS from the project's own release host, you see the URL and size before a
byte moves, and it lives under the game's data directory so uninstalling is
deleting one folder. An unpinned build refuses to install at all rather
than running something nobody verified.
Reporting a player, and moderation. A message can be reported two ways,
and they do different things: to the server owner, who can remove that
person from their game, or to the account service, which can ban or time
out the account everywhere. An account carrying the developer badge gets a
review queue in-game, with the message, whatever context the reporter
attached, and three outcomes — ban, timeout, or dismiss.

Dismissing is a real outcome. A queue whose only clearing action is a
punishment is a queue that punishes people.

Bans and timeouts are announced to a Discord channel, saying who and until
when and nothing else — never the reported message, never the reporter. A
moderation channel is still a room full of people, and republishing abuse in
order to announce that it was dealt with hands it a second audience. A
dismissal is never announced at all.

Reports are readable only with the developer badge and are deleted after 90
days. The terms now say plainly that player conduct is not ours, and the
privacy policy says exactly what a report stores and why it must.
A public tournaments noticeboard, readable without signing in — a
noticeboard you must log in to read is not one.
An optional, local, self-declared age prompt, off by default and toggled
by the host. It is stored on the player's machine and transmitted nowhere.
It is not described anywhere as satisfying the Online Safety Act, because
self-declaration is not "highly effective age assurance" and saying otherwise
would be worse than not having it.
You can report a bug from inside the game. Escape → Report a problem, or
the Report button in the map editor's toolbar. Pick what it is about — the
interface, the AI, maps and data, multiplayer, map scripting, mods and the
SDK, security, or something else — write a couple of sentences, and send. The
same form sends a suggestion: it is the first choice on it.

The reason it is in the game rather than on a forum is that a bug reported
where it happened arrives with the version, the platform and the state the
game was in, and the same bug reported an hour later arrives as "it crashed
sometimes". That gap is most of the work of fixing it, and it is currently
paid by the person least able to pay it.

Ticking "attach what the game knows about itself" adds the version, the AI
version, the window size and language, the turn, the map's size, which country
you are playing, your difficulty, every installed mod and any errors the map's
scripts reported. "See exactly what" shows that text in full, scrollable,
before anything is sent — there is no summary, because the promise is that you
can read what leaves your machine and a promise about a summary is not that
promise. Save and mod paths are shown as ~/...: on all three desktop
platforms the home directory contains your account name, and a report that
mentions a file would otherwise mention a person.

A security report is never filed as a public issue. It becomes a private
draft security advisory on the repository — readable only by people with admin
access, and the same object a coordinated disclosure is published from. Not a
public issue, and not the ordinary bugs channel either: an unfixed exploit
should not be readable by everyone who can see a chat channel. If it cannot be
filed there it is refused outright and you are told so, rather than being
delivered somewhere less private to make it succeed. The form says all of this
when you pick the category.

Spam is bounded in two places, because a client-side limit is a courtesy
rather than a defence: the game refuses a second report within 45 seconds and
an eleventh in a day, telling you before you type it rather than after; and
the relay applies its own per-address limit, a per-install daily cap, a
six-hour window that swallows the same report sent twice, and size caps
enforced before the body is parsed. Nothing is sent unless you press Send —
there is no background telemetry anywhere in the game. See net/PRIVACY.md.

Reporting is optional infrastructure: with no destinations configured the
endpoint still validates and rate-limits and simply forwards nowhere, which is
the right behaviour for a fork that has not set up a tracker of its own.
Closing a GitHub issue tells the Discord thread it came from. A report
opens a forum post and an issue; closing or reopening the issue now posts
"Closed by ", with a link, into that thread. The pain it removes is
people still arguing in a thread about something fixed a week ago.

One direction, deliberately. GitHub can push events to a URL and Discord
cannot, so the reverse would need a bot holding a gateway connection — a
permanently-connected Durable Object and a second always-on credential, which
is a different project. And the note is posted rather than the thread being
archived, because a webhook can only send: locking a thread needs a bot token
too, for a tidier result rather than a different one.

Setting it up is a step in add-feedback.sh, which generates the shared
secret and creates the webhook.
Reporting needs an account, and reports are signed. Every bug and
suggestion is published — to the community channel, and a bug to the public
issue tracker — so each carries the reporter's nickname, and the form says so
before you type. "Do not show my name" publishes "Anonymous (signed in)"
instead: the service still knows who you are, so the limits and the ban list
are untouched, you have just asked not to be named in public. The byline comes
from the sign-in token and never from anything the game sends, so nobody can
file a report under someone else's name.

The obvious cost is that the reports worth most — from players who cannot get
past the menu — are the ones least able to sign in. That is a real trade and
it was made deliberately.
The game asks once whether you are enjoying it. After 45 minutes, between
turns, in the corner: Rate on itch.io, Something's wrong, or Not
now, which means never. It is not shown again whether or not it is answered,
because the second time of asking gets a worse answer than the first.

It sends you to itch.io rather than collecting five stars in the game, because
a number only the maintainer sees does nothing for anybody, while a rating on
the page is what somebody deciding whether to try the game actually reads.
"Something's wrong" opens the report form — a player who is unhappy is about
to say so somewhere, and a bug report is a better destination than a one-star
review, for them and for the game.

It also closed a hole. A rating needed no account and still posted to the chat
channel, rationed only by an install ID the client invents — an open pipe into
a Discord server. There is now no unauthenticated write path at all.
The pause menu is translated. It measured and drew its four items as raw
English while every other menu in the game goes through T(). Nobody noticed
because the words are short.
Loading a save no longer doubles every doctrine. The load sequence is
unload → map load → save load, and the map load runs applyStartingPolicies,
so both the active-policy list and its per-country index were already full
when the save was read — and the save's policies were APPENDED to them. Every
continued game therefore held two of each doctrine: doubled doctrine upkeep,
and every "reduces unrest" doctrine counted twice, for the player and for
every AI country. The only clear lived in unloadGameData, which runs before
the map load rather than between it and the save.

It was invisible until the country profile started listing doctrines BY NAME:
a cost that is summed silently stays hidden until something enumerates it.
The rest of the load was audited for the same shape — a container the map load
fills and the save then appends to — and the policy pair is the only one; the
others assign or insert into sets, which are idempotent.

