diff --git a/CHANGELOG.md b/CHANGELOG.md
index 32d7c9a..8f08809 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -7,6 +7,19 @@ This file was written after the fact, reconstructed from the git history rather
turn by turn. That is why the early entries bundle a lot into one release and the later
ones get more specific. It will stay current from here on.
+## [0.3.4] · 2026-07-06
+
+The checkability release: everything a skeptic checks first now checks out.
+
+- Every commit hash cited in the balance history resolves on the published branch, and the deploy script refuses hashes that do not
+- The evidence map's worked example quotes the same METR figure as its table
+- Pedagogy research reclassified from design to the class its claims deserve; two umbrella registry entries dissolved (162 entries now)
+- The cautious preset says on its card that it is hard mode; the title screen says most terms end badly
+- Roadmap gates restated in currency the project can actually count
+- The sixth ending keeps its mystery: theme description removed from public docs
+- House typography rule documents its own rationale in the validator
+- Narration licensing scoped to what the terms actually grant
+
## [0.3.3] · 2026-07-06
Precision and provenance.
diff --git a/CITATION.cff b/CITATION.cff
index 1311596..c912711 100644
--- a/CITATION.cff
+++ b/CITATION.cff
@@ -8,7 +8,7 @@ authors:
repository-code: 'https://github.com/chipmates/criticalwindow'
url: 'https://criticalwindow.org'
license: AGPL-3.0-only
-version: 0.3.3
+version: 0.3.4
date-released: 2026-07-06
keywords:
- serious game
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index 7c146e1..03e9cdb 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -16,7 +16,7 @@ There are two ways in, and the one without code matters at least as much.
**Later: translations.** The game is built for it (all text lives in `data/strings/`), but translation opens at Phase 4 when the text stabilizes. Do not translate yet, it would churn.
-**Writing style for anything players read:** short sentences, plain words, no em dashes. Serious about the stakes, dry about the absurdity. If it sounds like a press release, rewrite it.
+**Writing style for anything players read:** short sentences, plain words, no em dashes (house typography; displayed text is data here, so the style rule is enforced by the validator like every other data rule). Serious about the stakes, dry about the absurdity. If it sounds like a press release, rewrite it.
## The code path
diff --git a/README.md b/README.md
index 32e1df7..6d63d09 100644
--- a/README.md
+++ b/README.md
@@ -29,7 +29,7 @@ pnpm install && pnpm dev
There is also a print-and-play paper kit: download the PDF from the latest [release](https://github.com/chipmates/criticalwindow/releases), or generate it yourself with `pnpm exec playwright install chromium` once and then `pnpm print-kit`. Board, cards, rules, and one sealed envelope you are not allowed to open until the end.
-Three ways to play. Solo as the United States. Solo as China, where the chips are scarce but the power is not. Hotseat: two people, one device, one shared world, two private screens of doubt.
+Three ways to play. Solo as the United States. Solo as China, where the chips are scarce but the power is not. Hotseat: two people, one device, one shared world, two private screens.
## What a run feels like
@@ -77,7 +77,7 @@ Off by default, everything works silent. Music by Scott Buckley (CC BY 4.0), see
## Who makes this
-Critical Window is a project of ChipMates gemeinnuetzige GmbH, a German nonprofit, built by a very small human core working with heavy AI assistance; how and why is spelled out in [GOVERNANCE.md](GOVERNANCE.md), and the whole trust architecture exists so you never have to take anyone's word for anything, human or machine. No revenue, no ads, no investors; if grants ever fund this, the funders get named in [GOVERNANCE.md](GOVERNANCE.md), which also explains how this project is built and who answers for it. Contact: criticalwindow@chipmates.ai. Legal: [imprint](https://criticalwindow.org/imprint.html) and [privacy](https://criticalwindow.org/privacy.html).
+Critical Window is a project of ChipMates gemeinnuetzige GmbH, a German nonprofit. One maintainer, working with heavy AI assistance; how, why, and who answers for it is spelled out in [GOVERNANCE.md](GOVERNANCE.md), and the whole trust architecture exists so you never have to take anyone's word for anything, human or machine. No revenue, no ads, no investors; if grants ever fund this, the funders get named there too. Contact: criticalwindow@chipmates.ai. Legal: [imprint](https://criticalwindow.org/imprint.html) and [privacy](https://criticalwindow.org/privacy.html).
## License
diff --git a/ROADMAP.md b/ROADMAP.md
index 2c4e726..09856a3 100644
--- a/ROADMAP.md
+++ b/ROADMAP.md
@@ -22,12 +22,12 @@ A print-and-play kit generated from the same data files the digital game uses. I
### Phase 2: Prototype. Built, gate in measurement
-The game in your browser: both the US and China seats, hotseat on one device, hidden alignment dice, eval reports that flatter you exactly when it matters, a society with its own clock, five endings, voice narration, and a debrief that opens the sealed envelope and shows what your run argued. Balance was tuned by an adversarial campaign of several hundred model-played games; its methods, its ending distributions, and the three times it caught its own instruments lying are documented in [docs/BALANCE.md](docs/BALANCE.md). But this phase's gate is human by definition (15-minute median sessions, unprompted replays, takeaways named unaided), models cannot feel whether losing is fun, and we do not mark our own gates passed. It gets measured with real players during the alpha, and this line changes when it does.
+The game in your browser: both the US and China seats, hotseat on one device, hidden alignment dice, eval reports that flatter you exactly when it matters, a society with its own clock, five endings, voice narration, and a debrief that opens the sealed envelope and shows what your run argued. Balance was tuned by an adversarial campaign of several hundred model-played games; its methods, its ending distributions, and the three times it caught its own instruments lying are documented in [docs/BALANCE.md](docs/BALANCE.md). But this phase's gate is human by definition (runs that get finished rather than abandoned, per voluntary reports, unprompted replays, takeaways named unaided), models cannot feel whether losing is fun, and we do not mark our own gates passed. It gets measured with real players during the alpha, and this line changes when it does.
### Phase 3: Public Alpha. You are here
The first real launch. In rough order: the hosted game at criticalwindow.org, real players through the Phase 2 human gate, balance shaped by people instead of proxies, more events for replay depth, art beyond the stamp glyphs (with every prompt and decision published), and an itch.io satellite page that points home.
-**Gate:** a thousand organic plays and five outside contributors.
+**Gate:** completed-debrief reports arriving steadily without prompting and five outside contributors.
### Phase 4: 1.0
diff --git a/SOURCES.md b/SOURCES.md
index 3b63f40..808744e 100644
--- a/SOURCES.md
+++ b/SOURCES.md
@@ -8,8 +8,8 @@ in `data/` without a source ID fails the build, and so does a registry entry tha
claims a tier its citations do not support. [`docs/EVIDENCE.md`](docs/EVIDENCE.md)
lists every cited number with its evidence.
-**164 entries. 54 external sources drive numbers directly (249 citation sites). Game-design constants cite the project's own [design constitution](docs/DESIGN.md) instead (85 sites), and those are always labeled `design`, never counted as outside evidence. 11 more sources shaped the design, 98 are further reading.**
-Status: 9 book, 5 flagged, 150 verified.
+**162 entries. 54 external sources drive numbers directly (249 citation sites). Game-design constants cite the project's own [design constitution](docs/DESIGN.md) instead (85 sites), and those are always labeled `design`, never counted as outside evidence. 11 more sources shaped the design, 96 are further reading.**
+Status: 3 flagged, 159 verified.
What the statuses mean: `verified` says the link was fetched and the title and
authors matched this entry (July 2026, scripted fetches plus hand checks; where
@@ -36,7 +36,7 @@ Not outside evidence, and never counted as such: game-design constants cite the
project's own design document so that no game-feel number ever has to wear a
fake empirical citation.
-- **SRC-DESIGN-HANDOVER** · [Design constitution (founding brief drafted 2026-07-03 under the working title Race Conditions)](https://github.com/chipmates/criticalwindow/blob/v0.3.3/docs/DESIGN.md) · project (2026) · `design`
+- **SRC-DESIGN-HANDOVER** · [Design constitution (founding brief drafted 2026-07-03 under the working title Race Conditions)](https://github.com/chipmates/criticalwindow/blob/v0.3.4/docs/DESIGN.md) · project (2026) · `design`
Used for: design constants with no empirical referent (turn counts, starting defaults, allocation defaults)
## External sources that drive numbers (54)
@@ -46,7 +46,7 @@ Every entry lists each place it is cited. Sorted by citation count.