And it compounded. Each load added the map's starting set to whatever the
save already held, so a game loaded and saved repeatedly accumulated a copy
per cycle. Measured across the 1,301 saves in data/saves: 48 carry
duplicates, the worst at ×7 — 572 policy entries where 529 are distinct, a
country paying seven times its doctrine upkeep. Stopping the doubling could
not have repaired those, because a save carrying the key is exactly what tells
the loader it holds the whole truth, so duplicates are now collapsed as the
save is read and such a game is repaired the first time it is opened.

The key is the whole tuple — country, policy, target province, target
minority — not the policy id. The same doctrine can legitimately be active
twice against different targets (an ethnic policy applied to two minorities),
and collapsing by id would silently repeal one of them.
Districts can split the budget by size, beside the existing equal split.
It is the neutral setting rather than a convenience: the resolver divides
budget share by ground share, so making them equal gives every district a
factor of exactly 1.0 — the same suppression everywhere, which is what the
country had before it was divided. Equal shares are a different thing and
often the wrong one: on a country cut into a 23-province heartland and a
3-province border strip, "equally" hands the strip eight times the suppression
per province.
A published country lists WHICH doctrines are in force, not just how many.
"3 in force" is a fact about a list rather than the list, and what a reader is
actually being told is what the government has enacted.
The country profile scrolls properly. It was clamped at the top only, so
the wheel could push the whole page off into empty space; and it now runs past
one screen. The district list is capped at six rows with "+N more" — eleven
districts is an ordinary number for a large country and the full list crowded
out everything below it. The Districts tab remains where all of them live.
The country profile shows the division on a map, framed on the country
rather than on the world. A list of names and percentages says how a country
is cut up without saying WHERE, and the same three rows describe a country
split east-to-west and one split between a heartland and an island chain.
Opening Politics → Districts no longer stalls. It cost 126 ms: 68 to build
a province-to-pixels index over all 33 million map pixels — 128 MB, held for
the rest of the session — and 58 more to clear, paint and upload a full-map
RGBA overlay, which it paid again on every stroke of the district brush.

None of that index was needed: the country's own pixels are already indexed at
load, and the only pixels this overlay touches are that country's. Painting
from the country instead, at half resolution (it is drawn into a panel a few
hundred pixels wide, scaled down twentyfold before the zoom), and uploading
only the rectangle the country occupies: 126 ms → 23 ms, measured on the
worst case in the game — a globe-spanning empire with 1.9 million pixels of
ground.
Districts are named the way countries are, and the name follows your
language. They were built as " " with the word translated at
CREATION time — "Українці Територія V", a plural noun jammed against a noun,
frozen into whichever language happened to be loaded, and unchanged when you
switched. Two faults in one string; it read wrong in English too
("Ukrainians Territory").

A district is a region, and the game already knows how to name places: the
canonical English form is stored and rendered per language at draw time
through the same patterns a country name uses. "Eastern Ukraine" is "Східний
Україна" in Ukrainian and "Ost-Ukraine" in German, and it changes the moment
you change language. A name you type yourself is yours and is never
re-rendered.

The compass also replaced the Roman numerals: five directions plus the bare
place is six names before anything has to repeat, and a district that would
have collided now takes a direction instead of a number.

Two things it learned the hard way: a name the string table KNOWS beats a
prettier one it does not ("Britain" transliterates to "Брітаін" — letters,
not a word — where "British Empire" is a real name in every language), and
the numeral has to come off before translation or " II" comes back as "ii".
New scripts start at the latest engine version, and there is a version 3.
A new script was created pinned to #OD/MapEngine/1, so every statement added
since — print, for, try, and everything from this release — was refused in a
file the editor had just made, and the author's first experience of the
language was a lint error on a line the documentation told them to write.
Version 3 covers wait N turns, set x to, {value} interpolation and
foreach district; version 1 and 2 files still get exactly their own
language, which is what pinning is for.
The map editor uses your accent colour. It was a gold constant, so the one
screen in the game that ignored the accent setting was the one people spend
hours in. It follows the setting live, without reopening anything.
Districts in the map editor are not a thing you switch on. They hid behind
"Districts: OFF (0)" beside the paint-mode toggles, which said a country
either has districts or does not. Every country has them — the game builds one
covering everything the moment anybody asks — and the only question is how
many. The list, the count and the controls are always there; the toggle that
remains is about which of the three things a drag on the map paints.
The country profile has a Close button, next to the ESC hint that used to
be the only way out. It is a screen you reach by clicking, from a panel you
close by clicking.
A model file the AI cannot read no longer looks like no model at all.
Every refusal in the loader — wrong magic, corrupt container, a format byte
from another architecture, a pre-trunk file — reached a caller that printed
"Fresh model (no file at X)" for a file that plainly existed. So pointing an
evaluation at a model the loader dislikes did not fail: it played an
untrained network and printed a perfectly ordinary score, and any number
taken that way was a measurement of random weights wearing a trained model's
name. A refusal now says which reason on stderr, and says that the run is
about to play on random weights.

The same fault in the other direction: OD_EVAL_MODEL pointing at a path that
does not exist used to fall through silently to data/ai/model.bin and run a
complete evaluation of a different model. It now stops.
The AI now plays by the player's information rules, and there was exactly
one place it did not. All eight of the features describing a neighbour were
audited against what the game will actually show a player: garrisons appear in
a province panel, industry in the industry view, borders on the map, wars and
alliances on a country, claims on their own screen, and a treasury only when
published. War weariness appeared nowhere — warWearinessOf had a single
caller, the resolver — so no amount of clicking could tell you how close
another country was to breaking, and the AI read it exactly.

It now reads the world's mean instead of the true figure, which is what
"everybody knows the war drags on, nobody knows who is closest to breaking"
looks like as a number. OD_AI_FOG_WEARINESS_OFF restores it.

What it costs depends on which model is playing, which is worth stating
because the last feature measured this way did not: on the shipped model it is
free (122 → 123 rating), and on the stronger N24 it costs 8 (187 → 179). Both
leave survival and the worst seat untouched to the digit, so the price is
entirely in the column that changes sign between world sets — and the reason
for the change is symmetry with the player rather than a number.

Everything else the AI reads about a rival is either visible per province or
derived from what is — the difference is effort, not access. Its own trend
features read only its own country.
You can see what your wars are costing you at home. War weariness is a
term in every province's rebellion chance and had no display in the game at
all, so a country could slide toward revolt because of a war it answered ten
turns ago with nothing on screen naming the cause. It now sits under the
unrest bar in Politics, when there is any.
The AI module can say whose AI it is. Neural stays observe-only, but a
mod reading the feature vector had no way to ask which AI produced it — and
that vector's layout is only stable within one ARCH. ai_version returns
"ParrotZero 8.4.0", ai_arch returns the number that gates it, and
country_stance reports what the AI has decided it is doing with a country
(expand, consolidate, defend, develop) — the one piece of its reasoning that
is a summary rather than a number.
Gearbox 1.2. The mod ABI learns the game the game has become: districts
(their names, budget shares, ground and regional law, readable and writable
through the game's own rebalance so the shares still sum to 100), what a
country publishes about itself, the army broken down by kind, a country's
expenses, what it is worth and how many people live in it, and the research
lock. 28 new imports; nothing renamed, re-signed or re-gated, and
sdk/compat/abi-1.1.json now freezes the 1.1 surface beside the 1.0 one so a
build that broke either fails the suite. Every 1.0 and 1.1 mod links and runs
unchanged.
A country's districts are public; how they are governed is not. The
profile shows the division and each district's share of the budget to anybody
who opens it — borders are visible, and so is what a government has drawn on
them. The regional law each district runs is a fourth publishable field,
beside expenses, doctrines and the treasury.
"Where the money goes" is a pie. Four numbers to divide in your head
became the same wedges the economy screen has always drawn — from the same
table, so the two charts cannot disagree about what a country spends.
AI countries govern their districts. They drew them and then did nothing
with them: the budget split was the whole of it, and since their pacification
budget is zero the split moved nothing. They now pass regional law where a
district is in trouble, priced as a share of gross income rather than out of
what is spare — an AI treasury is near zero as a matter of course, so a rule
gated on cash in hand is a prohibition dressed up as a budget.

The first version of that pricing counted the bill and ignored the income:
a tax holiday costs 0.00 per province in cash and a tenth of the district's
earnings, so it looked free and won 45 times out of 45. Counting income
forgone gives a spread — tax holidays where there is little to lose, language
rights and public works where there is.

Measured four ways — two seed sets at two horizons, one binary with
OD_AI_DLAW_OFF separating the arms:

instrument rating survival worst seat

120 turns, fixed seeds +16 +8 −3

120 turns, hold-out +4 −1 −13

400 turns, hold-out +29 +12 −10

400 turns, fixed seeds +15 +11 −12

Rating is up in all four and survival in three, and the floor is down in all
four — which looked like "lifts the median seat and costs the worst one" until
the per-seat tables were read instead of the summary line. The floor is the
same seat in all eight arms, Norway in a world that rushes it, and what the
rule does to it is this: 0.3 → 0.3, 0.4 → 0.2, 0.4 → 0.3, 0.5 → 0.3 percent of
the world's land. A seat holding half a percent has a tiny par, so a tenth of
a percent of land becomes fifteen points of score.

Meanwhile the other five seats improve in most arms, by up to 3.9 points of
land. So the rule helps five seats materially and takes a fifth of a percent
from one that is already collapsing. Left on; OD_AI_DLAW_OFF turns it off in
one variable.
Your chat can play a country. Hand a country to your stream and let
Twitch, YouTube or Kick decide its turn. Chat types commands, the game
tallies them for a window, and when the window closes the winner is applied
with everyone able to see why.

One vote each, because a tally where the loudest typist wins is a keyboard
test rather than a vote — a viewer may change their mind and the last thing
they typed counts, but they cannot stack. A command is a choice from a list
the game defines, never free text: chat is untrusted input and is treated
that way. The reader connects anonymously, holds no credential and cannot
post.

The overlay feed writes files for OBS instead of opening a port. The
obvious build is an HTTP endpoint on localhost, and it is the wrong one here:
it triggers a firewall prompt on first run, during a stream, on camera.
Discord rich presence. What you are playing appears under your name: the
screen, the scenario, the country. Never a save path, an invite code, an
account id or a session code.

That is not a privacy setting, it is the shape of the data — the only inputs
the code takes are a screen, a scenario name and a country name, so there is
nothing else it could leak. Presence is read by strangers in every server
you are in, which is the reason to build it that way rather than to filter it
afterwards.
An announcement board on the main menu. News, releases and tournaments,
fetched from the account service.

It is a sealed format, and that is the whole point: content from a server
drawn by a client is the exact shape of a remote code execution bug. So a body
is text marked up with the same parser the tutorial dialogue uses, which
produces styled glyph runs and can do nothing else. A button cannot name an
action — it picks one from a closed list defined in the game, and the document
supplies at most one short parameter whose meaning the game decides. Nothing
fetched can ask the game to open a URL, run a file or load a mod. Anything
unrecognised is refused whole rather than half-drawn, and no network at all
simply means no board.

instrument	rating	survival	worst seat
120 turns, fixed seeds	+16	+8	−3
120 turns, hold-out	+4	−1	−13
400 turns, hold-out	+29	+12	−10
400 turns, fixed seeds	+15	+11	−12

Infantry is a choice now, not a quantity. Every soldier used to be the
same soldier, so recruiting was one number and every battle was arithmetic on
it. There are four kinds now — three of them infantry — and each answers
differently in every column:

	money	manpower	frontage	attack	defence	fuel
Line Infantry	1.00	1.00	1.00	1.00	1.00	1.00
Militia	0.50	0.60	1.00	0.70	1.15	0.80
Assault Infantry	2.00	2.50	0.80	1.35	0.95	1.20
Mechanised	3.50	4.00	0.60	1.25	1.15	3.00

Frontage is the column that changes the game. A province fights across a
limited width, and above it both sides' counts cancel — power reduces to
width x stat/frontage, so the fight is decided by attack and defence PER
METRE rather than by how many men arrived. Bringing more stopped being the
only answer; bringing better now competes for the same manpower pool.

Line Infantry is exactly 1.00 in every column on purpose. Every soldier in
every existing save is line infantry, so the multipliers changed nothing at
all until a second type existed — the first stage shipped as a byte-identical
evaluation rather than a hope.

Militia was wrong once, and the mistake is worth keeping. It was drafted
at 1.20 frontage on the reasoning that an untrained rabble uses a frontage
badly. That sounds right and is arithmetically fatal: it made militia worse
than line at both attack and defence per metre, so it was strictly dominated
in the quarter of all assaults that are width-bound. An AI rule buying the
best defence per unit of money duly picked it and lost 34 and 28 rating points
on two models, with the floor falling from 110 to 26. Frontage is footprint,
not skill: a militiaman occupies a man's width like anybody else, and what he
lacks is training, which is the attack column.