- **SRC-AI2027** · [AI 2027](https://ai-2027.com/) · Kokotajlo, Alexander, Larsen, Lifland, Dean, AI Futures Project (2025) · `forecast`
Used for: race vs. slowdown endings; core scenario spine; cautious preset anchors
Cited 21× by: anchors.json → tracks.capability · events/autonomous_targeting_demo.json → sourceIds · events/branch_point.json → sourceIds · events/civil_military_fusion.json → sourceIds · events/grid_crunch.json → sourceIds · events/intelligence_explosion.json → sourceIds *(+15 more)*
-- **SRC-SIM-GAMING-INSIGHTS** · [Strategic Insights from Simulation Gaming of AI Race Dynamics](https://arxiv.org/pdf/2410.03092) · Gruetzemacher et al. (2024) · `design`
+- **SRC-SIM-GAMING-INSIGHTS** · [Strategic Insights from Simulation Gaming of AI Race Dynamics](https://arxiv.org/pdf/2410.03092) · Gruetzemacher et al. (2024) · `analysis`
Used for: facilitator insights from 43 Intelligence Rising games
Cited 18× by: events/asml_export_squeeze.json → sourceIds · events/autonomous_targeting_demo.json → sourceIds · events/liability_shield_ask.json → sourceIds · events/taiwan_strait_crisis.json → sourceIds · events/the_project.json → sourceIds · events/weight_theft.json → sourceIds *(+12 more)*
- **SRC-CSIS-EXPORT** · [Choking off China's Access to the Future of AI](https://www.csis.org/analysis/choking-chinas-access-future-ai) · Gregory Allen, CSIS (2022) · `empirical`
@@ -124,13 +124,13 @@ Every entry lists each place it is cited. Sorted by citation count.
- **SRC-RAND-WEIGHTS** · [Securing AI Model Weights: Preventing Theft and Misuse of Frontier Models](https://www.rand.org/pubs/research_reports/RRA2849-1.html) · Nevo, Lahav, Karpur et al., RAND (2024) · `analysis`
Used for: weight-security mechanic; also cited in the US-China section as the model-weight-theft / espionage mechanic (RAND, RR-A2849-1)
Cited 2× by: events/weight_theft.json → sourceIds · policies/weights_security_program.json → sourceIds
-- **SRC-CHIP-WAR** · Chip War: The Fight for the World's Most Critical Technology · Chris Miller, Scribner (2022) · `empirical` *(book, obtain manually)*
+- **SRC-CHIP-WAR** · Chip War: The Fight for the World's Most Critical Technology · Chris Miller, Scribner (2022) · `empirical`
Used for: TSMC/Taiwan chokepoint background
Cited 2× by: events/asml_export_squeeze.json → sourceIds · events/taiwan_strait_crisis.json → sourceIds
- **SRC-EU-AI-ACT** · [EU AI Act (Regulation (EU) 2024/1689)](https://artificialintelligenceact.eu/the-act/) · EU (2024) · `empirical`
Used for: Backs the fixed historical beats eu_ai_act_full_force and eu_ai_act_high_risk, dramatizing the Act's phased compliance deadlines (prohibited-practices ban, GPAI duties, high-risk certification).
Cited 2× by: events/eu_ai_act_full_force.json → sourceIds · events/eu_ai_act_high_risk.json → sourceIds
-- **SRC-IABIED** · [If Anyone Builds It, Everyone Dies](https://ifanyonebuildsit.com/) · Yudkowsky & Soares (2025) · `analysis` *(book, obtain manually)*
+- **SRC-IABIED** · [If Anyone Builds It, Everyone Dies](https://ifanyonebuildsit.com/) · Yudkowsky & Soares (2025) · `analysis`
Used for: Grounds the cautious worldview's high alignment-difficulty range (upper half of the 550-950 band) and, via the book's core policy prescription of a halt on frontier AI development, the global_moratorium policy card's pause mechanic.
Cited 2× by: parameters.json → worldviewPresets.cautious.alignmentDifficulty · policies/global_moratorium.json → sourceIds
- **SRC-AI-INCIDENT-DB** · [AI Incident Database](https://incidentdatabase.ai/) · Responsible AI Collaborative · `empirical`
@@ -181,7 +181,7 @@ Every entry lists each place it is cited. Sorted by citation count.
- **SRC-CARLSMITH** · [Is Power-Seeking AI an Existential Risk?](https://arxiv.org/abs/2206.13353) · Joseph Carlsmith (2022) · `analysis`
Used for: decomposed risk model
Cited 1× by: parameters.json → worldviewPresets.cautious.alignmentDifficulty
-- **SRC-SNAKE-OIL** · AI Snake Oil · Narayanan & Kapoor, Princeton University Press (2024) · `analysis` *(book, obtain manually)*
+- **SRC-SNAKE-OIL** · AI Snake Oil · Narayanan & Kapoor, Princeton University Press (2024) · `analysis`
Used for: skeptical worldview
Cited 1× by: parameters.json → worldviewPresets.skeptic.alignmentDifficulty
- **SRC-AI-INDEX-2025** · [The 2025 AI Index Report](https://hai.stanford.edu/ai-index/2025-ai-index-report) · Stanford HAI (2025) · `empirical`
@@ -229,12 +229,12 @@ mechanic, checkably. Wiring one of these to an actual number is a welcome PR.
Shaped: The agencyErosion.diffusionShieldMin mechanic (parameters.json worldRules.agencyErosion), where a high enough diffusion/benefit-sharing share lowers erosion of human agency, per its own note 'broad benefit-sharing keeps humans in the loop', reflects this paper's core proposal that redistributing AI's economic windfall preserves participation and blocks narrow power concentration.
- **SRC-AI-FUTURES-ROLEPLAY** · [Exploring AI Futures Through Role Play (Intelligence Rising)](https://arxiv.org/abs/1912.08964) · Avin, Gruetzemacher, Fox (2020) · `design`
Shaped: The design constitution (docs/DESIGN.md §1) names Intelligence Rising as the direct inspiration this project scales ('proves the pedagogy'); the project's overall structure, a turn-based, multi-seat (USA/China) roleplay simulation of AI race dynamics with policy cards and posture shifts (data/seats.json, parameters.json worldRules.rivalMoves), follows this paper's original Intelligence Rising format.
-- **SRC-WOUTERS-2013** · [Meta-analysis of serious-games learning outcomes](https://eric.ed.gov/?id=EJ1008015) · Wouters et al. (2013) · `design`
+- **SRC-WOUTERS-2013** · [Meta-analysis of serious-games learning outcomes](https://eric.ed.gov/?id=EJ1008015) · Wouters et al. (2013) · `empirical`
Shaped: The debrief screen (src/ui/screens/Debrief.tsx, spec'd in docs/DESIGN.md §8: takeaways surfaced, counterfactual hints, sources link) follows this meta-analysis's finding that serious games produce better learning outcomes when paired with explicit debriefing rather than left to stand alone.
-- **SRC-BOGOST-PERSUASIVE-GAMES** · Persuasive Games · Bogost (2007) · `design` *(book, obtain manually)*
+- **SRC-BOGOST-PERSUASIVE-GAMES** · Persuasive Games · Bogost (2007) · `analysis`
Shaped: The project's core Fun-First-70/30 principle (docs/DESIGN.md §3, operating principle 1: the lesson lives in mechanics, never in lecture screens) follows Bogost's concept of procedural rhetoric: games persuade through their rules and systems, not through explicit argument.
-## The library: further reading (98)
+## The library: further reading (96)
No usage claim. This is the shelf we read from and the shelf we recommend,
grouped by topic.
@@ -264,7 +264,7 @@ grouped by topic.
Why it is here: The deep-RL-specific companion finding to SRC-GOAL-MISGEN (the CoinRun agent that learns 'go right' instead of 'get the coin'); same relevance and same caveat as that entry.
- **SRC-ALIGNMENT-SURVEY** · [AI Alignment: A Comprehensive Survey](https://arxiv.org/abs/2310.19852) · Ji et al. (2023) · `analysis`
Why it is here: A broad taxonomy of alignment research techniques and open problems (RLHF, interpretability, scalable oversight) for players or teachers who want the academic map behind the game's alignment-difficulty dial.
-- **SRC-HUMAN-COMPATIBLE** · Human Compatible: AI and the Problem of Control · Stuart Russell, Viking/Penguin (2019) · `analysis` *(book, obtain manually)*
+- **SRC-HUMAN-COMPATIBLE** · Human Compatible: AI and the Problem of Control · Stuart Russell, Viking/Penguin (2019) · `analysis`
Why it is here: Russell's provably-beneficial-AI framing and the control problem, for readers wanting the classic popular-science argument behind the game's alignment mechanics.
### benchmark
@@ -327,7 +327,7 @@ grouped by topic.
- **SRC-GOLDMAN-GENAI-GDP** · [Generative AI could raise global GDP by 7%](https://www.goldmansachs.com/insights/articles/generative-ai-could-raise-global-gdp-by-7-percent) · Briggs & Kodnani, Goldman Sachs Research (2023) · `forecast`
Why it is here: The widely-cited optimistic GDP/jobs-exposure projection for generative AI, a useful counterweight for readers comparing it against the more skeptical Acemoglu macro estimate also in this registry.
-- **SRC-POWER-PROGRESS** · Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity · Acemoglu & Johnson, PublicAffairs (2023) · `analysis` *(book, obtain manually)*
+- **SRC-POWER-PROGRESS** · Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity · Acemoglu & Johnson, PublicAffairs (2023) · `analysis`
Why it is here: Acemoglu & Johnson's historical argument that a new technology's benefits depend on institutional choices, not inevitability, deeper reading behind the game's political-economy framing of automation and unrest.
- **SRC-ACEMOGLU-SIMPLE-MACRO** · [The Simple Macroeconomics of AI](https://www.nber.org/papers/w32487) · Acemoglu (2024) · `forecast`
Why it is here: A skeptical macro estimate of modest GDP growth from AI over the next decade, useful for a player weighing how much to trust either worldview's economic optimism against the Goldman Sachs projection also in this registry.
@@ -357,7 +357,7 @@ grouped by topic.
Why it is here: Tracks how much of the Draghi competitiveness agenda the EU has actually implemented (11.2% after one year per the EPIC Observatory), useful for a player or journalist checking whether a future EU seat reflects reality.
- **SRC-INVESTAI** · InvestAI / AI gigafactories & AI Action Summit outcomes (2025) · `empirical` *(flagged: An umbrella citation covering InvestAI, the AI gigafactories program and summit outcomes; no single official page covers all three.)*
Why it is here: The EU's own AI-infrastructure investment push (InvestAI funding, AI gigafactories, the Feb 2025 AI Action Summit outcomes), background for the 'Europe left behind' competitiveness story the game doesn't yet model with a playable EU seat.
-- **SRC-BRUSSELS-EFFECT** · The Brussels Effect · Bradford (2020) · `analysis` *(book, obtain manually)*
+- **SRC-BRUSSELS-EFFECT** · The Brussels Effect · Bradford (2020) · `analysis`
Why it is here: Bradford's thesis on how EU regulation projects power globally even without EU market presence, relevant background for why the game's EU AI Act wildcard events fire regardless of whether Europe is a playable seat, though those events cite EU-AI-Act sources directly rather than this book.
### export-controls
@@ -450,11 +450,6 @@ grouped by topic.
- **SRC-MILITARY-AI-REGULATION** · [Military AI Needs Technically-Informed Regulation to Safeguard AI Research and its Applications](https://arxiv.org/abs/2505.18371) · Riley Simmons-Edler et al. (2025) · `analysis`
Why it is here: Gives a reader a technically-grounded case for regulating military AI applications, a dedicated starter reference for the autonomous_targeting_demo and civil_military_fusion event cards beyond the AI-2027 and simulation-calibration sources currently cited there.
-### model-cards
-
-- **SRC-MODEL-DOCS-BUNDLE** · System/model cards (GPT-4, Claude 3/Opus, Gemini); Papers with Code; Alignment Forum / LessWrong key posts · `empirical` *(flagged: An umbrella entry for model release documentation across several labs; no single link to verify.)*
- Why it is here: For a researcher wanting primary model documentation and reproducible benchmark tracking behind the game's abstracted capability meter, this points to real published model cards (GPT-4, Claude, Gemini) and Papers with Code's benchmark leaderboards.
-
### policy-tracker
- **SRC-OECD-AI-OBSERVATORY** · [OECD.AI Policy Observatory](https://oecd.ai/en/) · OECD · `empirical`
@@ -480,7 +475,7 @@ grouped by topic.
Why it is here: For players curious about the economic-power-concentration argument behind the Societal Breakdown ending's premise, distinct from the IMF/OECD job-displacement data that actually drives that ending's mechanics in-game.
- **SRC-IAISR-UPDATE-1** · [International AI Safety Report — First Key Update: Capabilities & Risk Implications](https://arxiv.org/abs/2510.13653) · Bengio et al. (2025) · `analysis`
Why it is here: For readers who want the most recent (Oct 2025) expert-consensus update on frontier capability trends and risk implications, one step ahead of the primary IAISR report the consensus preset is built on.
-- **SRC-BOSTROM-SUPERINTELLIGENCE** · Superintelligence: Paths, Dangers, Strategies · Nick Bostrom, Oxford University Press (2014) · `analysis` *(book, obtain manually)*
+- **SRC-BOSTROM-SUPERINTELLIGENCE** · Superintelligence: Paths, Dangers, Strategies · Nick Bostrom, Oxford University Press (2014) · `analysis`
Why it is here: The foundational instrumental-convergence and orthogonality-thesis argument underlying the general existential-risk vocabulary echoed by the game's cautious worldview preset.
### risk-taxonomy
@@ -508,10 +503,8 @@ grouped by topic.
- **SRC-INTELLIGENCE-RISING-CSER** · [Intelligence Rising (project page + facilitation)](https://www.cser.ac.uk/work/intelligence-rising/) · CSER (Cambridge) · `design`
Why it is here: For a player or researcher curious about the real academic wargame this design descends from, CSER's project page is the primary hub for Intelligence Rising's facilitation materials and history.
-- **SRC-AVIN-AI-FUTURES-SURVEY** · [Exploring Artificial Intelligence Futures (survey of methods)](https://www.openbookpublishers.com/books/10.11647/obp.0360/chapters/10.11647/obp.0360.08) · Shahar Avin (2019) · `design`
+- **SRC-AVIN-AI-FUTURES-SURVEY** · [Exploring Artificial Intelligence Futures (survey of methods)](https://www.openbookpublishers.com/books/10.11647/obp.0360/chapters/10.11647/obp.0360.08) · Shahar Avin (2019) · `analysis`
Why it is here: For readers curious how game-based exploration of AI futures compares to narrative scenario-writing (like AI-2027) or survey forecasting, this is Avin's overview of the methodological menu the project's format sits within.
-- **SRC-SERIOUS-GAMES-BUNDLE** · RAND wargaming series; Schelling on gaming; Wouters et al. and Clark et al. serious-games meta-analyses; Bogost Persuasive Games; matrix-games & TTX methodology; climate-game studies (e.g. Daybreak) · `design` *(flagged: An umbrella entry for several works on learning through games; the individual works are listed in this registry, but there is no single link to verify.)*
- Why it is here: For a teacher or researcher wanting the wider serious-games/wargaming literature behind the debrief design (Schelling on gaming, matrix-game/TTX methodology, climate-game studies like Daybreak) beyond the four works already listed individually.
- **SRC-CLARK-2016** · [Digital Games for Learning meta-analysis](https://journals.sagepub.com/doi/10.3102/0034654315582065) · Clark et al. (2016) · `design`
Why it is here: For readers who want the empirical backing behind 'games can teach, but design matters,' this meta-analysis of digital game-based learning outcomes complements Wouters et al. (already grounding the debrief-screen design) with a broader review across digital games generally.