The AI stops cutting research first when money runs short. Austerity took
the research slider down before anything else, on the reasoning that a slider
comes back up for free the moment income recovers. For a country that is merely
tight rather than failing, that is the growth engine being switched off to pay a
bill it could have met another way. Research now comes down last, after
everything else has been given.

Measured in the shipped binary with the shipped model, against its own absence,
on two hold-out seed sets it was never tuned against. Six of six seats scored in
every arm, no seeds lost:

rating survival worst seat

hold-out C 241 → 311 81 → 77 36 → 17

hold-out D 255 → 297 100 → 80 105 → 38

Both halves of that replicate, and the second half is not a small-par
artefact. The four seats not under existential pressure gain between 2.3 and
17.6 points of the world's land each. The seat being invaded loses ground on
both sets — several points of raw land share on a seat whose par is 6.7, so
real territory rather than a ratio effect on a sliver. On hold-out D the
previous behaviour held every seat above par and this one does not.

That loss is deliberately not quoted to the decimal. The invaded seat is
bistable: across 22 recorded runs it either holds roughly 6 to 11 percent of
the world or collapses below 1.5, with nothing in between, so a three-seed
mean of it is nearer three coin flips than a measurement — it prints to two
decimals and moves by a factor of three on the draw.

Nor does playing both arms on the same worlds rescue it. Pairing controls how
hard the world is; it does not control which side of the cliff a run lands on,
and a seat this close to a tipping point can be flipped by a small change to
the policy alone — the same seed has put two different models on opposite
sides of it. What holds the direction up is replication: the sign is the same
on two independent seed sets, and the rating gaps behind it (+71 and +42) sit
well outside the noise floor of a three-seed run, which is about a dozen
points of standard error. Expect a re-run on fresh seeds to reproduce the
direction and not the number.

Said plainly, because the rating alone would not say it: the AI expands harder
and holds considerably more ground when it is not fighting for its life, and a
country being overrun does worse than it did before.
OD_AUSTERITY_RESEARCH_LAST=0 restores the old order in one variable.
Publishing your books now matters to the AI, not just to migrants. It
read the true treasury, army, industry and ground of every country directly,
so a country's decision to publish or withhold changed nothing about how the
AI treated it. The relational feature that meant "their treasury" now means
"their treasury if they publish it, otherwise a guess from their industry and
their ground" — a country that keeps its books shut is genuinely harder for
the AI to read.

This wants a retrain, and the shipped model has not had one. The model
file format is unchanged, so a policy trained before today loads and reads the
new input as though nothing happened. Measured against that frozen policy the
switch costs 2 rating points, with survival and worst seat unchanged, which is
inside the noise floor — and the reason it is that small is that 11.1% of
relational reads are fogged, because AI countries publish their treasury about
73% of the time. OD_AI_FOG_TREASURY_OFF restores perfect information, which
is what a training control arm wants.

Two measurement knobs come with it, both default-off:
OD_AI_REL_ABLATE_TREASURY=<v> replaces that feature with a constant, which
bounds what the fog can cost (destroying it outright costs 5 rating points, so
fogging one read in nine cannot cost more), and
OD_AI_REL_ABLATE_WHERE=fogged|published destroys only one half of the reads,
which is what can tell a policy that learned the mechanic from one that leaned
on the guess — the unconditional version cannot, since both would lose more.
AI countries still do not spend on suppression, and now it is understood
why. The rule that funds it was documented as blocked by the AI's empty
treasury. Measured: with the risk asked about the right government, the worst
province anywhere in 203 reviews scores 10.14 against a bar of 12, so the rule
could not fire for any country however rich. Lowering the bar and funding a
floor out of gross income makes it fire (19 of 217 reviews, ~6% of income),
and costs the player 4 rating points at bar 5 or 13 at bar 3 — but the world
aggregate cannot say what the money buys at three maps, so both remain behind
OD_AI_PACIFY_BAR and OD_AI_PACIFY_GROSS rather than becoming the default.
The AI was judging its own provinces by the player's unrest. The reflex
that cuts a country into districts and decides where its suppression money
goes asked getProvinceRebellionChance(pid) — the one-argument form, which
answers "how likely is this province to revolt against the PLAYER", because
that is the only government the panels calling it ever ask about. Every AI
country was therefore cutting its districts and aiming its money by a figure
belonging to somebody else's government. It asks about the right one now.
Map scripts can see the game the game has become. The language knew about
treasuries, populations and borders, and nothing that has been added since:
a script could not ask how many soldiers a country had, what it earned, what
it was worth, or how it was governed. It can now — country.X.troops and
country.X.troops.militia for the army by kind, income, expenses,
national_value and population read from the same snapshot the economy
screen draws, district_count and district.N.share, and on a province
troops, district and rebellion_chance. Garrisons are writable per kind
(set province.42.troops.militia 12000).

Research groups can be forced open or shut — set country.USA.research_groups 1 holds a country to a single programme however rich it gets, 3 gives it
three however poor, 0 hands the decision back to its economy. The override
outranks the economic gate in both directions and is saved with the game, so
a lock a turn-zero script sets is still there after a reload.

New sugar: wait 5 turns (the target is worked out when the script parks, so
label / wait / jump really does wait five turns every time round rather than
racing past a date that has gone), set x to <expr>, {value} interpolation
inside print, and foreach district in country.X. Everything above has a
block in the editor's palette — which now scrolls, having outgrown its own
panel at seventeen entries.
A map's scripts no longer depend on its metadata being honest. The loader
scanned the archive for scripts/ entries and then let metadata.json
overrule what it found, so a map whose metadata was written before its scripts
were added carried scripts that silently never ran. Metadata can now only turn
scripting on, never off.
Districts are authorable in the map editor. The Countries tab gains a
Districts brush beside the claims one: add districts, name them, colour them,
set each one's claim on the pacification budget, and drag over provinces to
assign them. A district is a partition rather than an overlay, so painting a
province into one takes it out of the one that held it, and a legend on the
map says how much of the country is still in no district at all. It travels
in the map as districts.json, and a country the author divided starts the
game divided that way — laws included.