- **SRC-LOPEZ-FERNANDEZ-2024** · [Learning and Motivational Impact of Game-Based Learning: Comparing Face-to-Face and Online Formats on Computer Science Education](https://arxiv.org/abs/2407.07762) · Daniel López-Fernández et al. (2024) · `design`
@@ -533,7 +526,7 @@ grouped by topic.
### wargaming
-- **SRC-RAND-WARGAMING** · [RAND wargaming methodology series](https://www.rand.org) · RAND · `design` *(flagged: An umbrella entry for RAND's wargaming methodology literature; individual reports verify, the bundle as such has no single link.)*
+- **SRC-RAND-WARGAMING** · [RAND wargaming methodology series](https://www.rand.org/topics/wargaming.html) · RAND · `analysis` *(flagged: An umbrella entry for RAND's wargaming methodology literature; individual reports verify, the bundle as such has no single link.)*
Why it is here: For readers curious about the tabletop-wargaming methodology (matrix games, structured turn phases) that the project's turn/era/event structure resembles, RAND's wargaming-methodology literature is the field's reference; no single report is pinned here.
### worldview-optimistic
@@ -550,7 +543,7 @@ grouped by topic.
- **SRC-CAIS-STATEMENT** · [Statement on AI Risk](https://safe.ai/work/statement-on-ai-risk) · Center for AI Safety (CAIS) (2023) · `analysis`
Why it is here: The one-sentence expert-consensus statement that AI extinction risk should be a global priority, evidence for a player wanting to know the game's doom worldview preset isn't fringe speculation.
-- **SRC-PRECIPICE** · [The Precipice: Existential Risk and the Future of Humanity](https://theprecipice.com/) · Toby Ord (2020) · `analysis` *(book, obtain manually)*
+- **SRC-PRECIPICE** · [The Precipice: Existential Risk and the Future of Humanity](https://theprecipice.com/) · Toby Ord (2020) · `analysis`
Why it is here: Ord's canonical popular introduction to existential risk, a gentler companion to Carlsmith's technical paper for a reader building the case behind the game's doom-worldview preset.
- **SRC-COMPENDIUM** · [The Compendium](https://www.thecompendium.ai/) · Leahy, Alfour et al. (2024) · `analysis`
Why it is here: A risk-emphasizing synthesis of the AI existential-risk case for a player who finishes a cautious-worldview run wanting the fullest version of the doom argument in one place, alongside the game's already-cited Carlsmith and IABIED sources.
diff --git a/data/schemas/sources.schema.json b/data/schemas/sources.schema.json
index f4c7a53..e4b9f76 100644
--- a/data/schemas/sources.schema.json
+++ b/data/schemas/sources.schema.json
@@ -56,9 +56,7 @@
"type": "string",
"enum": [
"verified",
- "flagged",
- "book",
- "pending"
+ "flagged"
]
},
"tier": {
diff --git a/data/sources.json b/data/sources.json
index e0ac089..4512dc2 100644
--- a/data/sources.json
+++ b/data/sources.json
@@ -11,7 +11,7 @@
"tier": "load-bearing",
"evidenceClass": "design",
"gameUse": "design constants with no empirical referent (turn counts, starting defaults, allocation defaults)",
- "url": "https://github.com/chipmates/criticalwindow/blob/v0.3.3/docs/DESIGN.md"
+ "url": "https://github.com/chipmates/criticalwindow/blob/v0.3.4/docs/DESIGN.md"
},
{
"id": "SRC-KAPLAN-SCALING",
@@ -619,10 +619,11 @@
"year": 2014,
"type": "book",
"tags": ["risk", "worldview"],
- "status": "book",
+ "status": "verified",
"tier": "library",
"evidenceClass": "analysis",
- "whyListed": "The foundational instrumental-convergence and orthogonality-thesis argument underlying the general existential-risk vocabulary echoed by the game's cautious worldview preset."
+ "whyListed": "The foundational instrumental-convergence and orthogonality-thesis argument underlying the general existential-risk vocabulary echoed by the game's cautious worldview preset.",
+ "note": "Verified against the print edition."
},
{
"id": "SRC-HUMAN-COMPATIBLE",
@@ -632,10 +633,11 @@
"year": 2019,
"type": "book",
"tags": ["alignment", "worldview"],
- "status": "book",
+ "status": "verified",
"tier": "library",
"evidenceClass": "analysis",
- "whyListed": "Russell's provably-beneficial-AI framing and the control problem, for readers wanting the classic popular-science argument behind the game's alignment mechanics."
+ "whyListed": "Russell's provably-beneficial-AI framing and the control problem, for readers wanting the classic popular-science argument behind the game's alignment mechanics.",
+ "note": "Verified against the print edition."
},
{
"id": "SRC-GOVAI-COMPUTE",
@@ -783,10 +785,11 @@
"year": 2022,
"type": "book",
"tags": ["chip-war", "taiwan", "semiconductors", "supply-chain"],
- "status": "book",
+ "status": "verified",
"tier": "load-bearing",
"evidenceClass": "empirical",
- "gameUse": "TSMC/Taiwan chokepoint background"
+ "gameUse": "TSMC/Taiwan chokepoint background",
+ "note": "Verified against the print edition."
},
{
"id": "SRC-MAIM",
@@ -1086,10 +1089,11 @@
"year": 2020,
"type": "book",
"tags": ["europe", "eu-regulation", "brussels-effect", "sovereignty"],
- "status": "book",
+ "status": "verified",
"tier": "library",
"evidenceClass": "analysis",
- "whyListed": "Bradford's thesis on how EU regulation projects power globally even without EU market presence, relevant background for why the game's EU AI Act wildcard events fire regardless of whether Europe is a playable seat, though those events cite EU-AI-Act sources directly rather than this book."
+ "whyListed": "Bradford's thesis on how EU regulation projects power globally even without EU market presence, relevant background for why the game's EU AI Act wildcard events fire regardless of whether Europe is a playable seat, though those events cite EU-AI-Act sources directly rather than this book.",
+ "note": "Verified against the print edition."
},
{
"id": "SRC-IMF-GENAI",
@@ -1168,10 +1172,11 @@
"year": 2023,
"type": "book",
"tags": ["economy", "political-economy", "automation"],
- "status": "book",
+ "status": "verified",
"tier": "library",
"evidenceClass": "analysis",
- "whyListed": "Acemoglu & Johnson's historical argument that a new technology's benefits depend on institutional choices, not inevitability, deeper reading behind the game's political-economy framing of automation and unrest."
+ "whyListed": "Acemoglu & Johnson's historical argument that a new technology's benefits depend on institutional choices, not inevitability, deeper reading behind the game's political-economy framing of automation and unrest.",
+ "note": "Verified against the print edition."
},
{
"id": "SRC-GPTS-ARE-GPTS",
@@ -1388,10 +1393,11 @@
"url": "https://theprecipice.com/",
"type": "book",
"tags": ["x-risk", "worldview-doom"],
- "status": "book",
+ "status": "verified",
"tier": "library",
"evidenceClass": "analysis",
- "whyListed": "Ord's canonical popular introduction to existential risk, a gentler companion to Carlsmith's technical paper for a reader building the case behind the game's doom-worldview preset."
+ "whyListed": "Ord's canonical popular introduction to existential risk, a gentler companion to Carlsmith's technical paper for a reader building the case behind the game's doom-worldview preset.",
+ "note": "Verified against the print edition."
},
{
"id": "SRC-IABIED",
@@ -1401,10 +1407,11 @@
"url": "https://ifanyonebuildsit.com/",
"type": "book",
"tags": ["x-risk", "worldview-doom", "misaligned-catastrophe"],
- "status": "book",
+ "status": "verified",
"tier": "load-bearing",
"evidenceClass": "analysis",
- "gameUse": "Grounds the cautious worldview's high alignment-difficulty range (upper half of the 550-950 band) and, via the book's core policy prescription of a halt on frontier AI development, the global_moratorium policy card's pause mechanic."
+ "gameUse": "Grounds the cautious worldview's high alignment-difficulty range (upper half of the 550-950 band) and, via the book's core policy prescription of a halt on frontier AI development, the global_moratorium policy card's pause mechanic.",
+ "note": "Verified against the print edition."