The AI treats an authored division as scenario content: it will not redraw a
country somebody cut up by hand, and decides only where that country's money
goes, bounded so no district the author drew is starved to nothing. It still
cuts its own districts freely for every country nobody drew.
The editor's bottom bar no longer prints its labels through its buttons.
It was laid out on fixed pixel offsets measured against English: "Tools" is
33px wide, "Інструменти" is 96, and in half the languages the game ships the
first tool button sat on top of the label — as did the brush-size number on
the slider. It now measures the labels and lays out from them, and the tool
names themselves are translated.
Country profiles. Any country's province panel now opens a profile of the
country that holds it: how long it has existed (in months, years or millennia,
or "existing since the start of the session" for a country that was on the map
when the game began), its population and where the graph has been going, its
gross income and its expenses, where it sits on the political compass, and the
flags it has flown, in order, with the turn each was raised.

Three further figures are the country's own to publish or withhold — the
composition of its expenses, the doctrines it has in force, and the treasury it
held at the start of last turn. Publishing is not free advertising and not
merely flavour: migrants read it. A country whose disclosed figures are good
draws people to it, up to a fifth again on the attractiveness a province
already had, so an honest, wealthy, calm country fills up faster than a
secretive one — and a country with bad figures is better off saying nothing.
The AI publishes and withholds on the same terms you do.
New territory joins the nearest district, not the first in the list. A
province it borders wins outright; otherwise the closest one, so an island or
a landing across a strait still lands somewhere sensible.
Districts can be removed, down to the last one, and their ground rejoins
the nearest surviving district. A district that loses all its provinces stops
existing.
Districts name themselves after the ground they hold. A people holding a
real majority lends its name — "Mongol County", "Norwegian County" — and
otherwise the district is named for where it sits in the country: Northern,
Eastern, Central. The word for a region is a property of the map, chosen
from its own content, so one world has counties and another has oblasts
rather than mixing them.
Every doctrine that says it reduces unrest now does. The value was stored
as a fraction and displayed as a percentage, but the resolver subtracted the
fraction straight from a figure measured in percentage points — so Secret
Police advertised "5%" and delivered 0.05, against a loyalty floor of 6. Every
such doctrine in the game did nothing at all. Rebellions per 1k country-turns
fall from 81 to 30 on the 1939 map and 77 to 30 on 1914, and the seat
benchmark goes from 72 to 121 with survival 49 to 87.
Districts pass regional law, which is its own body of content rather than
the national doctrine list under a smaller heading. Ten of them: a curfew, a
tax holiday, language rights in the courts, martial law, company towns, grain
requisition. Each is a trade — calm bought with money, or money bought with
resentment — and several make things worse on purpose. They are priced per
province, so the same law is a different bill in a big district than in a
small one, and their effects on unrest, income and population growth apply
only to the ground that district holds.
Districts: a new Politics tab where you divide the country up. Pacification
used to be one slider for a whole country — the same weight on a quiet county
as on the one that is arming, so you paid for the calm half in order to reach
the loud one. Now you can cut the country into as many districts as you like:
paint provinces into them on an inline map, colour them, and give each a share
of the budget. The shares always add to 100, and a button splits them evenly.
The pacification budget itself has moved to this tab.

What matters is a district's share of the money against its share of the
ground. One holding a tenth of the country on a tenth of the budget is
policed exactly as before — so dividing changes nothing until the shares are
uneven, and an undivided country plays the game it always played. Each row
shows the multiplier it comes to.
AI countries draw their own districts, splitting the worst fifth of their
provinces off from the rest and aiming the budget at the trouble.
The unrest tutorial covers it, in all twenty languages.
A new game is a new world. The world seed only ever reached the turn
resolver, so two new games on the same map started from an identical position
— same countries, same politics, same flags — and differed only in what
happened afterwards. A fresh world now nudges each country's political compass
by up to 18 points of 200, which is enough to carry the ones sitting near a
threshold across it and shows up as different ideologies, names and flags. The
same seed still reproduces the same world exactly.
Fixed the black screen after Process Turn. Regenerating the political map
is seconds of work and had been moved out of the turn's loading screen into
the draw loop, where nothing was on screen while it ran. It happens inside the
turn again, behind its progress bar.
Orders are shown in the view they belong to. Marches and artillery in Army
Navigation, voyages and naval guns in Navy — matching the recruitment and
construction marks, which already worked that way. The banner says which view
carries what.
Fixed the claims tabs overlapping. They sat on a fixed 140px pitch whatever
they said, so any language whose words are longer than English's ran one label
into the next.
The game renders at the display's real pixel density. Until now the
framebuffer was the window's logical size and the system stretched it to the
panel, so on a 2x display every pixel of the map, the interface and the text
was drawn once and shown as four. That is what "blurry on a big display" was,
and the font was only the part that showed it first. Set "highDpi": false in
config.json to go back.
Text is sharper in every script that can afford it. The glyph atlas was
rasterised at 16 pixels and scaled up for every size drawn, which on a large
display is a 16-pixel bitmap smeared to fit. Cyrillic, Latin, Greek, Arabic,
Devanagari and Turkish are now rasterised at 32. Japanese, Chinese and Korean
have several times as many glyphs and stay at 16, so that no atlas is ever
larger than the biggest one the game already built — nothing asks a machine
for memory it was not already asked for.
A nation can be freed as part of a settlement. Releasing was a button you
pressed on yourself and nobody could ever ask you to — which is the shape of
most real peace treaties. It is now a term like any other in a ceasefire or a
trade: offer to free one of your own peoples, or demand the other side frees
one of theirs. Received offers name it plainly, and the offer map draws the
ground in its own colour, because land leaving for a new country is not the
same event as land changing hands between the two of you.
You choose which territories a freed nation gets. Releasing used to hand
over the whole region a people held, all or nothing. Now pressing Release
opens the choice: the region is marked on the map, you keep or give each
province, and the game says why a selection will not do — too small, or in two
pieces. A country in two halves with someone else's ground between them is not
a border anybody drew.
The orders view shows naval bombardment. A carrier group working over a
coast was the one kind of attack it could not draw, so a shelled shoreline
looked like a quiet turn. Shown as a shell's flight from the hull, with a mark
at the firing end so it reads apart from a battery inland.
All twenty languages are complete again. The turn phase, research groups,
troop kinds and economy graphs are translated, along with a set of older map
editor and goods strings that had never been done — 86 strings, 1507/1507 in
every language.
Research groups are cards, not a strip. Each one now shows what it is
building, a progress bar for it, its share of the budget and its Auto switch,
side by side so you can compare the three at a glance. They were 21px rows: a
name clipped to fit, no progress at all, a slider the width of a thumbnail —
the primary control of a whole subsystem drawn as the smallest thing on the
screen. On a narrow canvas they drop to a row of their own rather than off
the edge.
"Show orders" is now a pause after the turn, not a layer over play. It
used to paint every order in the world onto the live map while you were still
giving orders — a labelled box on every province on Earth, on top of the
troop counts already there. It now means "stop after each turn and show me
what everyone did", which is the same setting the phase's own "Skip this from
now on" writes, and it works from the first turn.
Orders use the map's own symbols. A levy, a works going up, a hull on the
slipway and a specialisation are marked with the same glyphs your own pending
orders have always worn, in the same view each belongs to — so the industry
map is not also an army map. Tinted by country, so whose order it is stays
readable when four countries are building along one border. Standing orders
appear once you look at a region rather than at world zoom, and the banner
says how many are waiting there.
Research budgets add up to 100%. The three group shares were each
independently 0–100, so two groups could both read 90% and the panel showed a
country spending 180% of its research. Nothing was overspent — the points were
normalised before being paid out — which is worse, because the numbers on
screen then meant nothing. Moving one share now pushes the difference onto the
others in proportion to what they hold. AI countries run the same rule.
AI countries run research groups too, on the same unlock rule as you.
Measured: research completions per 1k country-turns rise from 7.4 to 11.3 on
the 1939 map and 10.3 to 15.4 on 1914.
Assault Infantry now follows Militia Levies instead of Combined Arms, so
the Formations tree is a ladder of its own rather than three nodes whose
prerequisites live in another tree.
The economy screen shows what the country is worth. Two new graphs: income
against expenses with the zero line drawn, and national value — everything
built and fielded priced from the same tables the build buttons charge from,
plus a per-head figure. The old five-line income graph flattened net income
into the baseline and put its legend underneath the pie charts.
Fixed the two pie charts overlapping. The second advanced 100px from the
first one's title, but the first is 110px of circle plus a legend that runs
further, so any country with more than about four kinds of expense had its
charts drawn on top of each other.
Research runs in up to three groups. Each has its own project, its own
share of the budget and its own Auto switch, so a country no longer queues
every technology in the game behind one. How many you get is decided by the
size of your economy against the rest of the world — a second group at twice
the world's median, a third at five times — and then gated by income per
head: an economy spread too thin over too many people supports one programme
however large it is in total. One is always available whatever your position,
and a locked group tells you which of the two is holding it back. On the 1939
map the great powers all run three; China, with the fifth largest economy on
the map and a fifth of the world's income per head, runs one.
A research group can follow its branch on its own. Turn Auto on and it
starts the next technology as soon as the last one finishes, stopping the
moment the branch offers a real choice — two nodes open at once — or runs out.
It will never pick which branch to start; that is the judgement it exists to
hand back to you.
Formations and Efficiency are their own research trees. Both were extra
columns bolted onto trees that were already full, with prerequisite lines
running the whole width of the canvas across two other branches to reach
them. A prerequisite in another tree now draws no line and is named in the
tooltip instead, with the tree it lives in.
A province can raise more than one kind of soldier a turn. Each kind
keeps its own slider and its own order, all drawing on the one population, and
a kind already on order is marked on its tab. Queueing militia used to lock
the province for the turn.
Fixed the recruitment ceiling reading a hundred times too high. Garrison
counts are stored a hundred to the soldier and the readout was not dividing,
so it offered "82% of 1.08m" beside a button raising 8.9k men. The button was
right all along.
Ships no longer sail across land. Sea routes were planned on a grid where
a cell counted as navigable if it held any water, so two cells on opposite
shores of a peninsula were treated as neighbours and the leg between them ran
overland. Russian hulls crossed Crimea; the same fault sent fleets over
Jutland, over Italy and across the base of Florida. Every leg is now checked
against the map before the router may use it. Measured on the shipped world,
the worst stretch of dry land under a route falls from 53 pixels — about 260
km — to 16, which is the width the game deliberately treats as an undrawn
strait so the Bosphorus and the Dardanelles stay open.
The land no longer changes while you are reading the orders. The Viewing
Orders screen showed the map as it was after the turn resolved, so borders
had already moved to their new owners while you were still being shown the
attacks that moved them. The political map now holds still until you press
Continue. As a side effect a turn that ends several wars regenerates the map
once instead of once per ceasefire.
The orders view says what countries are building and raising, not only
where their armies went — recruitment and construction now appear on the
province they are happening in.
Army orders are drawn as army arrows, weighted and filled like the ones in
the army view, and each carries the share of the garrison that marched. The
draggable knob is gone from this view: it was a handle on an order that had
already resolved.
The recruitment slider shows what its ceiling actually is. The cap has
always been counted in people, so a province that raises 120k line infantry
raises only 30k mechanised — true from the day the troop types landed, and
invisible, because the readout showed a percentage and nothing else. It now
reads "50% of 1.65b", and the figure moves when you change the kind.
Show orders no longer takes your sidebar away. Suppressing the tabs is
right for the Viewing Orders phase, where no order can be given; it was wrong
for the toggle, which is a lens held over ordinary play.
Fixed the army research tree drawing through the navy column. The new
Formations branch was placed at the far right of the canvas, so its
prerequisite lines ran the full width of the tree and across two other
branches. It now sits beside Army, where everything it depends on is.
The turn now has a middle. When a turn resolves, the game pauses on a
"Viewing Orders" screen showing what every country actually did — the marches,
the bombardments, the voyages — before handing the map back to you. Nothing
takes orders while it is up; a Continue button moves you on, and "Skip this
from now on" turns it off for good if you would rather not have the beat.
You can still open the same view any time from the Show orders option under
Process Turn.
Your army can be made of different things. A new Formations branch in the
army research tree unlocks militia, assault infantry and mechanised troops.
Each is a different bargain: militia are cheap in money and in men and hold
ground well but crowd a battle line; mechanised cost four times the manpower