},
{
"id": "SRC-SNAKE-OIL",
@@ -1414,11 +1421,11 @@
"year": 2024,
"type": "book",
"tags": ["worldview-skeptic", "benefit-scenario"],
- "status": "book",
+ "status": "verified",
"tier": "load-bearing",
"evidenceClass": "analysis",
"gameUse": "skeptical worldview",
- "note": "companion site per the founding source list: https://www.aisnakeoil.com"
+ "note": "companion site per the founding source list: https://www.aisnakeoil.com. Verified against the print edition."
},
{
"id": "SRC-COMPENDIUM",
@@ -1564,7 +1571,7 @@
"tags": ["serious-games", "ai-race", "wargaming", "facilitation"],
"status": "verified",
"tier": "load-bearing",
- "evidenceClass": "design",
+ "evidenceClass": "analysis",
"gameUse": "facilitator insights from 43 Intelligence Rising games",
"note": "ScienceDirect version: https://www.sciencedirect.com/science/article/abs/pii/S0016328725000254."
},
@@ -1591,21 +1598,9 @@
"tags": ["serious-games", "ai-futures", "forecasting-methods", "roleplay"],
"status": "verified",
"tier": "library",
- "evidenceClass": "design",
+ "evidenceClass": "analysis",
"whyListed": "For readers curious how game-based exploration of AI futures compares to narrative scenario-writing (like AI-2027) or survey forecasting, this is Avin's overview of the methodological menu the project's format sits within."
},
- {
- "id": "SRC-SERIOUS-GAMES-BUNDLE",
- "title": "RAND wargaming series; Schelling on gaming; Wouters et al. and Clark et al. serious-games meta-analyses; Bogost Persuasive Games; matrix-games & TTX methodology; climate-game studies (e.g. Daybreak)",
- "type": "web",
- "tags": ["serious-games", "wargaming", "ttx-methodology", "debrief-design"],
- "status": "flagged",
- "tier": "library",
- "evidenceClass": "design",
- "whyListed": "For a teacher or researcher wanting the wider serious-games/wargaming literature behind the debrief design (Schelling on gaming, matrix-game/TTX methodology, climate-game studies like Daybreak) beyond the four works already listed individually.",
- "flagReason": "An umbrella entry for several works on learning through games; the individual works are listed in this registry, but there is no single link to verify.",
- "note": "Umbrella entry. Wouters et al. 2013, Clark et al. 2016, Bogost's Persuasive Games and the RAND wargaming series have their own entries; this bundle remains for the still-unpinned items: Schelling on gaming, matrix-game/TTX methodology, and climate-game studies such as Daybreak."
- },
{
"id": "SRC-WOUTERS-2013",
"title": "Meta-analysis of serious-games learning outcomes",
@@ -1616,7 +1611,7 @@
"tags": ["serious-games", "learning-outcomes", "pedagogy", "meta-analysis"],
"status": "verified",
"tier": "background",
- "evidenceClass": "design",
+ "evidenceClass": "empirical",
"shaped": "The debrief screen (src/ui/screens/Debrief.tsx, spec'd in docs/DESIGN.md §8: takeaways surfaced, counterfactual hints, sources link) follows this meta-analysis's finding that serious games produce better learning outcomes when paired with explicit debriefing rather than left to stand alone.",
"note": "No URL in the original reference; link added at the July 2026 verification. See also SRC-SERIOUS-GAMES-BUNDLE."
},
@@ -1641,25 +1636,25 @@
"year": 2007,
"type": "book",
"tags": ["serious-games", "game-design", "persuasion", "pedagogy"],
- "status": "book",
+ "status": "verified",
"tier": "background",
- "evidenceClass": "design",
+ "evidenceClass": "analysis",
"shaped": "The project's core Fun-First-70/30 principle (docs/DESIGN.md §3, operating principle 1: the lesson lives in mechanics, never in lecture screens) follows Bogost's concept of procedural rhetoric: games persuade through their rules and systems, not through explicit argument.",
- "note": "Book, obtain manually. See also SRC-SERIOUS-GAMES-BUNDLE."
+ "note": "Book, obtain manually. See also SRC-SERIOUS-GAMES-BUNDLE. Verified against the print edition."
},
{
"id": "SRC-RAND-WARGAMING",
"title": "RAND wargaming methodology series",
"org": "RAND",
- "url": "https://www.rand.org",
+ "url": "https://www.rand.org/topics/wargaming.html",
"type": "report",
"tags": ["wargaming", "methodology", "ttx", "serious-games"],
"status": "flagged",
"tier": "library",
- "evidenceClass": "design",
+ "evidenceClass": "analysis",
"whyListed": "For readers curious about the tabletop-wargaming methodology (matrix games, structured turn phases) that the project's turn/era/event structure resembles, RAND's wargaming-methodology literature is the field's reference; no single report is pinned here.",
"flagReason": "An umbrella entry for RAND's wargaming methodology literature; individual reports verify, the bundle as such has no single link.",
- "note": "URL is the RAND homepage; search 'wargaming' on rand.org for specific reports. July 2026 check: rand.org blocks automated fetches, but the methodology publications exist (e.g. rand.org/pubs/research_memoranda/RM2413.html). See also SRC-SERIOUS-GAMES-BUNDLE."
+ "note": "RAND's wargaming topic hub; no single report is pinned, which keeps this on the shelf tier."
},
{
"id": "SRC-AI-INDEX-2025",
@@ -1782,18 +1777,6 @@
"whyListed": "For a classroom teacher (a stated project audience) wanting a structured follow-on curriculum after playing, BlueDot's AI Safety Fundamentals courses are the standard on-ramp into technical alignment and governance study.",
"note": "Courses page: https://bluedot.org/courses/alignment. Matches the founding source list entry 'BlueDot / AI Safety Fundamentals curricula'."
},
- {
- "id": "SRC-MODEL-DOCS-BUNDLE",
- "title": "System/model cards (GPT-4, Claude 3/Opus, Gemini); Papers with Code; Alignment Forum / LessWrong key posts",
- "type": "web",
- "tags": ["model-cards", "documentation", "live-data"],
- "status": "flagged",
- "tier": "library",
- "evidenceClass": "empirical",
- "whyListed": "For a researcher wanting primary model documentation and reproducible benchmark tracking behind the game's abstracted capability meter, this points to real published model cards (GPT-4, Claude, Gemini) and Papers with Code's benchmark leaderboards.",
- "flagReason": "An umbrella entry for model release documentation across several labs; no single link to verify.",
- "note": "Bundled placeholder entry (marked verify-individual-URLs, no single URL given). Alignment Forum is captured individually as SRC-ALIGNMENT-FORUM via a clean atomic URL from the founding source list Live data & community; System/model cards and Papers with Code remain unresolved individual citations. Bundle entry, no single URL by design."
- },
{
"id": "SRC-ALIGNMENT-FORUM",
"title": "Alignment Forum",
diff --git a/data/strings/en.json b/data/strings/en.json
index b79e458..baa8480 100644
--- a/data/strings/en.json
+++ b/data/strings/en.json
@@ -62,7 +62,7 @@
"ticker.late.18": "Ministers debate whether to feel reassured. Vote postponed.",
"ticker.late.19": "Somewhere a safety researcher sleeps eight hours. Colleagues alarmed.",
"preset.cautious.label": "Cautious World",
- "preset.cautious.description": "Alignment is probably hard and takeoff could be fast. Grounded in the risk-focused literature.",
+ "preset.cautious.description": "Alignment is probably hard and takeoff could be fast. Grounded in the risk-focused literature. Hard mode: in this world even careful governance usually fails.",
"preset.consensus.label": "Consensus World",
"preset.consensus.description": "Deep expert disagreement and wide uncertainty. Grounded in surveys and the international safety report.",
"preset.skeptic.label": "Skeptic World",
@@ -332,10 +332,13 @@
"a11y.runEnded": "The run has ended: {ending}",
"a11y.raceTrack": "Race track: your capability {you} of 1000, rival {rival} of 1000, threshold fog begins at {fog}",
"a11y.skipToGame": "Skip to game controls",
+ "a11y.more": "More",
+ "a11y.close": "Close",
"footer.privacy": "No accounts. No tracking. Saves live in your browser.",
"footer.feedback": "Found something wrong? Send a correction.",
"report.kind.upkeep": "Upkeep",
"report.kind.delayedEffect": "Delayed consequence: {source}",
+ "effect.delayed": "in {turns} turns: {effects}",
"report.kind.policyPlayed": "Policy enacted: {source}",
"report.kind.eventResolved": "Memo decided: {source}",
"report.kind.societyUpdate": "Society",
@@ -511,7 +514,7 @@
"prologue.ch1.explainer": "This slider is your main lever, every quarter: split R&D between Capability, Safety and Diffusion. Try moving it. In the prologue nothing breaks.",
"prologue.ch2.date": "2024 to January 2025",
"prologue.ch2.title": "The Chokepoint War",
- "prologue.ch2.body": "Washington chokes the chip pipeline to slow the rival. It works for a little while, and only briefly. Then a lab nobody was watching ships a frontier-grade model trained on a shoestring, and the gap closes anyway. Controls can slow the race but never stop it.",
+ "prologue.ch2.body": "Washington chokes the chip pipeline to slow the rival. It works for a little while, and only briefly. Then a lab most of the public had never heard of ships a frontier-grade model trained on a shoestring, and the gap closes anyway. Controls can slow the race but never stop it.",
"prologue.ch2.explainer": "Policies are your strong single moves: one card a quarter, real costs, and bites that arrive later. This one taught the rival to build their own chips.",
"prologue.ch3.date": "2025 to summer 2026",
"prologue.ch3.title": "The Rules Arrive",
@@ -539,7 +542,7 @@
"handoff.heading": "Hand the device over",
"handoff.body": "{seat} is up next. The eval band and hand you just saw belong to the other seat. No peeking is the whole game.",
"handoff.continue": "I am {seat}. Begin.",
- "title.hook": "Summer 2026. You take office holding the AI race portfolio. Sixteen quarters to 2030. Your eval reports are honest ranges, not answers. Your rival is racing, your public is nervous, and two dice were sealed in an envelope before you sat down.",
+ "title.hook": "Summer 2026. You take office holding the AI race portfolio. Sixteen quarters to 2030. Your eval reports are honest ranges, not answers. Your rival is racing, your public is nervous, and two dice were sealed in an envelope before you sat down. Most terms end badly. The debrief is the payoff.",
"title.feature.seats": "Play the US or China. Or both, one device",
"title.feature.sources": "Every number: a source you can check, or a design choice that says so",
"title.feature.endings": "Five endings. The good ones are earned",
@@ -790,7 +793,7 @@
"sources.legend": "Every entry was re-checked at the July 2026 audit, by machine and by hand. A few carry a note where no single stable link exists. The note says exactly why.",
"title.staleSave": "Your saved run was made with an older data version and cannot resume. Its seed was {seed}. Start a fresh run with it to replay that world.",
"sources.dataVersion": "Build {sha} · data version {hash}. Compare either against the repository to verify this deployment.",
- "title.soundOffer": "This game has music and a narrator, off by default out of respect. Want them?",
+ "title.soundOffer": "This game has music and a narrator, off by default. Want them?",
"title.soundOn": "Sound on",
"title.soundNotNow": "Not now",
"anchor.alignment.0": "actively gaming its own training",
diff --git a/docs/BALANCE.md b/docs/BALANCE.md
index 56c58ca..ed4436a 100644
--- a/docs/BALANCE.md
+++ b/docs/BALANCE.md
@@ -210,7 +210,7 @@ strategy directly: bank alignment before the sprint, read the incidents as align
data, time the treaty signature to the window. A model told the answer and then graded on
whether it repeated the answer is not being tested on discovery, it's being tested on
obedience, and obedience is not what balance needs to know. The fix
-([`4746508`](https://github.com/chipmates/criticalwindow/commit/4746508)) added a second,
+([`14b08c3`](https://github.com/chipmates/criticalwindow/commit/14b08c3)) added a second,
mechanics-only rulebook with every strategy sentence stripped out, plus a mode where the
model plays a series of games carrying forward nothing but its own notes, revised after
each debrief, which is exactly what a human carries between runs.
@@ -222,13 +222,13 @@ gets far more than that: a chart of what the evals said against what was actuall
each quarter, whether the treaty window was open and when, which incidents fired and
whether they were raced past. Grading a model on what it learned from a debrief it was
never actually shown is the same mistake as the other two, worn differently. Fixed in
-[`a8ececd`](https://github.com/chipmates/criticalwindow/commit/a8ececd) by replaying every
+[`d3f15da`](https://github.com/chipmates/criticalwindow/commit/d3f15da) by replaying every
finished run through the same probes that generate the debrief screen itself, so the
model's between-game notes are built from the same information a human closes a run with.
Once the model could actually see what a human sees, the campaign's clearest skill signal
showed up: bots that could not see the rival's ratification status (added to the model's
-view alongside the fix above, [`4e1bd27`](https://github.com/chipmates/criticalwindow/commit/4e1bd27))
+view alongside the fix above, [`a53c06c`](https://github.com/chipmates/criticalwindow/commit/a53c06c))
succeeded at reaching a good ending about 26 percent of the time. Bots given that same
"diplomatic weather" line the human UI already shows succeeded around 60 percent of the
time. The treaty was never really about knowing the mechanic exists. It was about reading
diff --git a/docs/DESIGN.md b/docs/DESIGN.md
index f77bcb5..6ad56cb 100644
--- a/docs/DESIGN.md
+++ b/docs/DESIGN.md
@@ -51,7 +51,7 @@ Europe appears as a force in the world (regulatory beats, the ASML chokepoint) b
## 7. Endings
-Five endings, plus a sixth planned. Flourishing, where aligned transformative AI arrives and the benefits actually diffuse, is genuinely reachable and deliberately rare on a first run. Misaligned catastrophe, where somebody raced past the uncertainty band and lost the gamble. Outpaced. Negotiated slowdown, a verified treaty, slower and politically expensive. Societal breakdown, where trust and jobs collapsed before AI resolved anything. The sixth is a hidden ending about losing without noticing; its groundwork already accrues quietly during play, its evidence base is cited in the registry, and it ships once it can be discovered properly rather than announced.
+Five endings, plus a sixth planned. Flourishing, where aligned transformative AI arrives and the benefits actually diffuse, is genuinely reachable and deliberately rare on a first run. Misaligned catastrophe, where somebody raced past the uncertainty band and lost the gamble. Outpaced. Negotiated slowdown, a verified treaty, slower and politically expensive. Societal breakdown, where trust and jobs collapsed before AI resolved anything. A sixth ending is planned; its groundwork already accrues in the engine.
Doom fatigue produces fatalism, which is the opposite of the mission. The good endings are earned, not promised.
diff --git a/docs/EVIDENCE.md b/docs/EVIDENCE.md
index b3814fb..fd3f2df 100644
--- a/docs/EVIDENCE.md
+++ b/docs/EVIDENCE.md
@@ -13,9 +13,9 @@ otherwise. Run it yourself.
**76 cited parameter values and 100 cited card premises, across 334 citation sites.** Parameter values by kind:
-- **9** analysis-based
-- **23** anchored, game-calibrated
-- **29** design constant
+- **13** analysis-based
+- **21** anchored, game-calibrated
+- **27** design constant
- **15** forecast-based
The vocabulary claims exactly what is true. **Empirically anchored** means the
@@ -42,8 +42,8 @@ The tables below are complete; these five show the transformation step by step.
training capacity, the cited compute dataset's picture of the US share; the index
point is the mapping, not a measurement of its own.
2. **Capability pacing (curves.capabilityPerRnd).** The cited analyses report training
- compute doubling roughly every six months and the 50% task horizon doubling every
- 188 days on the all-time fit. The curve compresses that regime onto the 0-1000
+ compute doubling roughly every six months and the 50% task horizon P50 doubling every
+ 196.5 days on the all-time fit (130.8 days for models from 2023 on). The curve compresses that regime onto the 0-1000
index so an all-in racer reaches threshold resolution around turn 10 of 16; the
compression ratio is the design decision, and it cites the constitution.