per soldier and drink fuel, but use a frontage better than anything else.
Pick which kind a province raises with the new selector above the Recruit
button — and because a better soldier represents more of your population, the
same province raises far fewer of them.
You can see what is standing in a province, and order it about by kind.
The army view listed one line per stack and ran off the bottom of the panel if
there were many. It now shows your garrison broken down by type, then anybody
else's troops by country — scrollable, and sized to what is actually there.
Click a type to aim your next order at just those troops; click "All our
troops" to go back to commanding the province as a whole. On the map, a
province holding more than one kind of soldier gets a small bar under its
marker showing the mix.
Groundwork: an army is made of something. Soldiers now have a kind. Nothing
in this release changes — every soldier in every existing campaign is line
infantry and behaves exactly as before, and a save written now is byte-for-byte
what the previous version would have written. What it buys is everything that
comes next: different troops, at different costs, drawing on a limited pool of
manpower. Old saves load correctly, and a save from this version can still be
read by an older one, which will see the right number of men.
A guarantee is worth something in the war you signed it for. Guarantees
only ever fired at the moment war was declared, so a country that won one on
turn three of a war it was losing got nothing from it — the pact was dead
paper for the only war it was obviously signed for. A guarantor can now be
called to arms like an ally, and can accept or refuse. And calling for help is
no longer something only the human player could do: every country can now ask
its friends to join a war already under way, which is the point of having
friends.
A beachhead is supplied by the fleet that put it there — while the fleet is
still there. Troops landed on a hostile shore have no road home, which the
supply rules would otherwise read as being encircled. A friendly or allied
fleet within landing range now supplies them. Lose that fleet, or sail it
away, and the beachhead is genuinely stranded and fights like it. Landing
under cover of a navy and then leaving is no longer free.
Battles now last, and you can feed them or pull out. An attack used to be
over the moment you ordered it: it took the province or it was thrown back and
the survivors walked home by themselves. Now an attack that neither wins nor
dies stands in the province and fights on, a round each turn. Send more
troops to that province and they join the fight instead of starting a separate
one; or order a withdrawal from the province panel and pull your men back to
where they came from. Retreating used to be free and automatic. It is a
decision now — and so is committing.
A failed landing still drowns: there is nowhere for it to fall back to.
Your army in a battle still eats, still draws ammunition, and still costs
upkeep, because a war has to be paid for while it is being fought.
Armies far from home fight worse, and cut-off armies fight badly. Nothing
in the game knew how far an attacker had marched: a stack twenty provinces
deep fought exactly as well as one defending its own capital, so pushing deep
cost nothing and defence in depth had no point beyond stacking forts. Supply
is now measured from your ports and your largest industrial province, along
roads through your own and your allies' territory. Two provinces out costs
nothing; beyond that your troops fight progressively worse; and a province
with no land route home at all is cut off and fights badly. Severing a
corridor is now a real operation with a real payoff — and so is holding one
open. A landing is supplied from the sea rather than treated as encircled.
Bringing more men to a battle matters again. Combat width caps how many
troops either side can bring to bear on a province — but it capped both
sides, so once two armies both filled the frontage the fight was decided by
research bonuses alone and returned the same answer every turn no matter how
many men you sent. An attack with 174,800 men against 93,048 was repulsed three
times running, each time losing exactly the same number of soldiers, because
the extra 139,000 counted for nothing. Men beyond the frontage are now the
reserve: they can't widen the fight, but they feed it, so a deeper army fights
harder. A narrow fortified pass still stops numbers — that is what the frontage
is for — but on open ground, weight of numbers tells.
A doctrine that lets you fight several countries at once. You could only
declare one war a turn, and the reason written in the code only ever justified
the narrower rule it sat next to (you still cannot say two things to the same
country in one turn, and that has not changed). War on Several Fronts raises
it to three: an authoritarian doctrine at 16/turn over six turns, paid for with
a fifth more army upkeep, weaker defence in the field and more unrest at home.
The old limit also lived in the diplomacy panel rather than in the rules, so it
bound you and not the countries you were playing against — it is one rule now,
and everyone is held to it.
The mobilisation doctrines now actually work. How a country raises its
army was supposed to change what the army costs to recruit and to keep —
conscription cheap to raise and dear to maintain, a professional army the
other way round. That was written, documented, and never reached a single
game: the shipped maps carry their own copy of the doctrine data, and it had
not been updated. Fourteen doctrines now charge the army upkeep and
recruitment costs they had been advertising all along, so armies are more
expensive than they were and a large one is a real commitment.
Going broke for one turn no longer breaks your country. An empty treasury
added a flat, very large amount of unrest to every province from the first
turn it happened — the same charge whether you had been insolvent for one turn
or twenty. It now builds over three turns: a third, two thirds, then the full
weight. Chronic bankruptcy is exactly as punishing as it always was; what has
gone is the case where an invader takes your industry and the resulting
one-turn cash shock fractures an otherwise stable country before any
government could have cut a budget. Measured on three worlds: 3.8 to 5.7 more
countries in a hundred survive, and rebellions fall by 17–30%.
You can see where your ships are actually going. A pending naval move drew
a straight line to the destination — which is the one route no ship ever
sails, because they steer around land. The map now shows the real path, with
the stretch your ship covers this turn drawn solid and the rest of the
voyage faint, a dot where it will stop tonight and a ring on the destination.
A destination with no sea route to it is marked with a red cross instead of a
line, rather than looking like a very long voyage.
A new Orders view shows what everyone did this turn. Toggle it from the
button above Find country: artillery, army movements and naval voyages for
every country, each in its own colours, drawn over whichever map view you were
already in. Politics, Economy and Research step out of the way while it is on
— it is a view for reading the board, not changing it — while Find country,
Settings and Claims stay put. It works in multiplayer, where the host sends
each turn's orders to everyone; an older client simply does not show the
overlay. It shows the turn that has resolved, never the orders other players
have not yet committed.
Every new game is now its own world. Starting a new game drew from the
same fixed random sequence every time, so the same rebellions arrived in the
same provinces on the same turns in every campaign anybody ever played. New
games now take a genuine random seed, and the load log records it — if a
world does something worth showing somebody, the number that reproduces it is
in your log. Saving and reloading continues the world's luck from where you
left it rather than rewinding it, so reloading is not a way to reroll a bad
turn. Set OD_WORLD_SEED to pin a world deliberately; the AI benchmarks pin
their own and are unaffected.
A province will only carry the industry its land can support. Level 10
used to be available anywhere on the map, so an uninhabited island with no
resources could out-produce a city. Every province now states its own ceiling
in the province panel — "Level II of VI", and it says so when it is full —
set by how many people live there, how tightly they live, how much ground it
covers, and what is buried in it. A factory in an empty province really is an
expensive shed now, which is what the tutorial has been telling you all along.
What a factory earns depends on where it stands. Industry income was a
flat rate per level everywhere. A factory built well beyond what its province
can support now earns proportionally less, and one built in a province that
can carry it earns exactly what it always did.
Provinces you have already overbuilt keep everything they have. The new
ceiling applies to what you build next, not to what you built before. An
existing factory above its province's ceiling is left standing and says so in
the panel — you will see a handful of these on the shipped maps, in the
historically industrial parts of France and Belgium.
Your population no longer runs away. It was growing 52-fold over a long
campaign — two billion people at the start of the 1939 map became a hundred
and ten billion — because population growth was being applied once per ethnic
group in a province instead of once per province, and the ceiling meant to
stop it was set 133 times higher than the largest province that has ever
existed. Growth is now applied once per province, ethnic policies are weighted
by each group's share of it, and the ceiling is the ground a province actually
covers.
Mobilisation doctrines change what an army costs. Raising one and keeping
one are now opposed: Mass Mobilisation makes recruits 30% cheaper and upkeep
25% dearer, Demobilisation the other way round, a Professional Army costs more
both ways and fights far better. Seventeen doctrines carry these now, and the
economy panel shows the rate you are paying.
Army maintenance discounts finally do something. Eight doctrines and
research nodes have advertised an army upkeep reduction — Demobilisation says
-25%, Austerity -20% — since the day they were written, and not one of them
was ever applied to anything.
You can let a nation go. Where a disaffected people holds the majority
across a run of your provinces, you can grant them independence: they leave as
a country at peace, under your guarantee. You lose the land, the people and
the industry; you lose the unrest with them, and gain a friend. Repressing a
minority is what makes their region releasable, and reconciling them is what
removes the need — so this is the last rung of the ethnic-policy ladder rather
than a lever of its own.
A country that cannot pay for itself for years on end will shed a region
rather than disintegrate. It is the final step of going broke, after budgets,
doctrines, minority programmes, ships and troops — and the only one that
cannot be undone.
A province can only hold so wide a front. Numbers stop being decisive past
what the ground will take: a huge army no longer automatically rolls over a
well-sited smaller one, and a fort now narrows the front as well as
strengthening the defenders. Wide provinces let more of an army bear, narrow
ones are passes.
A failed attack no longer destroys your whole army. The men who fought are
lost; the ones still in reserve fall back to the province they came from. A
failed amphibious landing still drowns — there is no ground behind it.