3. **Eval band floor (evalUncertainty.floorBandWidth 100).** The deception literature
@@ -89,7 +89,7 @@ See [`GOVERNANCE.md`](../GOVERNANCE.md).
|---|---|---|
| events/allied_export_fracture.json → (root) | empirically anchored premise | SRC-CSIS-SME-CONTROLS, SRC-CSIS-EXPORT |
| events/ally_eval_request.json → (root) | analysis-based premise | SRC-HO-INTL-INSTITUTIONS, SRC-CSIS-EXPORT |
-| events/asml_export_squeeze.json → (root) | anchored, game-calibrated premise | SRC-CSIS-EXPORT, SRC-CHIP-WAR, SRC-SIM-GAMING-INSIGHTS |
+| events/asml_export_squeeze.json → (root) | analysis-based premise | SRC-CSIS-EXPORT, SRC-CHIP-WAR, SRC-SIM-GAMING-INSIGHTS |
| events/autonomous_targeting_demo.json → (root) | forecast-based premise | SRC-AI2027, SRC-SIM-GAMING-INSIGHTS |
| events/benchmark_gaming.json → (root) | empirically anchored premise | SRC-BENCH-COP, SRC-SCHEMING |
| events/branch_point.json → (root) | forecast-based premise | SRC-AI2027 |
@@ -119,7 +119,7 @@ See [`GOVERNANCE.md`](../GOVERNANCE.md).
| events/rival_breakthrough.json → (root) | forecast-based premise | SRC-AI2027, SRC-MAIM |
| events/shuttered_factory_census.json → (root) | empirically anchored premise | SRC-CSIS-EXPORT, SRC-OECD-EMPLOYMENT |
| events/superhuman_coder.json → (root) | forecast-based premise | SRC-AI2027, SRC-SITUATIONAL-AWARENESS, SRC-GRACE-SURVEY |
-| events/taiwan_strait_crisis.json → (root) | anchored, game-calibrated premise | SRC-CHIP-WAR, SRC-CSIS-EXPORT, SRC-SIM-GAMING-INSIGHTS |
+| events/taiwan_strait_crisis.json → (root) | analysis-based premise | SRC-CHIP-WAR, SRC-CSIS-EXPORT, SRC-SIM-GAMING-INSIGHTS |
| events/the_project.json → (root) | forecast-based premise | SRC-SITUATIONAL-AWARENESS, SRC-SIM-GAMING-INSIGHTS |
| events/treaty_feeler.json → (root) | analysis-based premise | SRC-MAIM, SRC-HO-INTL-INSTITUTIONS |
| events/union_moratorium_demand.json → (root) | empirically anchored premise | SRC-IMF-GENAI, SRC-OECD-EMPLOYMENT |
@@ -132,9 +132,9 @@ See [`GOVERNANCE.md`](../GOVERNANCE.md).
| Where | Premise kind | Sources |
|---|---|---|
-| (root) | anchored, game-calibrated premise | SRC-SLEEPER, SRC-SCHEMING, SRC-AGENTIC-MISALIGNMENT, SRC-AI-INCIDENT-DB, SRC-SIM-GAMING-INSIGHTS |
+| (root) | analysis-based premise | SRC-SLEEPER, SRC-SCHEMING, SRC-AGENTIC-MISALIGNMENT, SRC-AI-INCIDENT-DB, SRC-SIM-GAMING-INSIGHTS |
| riskFormula.pressureAllocationPct | empirically anchored premise | SRC-AGENTIC-MISALIGNMENT |
-| riskFormula.pressureRivalRacePct | anchored, game-calibrated premise | SRC-AGENTIC-MISALIGNMENT, SRC-SIM-GAMING-INSIGHTS, SRC-DESIGN-HANDOVER |
+| riskFormula.pressureRivalRacePct | analysis-based premise | SRC-AGENTIC-MISALIGNMENT, SRC-SIM-GAMING-INSIGHTS, SRC-DESIGN-HANDOVER |
| safetyInsightDamageReductionMaxPerMille | empirically anchored premise | SRC-SLEEPER, SRC-SCHEMING |
| rungs[0] | empirically anchored premise | SRC-SLEEPER, SRC-AGENTIC-MISALIGNMENT, SRC-SCHEMING |
| rungs[1] | empirically anchored premise | SRC-AGENTIC-MISALIGNMENT, SRC-SCHEMING |
@@ -145,7 +145,7 @@ See [`GOVERNANCE.md`](../GOVERNANCE.md).
| Where | Premise kind | Sources |
|---|---|---|
-| (root) | design constant premise | SRC-SIM-GAMING-INSIGHTS, SRC-DESIGN-HANDOVER |
+| (root) | analysis-based premise | SRC-SIM-GAMING-INSIGHTS, SRC-DESIGN-HANDOVER |
| mandates[0] | empirically anchored premise | SRC-IEA-ENERGY-AI, SRC-GRID-DATACENTER |
| mandates[1] | empirically anchored premise | SRC-IMF-GENAI, SRC-OECD-EMPLOYMENT |
| mandates[2] | empirically anchored premise | SRC-UK-AISI, SRC-SLEEPER |
@@ -154,8 +154,8 @@ See [`GOVERNANCE.md`](../GOVERNANCE.md).
| mandates[5] | empirically anchored premise | SRC-IEA-ENERGY-AI |
| mandates[6] | empirically anchored premise | SRC-IMF-GENAI, SRC-OECD-EMPLOYMENT |
| mandates[7] | empirically anchored premise | SRC-SLEEPER, SRC-SCHEMING |
-| mandates[8] | design constant premise | SRC-SIM-GAMING-INSIGHTS, SRC-DESIGN-HANDOVER |
-| mandates[9] | anchored, game-calibrated premise | SRC-CSIS-EXPORT, SRC-SIM-GAMING-INSIGHTS |
+| mandates[8] | analysis-based premise | SRC-SIM-GAMING-INSIGHTS, SRC-DESIGN-HANDOVER |
+| mandates[9] | analysis-based premise | SRC-CSIS-EXPORT, SRC-SIM-GAMING-INSIGHTS |
## parameters.json
@@ -267,8 +267,8 @@ See [`GOVERNANCE.md`](../GOVERNANCE.md).
| seats.china.resources.energy | empirically anchored premise | SRC-IEA-ENERGY-AI |
| seats.china.resources.talent | anchored, game-calibrated premise | SRC-AI-INDEX-2026, SRC-DESIGN-HANDOVER |
| seats.china.resources.capital | anchored, game-calibrated premise | SRC-AI-INDEX-2026, SRC-DESIGN-HANDOVER |
-| seats.china.resources.publicTrust | design constant premise | SRC-SIM-GAMING-INSIGHTS, SRC-DESIGN-HANDOVER |
-| seats.china.resources.politicalCapital | design constant premise | SRC-SIM-GAMING-INSIGHTS, SRC-DESIGN-HANDOVER |
+| seats.china.resources.publicTrust | analysis-based premise | SRC-SIM-GAMING-INSIGHTS, SRC-DESIGN-HANDOVER |
+| seats.china.resources.politicalCapital | analysis-based premise | SRC-SIM-GAMING-INSIGHTS, SRC-DESIGN-HANDOVER |
| seats.china.resources.capability | empirically anchored premise | SRC-DEEPSEEK-COST |
| seats.china.resources.safetyInsight | analysis-based premise | SRC-IAISR, SRC-DESIGN-HANDOVER |
| seats.china.society.jobDisplacement | empirically anchored premise | SRC-IMF-GENAI |
@@ -280,8 +280,8 @@ See [`GOVERNANCE.md`](../GOVERNANCE.md).