	rating	survival	worst seat
hold-out C	241 → 311	81 → 77	36 → 17
hold-out D	255 → 297	100 → 80	105 → 38

Fixes

Ships stop sailing the wrong way round the world. Near the date line a
one-degree hop measured as three hundred and fifty-nine, so a fleet in the
Aleutians would spend its whole turn steaming east across the Pacific instead
of taking one step west, and the route preview quoted twenty turns for a
two-turn crossing. The router always knew the way; nothing that measured or
followed it did. Ship moves stopped dead by land fell from roughly a third to
under a tenth.
The move preview no longer draws a route that does not exist. When the
router could not find a path it drew a single straight line to the target,
across whatever land lay between. Unreachable destinations are now shown as
unreachable.
A conquered country's claims die with it. Destroy a neighbour and every
claim it had ever pressed stayed on the board for the rest of the game,
agitating your provinces — visibly, and with no way to answer them, since the
country that could drop them no longer existed. On a long game two thirds of
all claims belonged to the dead. Rebellions fell by well over half.
"Declare War" tells you why it is greyed. You may only declare one war a
turn, and the button did not know it: after your first declaration every other
one accepted the click, did nothing, and said nothing. You may still only
declare one, but now the game says so.
Splitting a garrison no longer strands a man behind. Dividing a province's
army in two rounded both halves down, so an odd garrison left a single soldier
standing — an army marker on the map reading "<1" after you had moved
everybody out. Half of all garrisons are odd, so it happened about every other
time you split one. Shares are now taken so they add up to the whole.
Orders die with the ground they were given for. A disband or move order on
a province you then lost survived, and fired if you ever took the province
back — an army recaptures a province and immediately disbands itself for a
decision made before the province changed hands twice.
The disband marker names the number. It said "Disbanding..." and nothing
else, while the order means "everything here when the turn resolves" — so
troops moved into a marked province went with it, invisibly. It now shows what
will actually go.
You can search your Implementing and Active doctrines, not only the ones
you have yet to enact. The search reads descriptions and effects as well as
names, so "upkeep" or "navy" finds the doctrines that do those things.

Experimental, off by default: the production economy

Set OD_GOODS=1 to play a world where factories make things instead of
money. Not finished, not balanced, and not on unless you ask for it — a world
started without it plays exactly as it did before, to the number.

Factories produce goods. Consumer goods, machinery, fuel and munitions,
made from the oil, metal, rubber and gemstones your provinces hold. Gold is
still money.
Your population eats. Consumer goods feed it, and how well fed it is shows
in the economy panel as living standards. Fall short and unrest rises
everywhere at once until you fix it — build more, take somewhere that can, or
trade for what you lack.
Undirected factories direct themselves, making whatever the country is
shortest of and can actually build. Choosing for yourself — the planned
economy, and the politics of who decides — is not in yet.
OD_AUTOSELL_PCT sets how much of your surplus raw material sells itself
each turn: 100 (the default) behaves much like the old economy, 0 means
nothing sells unless you trade it, and the interesting settings are between.

Armies now cost things, not just money. Recruits need munitions, artillery
and naval bombardment need fuel and munitions, and a standing army burns fuel
every turn — if you have none, you buy it in at a painful price, and the economy
panel names that bill. Your mobilisation doctrines set both exchange rates: the
one that makes recruits cheaper makes their munitions cheaper too, and the one
that cuts army upkeep cuts its fuel. So machinery, fuel and munitions are worth
making at last, and a large standing army is a burden on your factories rather
than only on your treasury.

Known gaps: national stockpiles are not saved yet (they reset when you load;
factory assignments do persist), armies and artillery still cost only money,
and the production rates have been tuned by hand rather than measured.

Navigation Menu

Uh oh!

v1.2.0a

Choose a tag to compare

Sorry, something went wrong.

Sorry, something went wrong.

Uh oh!

No results found

game 1.2.0a

Fixes

Experimental, off by default: the production economy

Uh oh!