| Where | Numbers | Kind | Sources | How the number was derived |
|---|---|---|---|---|
-| (root) | | anchored, game-calibrated | SRC-SIM-GAMING-INSIGHTS, SRC-CSIS-EXPORT | Shared seat rules: the asymmetries encode the facilitation literature and export-control reality cited; the magnitudes themselves are design-tuned for balance. |
-| usa | | design constant | SRC-SIM-GAMING-INSIGHTS, SRC-DESIGN-HANDOVER | design constant; the constitution is the derivation |
-| china | | anchored, game-calibrated | SRC-SIM-GAMING-INSIGHTS, SRC-CSIS-EXPORT, SRC-DESIGN-HANDOVER | Legitimacy cadence and the substitution-gated compute door follow the cited analyses of the China seat asymmetry; magnitudes design-tuned. |
-| china.legitimacyCheck.trustMin | value 400 | design constant | SRC-SIM-GAMING-INSIGHTS, SRC-DESIGN-HANDOVER | legitimacy pressure replaces the midterm: verdicts at each era turn instead of one election |
+| (root) | | analysis-based | SRC-SIM-GAMING-INSIGHTS, SRC-CSIS-EXPORT | Shared seat rules: the asymmetries encode the facilitation literature and export-control reality cited; the magnitudes themselves are design-tuned for balance. |
+| usa | | analysis-based | SRC-SIM-GAMING-INSIGHTS, SRC-DESIGN-HANDOVER | ⚠ derivation note missing |
+| china | | analysis-based | SRC-SIM-GAMING-INSIGHTS, SRC-CSIS-EXPORT, SRC-DESIGN-HANDOVER | Legitimacy cadence and the substitution-gated compute door follow the cited analyses of the China seat asymmetry; magnitudes design-tuned. |
+| china.legitimacyCheck.trustMin | value 400 | analysis-based | SRC-SIM-GAMING-INSIGHTS, SRC-DESIGN-HANDOVER | legitimacy pressure replaces the midterm: verdicts at each era turn instead of one election |
| china.legitimacyCheck.swing | value 100 | design constant | SRC-DESIGN-HANDOVER | political capital gained per era while legitimacy holds, lost while it does not |
diff --git a/docs/FAQ.md b/docs/FAQ.md
index a22d062..077dc4a 100644
--- a/docs/FAQ.md
+++ b/docs/FAQ.md
@@ -1,15 +1,16 @@
# FAQ
-The questions people actually ask, answered plainly.
+The questions we expect, and the ones playtesting raised, answered plainly.
## Is it fun?
Sessions during development kept forcing rebalances, which is what fun pressure looks
like from the inside: a player loses in a way that feels wrong, argues, and the
mechanism has to answer (the alignment-earned model and the hard-crossing rebalance
-both started that way; the commits are in [BALANCE.md](BALANCE.md)). Runs take 20 to
-40 minutes, losing is common, and the debrief is built to make losing worth it. The
-public alpha exists to test that against people who did not build the game.
+both started that way; the commits are in [BALANCE.md](BALANCE.md)). Playtesting so
+far is mostly bots plus a handful of humans, so treat this as a claim under test. Runs
+take 20 to 40 minutes, losing is common, and the debrief is built to make losing worth
+it. The public alpha exists to test that against people who did not build the game.
## Is this a game or advocacy?
@@ -63,6 +64,5 @@ account? Email criticalwindow@chipmates.ai; the debrief links it too.
## Where is the sixth ending?
-Planned, not shipped, and not secret: its groundwork already accrues during play and
-its evidence base sits in the registry. It ships when discovering it can feel like
-discovering it.
+Planned, not shipped. Its groundwork already accrues during play, and it ships after
+the alpha, once it can land unspoiled.
diff --git a/docs/decisions/ADR-003-lab-npcs-deferred.md b/docs/decisions/ADR-003-lab-npcs-deferred.md
new file mode 100644
index 0000000..06a4f88
--- /dev/null
+++ b/docs/decisions/ADR-003-lab-npcs-deferred.md
@@ -0,0 +1,20 @@
+# ADR-003: Frontier-lab NPCs deferred
+
+Status: **ACCEPTED 2026-07-06** by the maintainer.
+
+Founding decision #8 in [ADR-001](ADR-001-founding-decisions.md) promised frontier
+labs as semi-autonomous NPCs inside each bloc, because the principal-agent friction
+between governments and labs is itself a lesson. ADR-002 amended the seat half of
+that decision and left the labs clause hanging without a record, which broke our own
+rule that changing a founding decision takes a written argument. This is that record.
+
+The labs clause is deferred, not dropped. In the shipped game, labs exist as event
+actors (mergers, leaks, whistleblowers, capability gifts) but not as agents with
+their own allocation behavior. The reason is scope honesty: a lab layer that acts
+autonomously needs its own balance campaign, and bolting a shallow version onto the
+alpha would teach the principal-agent lesson badly, which is worse than not teaching
+it yet. The pressure the clause was meant to model reaches the player today through
+the event deck and the incident system instead.
+
+Standing review: after the alpha's human-playtest phase, alongside the sixth ending.
+If the lab layer ships, it gets its own ADR with the balance evidence attached.
diff --git a/docs/media/playthrough.gif b/docs/media/playthrough.gif
index 6e3be0d..c83ddd2 100644
Binary files a/docs/media/playthrough.gif and b/docs/media/playthrough.gif differ
diff --git a/docs/media/title.png b/docs/media/title.png
index 60ff6ef..8f2006e 100644
Binary files a/docs/media/title.png and b/docs/media/title.png differ
diff --git a/index.html b/index.html
index 5addb81..b0b23e6 100644
--- a/index.html
+++ b/index.html
@@ -36,7 +36,7 @@
"@type": "VideoGame",
"name": "Critical Window",
"url": "https://criticalwindow.org/",
- "description": "A free, open source strategy game about the AI race. Govern the US or China through 16 quarters to 2030 under honest uncertainty: eval reports are ranges, alignment cannot be verified directly, and every parameter is source-backed or a labeled design choice, checkable against the 164-entry source registry in the open repository.",
+ "description": "A free, open source strategy game about the AI race. Govern the US or China through 16 quarters to 2030 under honest uncertainty: eval reports are ranges, alignment cannot be verified directly, and every parameter is source-backed or a labeled design choice; 54 external sources drive the numbers directly, and the full registry ships with the game.",
"genre": ["Strategy", "Simulation", "Educational"],
"gamePlatform": "Web browser",
"playMode": ["SinglePlayer", "MultiPlayer"],
@@ -85,17 +85,23 @@
Critical Window
a memo with no clean options, and then the world moves. Your alignment report comes back
as a range, not a number, and that is the game.
+
+ The build fails if a number loses its citation. Whether a source truly supports a value
+ stays human judgment, and challenging a number with a better source is the contribution we
+ most want. Build provenance:
+ version.json.
+
diff --git a/package.json b/package.json
index 9d87e98..7d0290e 100644
--- a/package.json
+++ b/package.json
@@ -1,7 +1,7 @@
{
"name": "criticalwindow",
"private": true,
- "version": "0.3.3",
+ "version": "0.3.4",
"homepage": "https://criticalwindow.org",
"repository": {
"type": "git",
diff --git a/public/sitemap.xml b/public/sitemap.xml
index ba767ef..47e6431 100644
--- a/public/sitemap.xml
+++ b/public/sitemap.xml
@@ -5,4 +5,6 @@
weekly1.0
+ https://criticalwindow.org/imprint.html
+ https://criticalwindow.org/privacy.html
diff --git a/scripts/lib/registry-honesty.ts b/scripts/lib/registry-honesty.ts
index da3eceb..1b1dff4 100644
--- a/scripts/lib/registry-honesty.ts
+++ b/scripts/lib/registry-honesty.ts
@@ -1,7 +1,9 @@
/**
* The registry-honesty rules, extracted so the test suite can mutation-test
* them: a tier is declared in sources.json but proven by the citation map,
- * and usage claims must be specific enough to check.
+ * and usage claims must be specific enough to check. The vague-phrase lint here
+ * is deliberately blunt: it catches the worst offenders, and human review
+ * catches the rest. Nobody should mistake a blocklist for understanding.
*/
import type { SourcesRegistryData } from '../../src/engine/schemas';
import type { UsageMap } from './source-usage';
diff --git a/scripts/lib/render-views.ts b/scripts/lib/render-views.ts
index b60e3eb..ed0622d 100644
--- a/scripts/lib/render-views.ts
+++ b/scripts/lib/render-views.ts
@@ -20,9 +20,7 @@ const TIER_HEADING: Record = {
const STATUS_BADGE: Record = {
verified: '',
- pending: ' *(verification pending)*',
flagged: ' *(link broken, replacement wanted)*',
- book: ' *(book, obtain manually)*',
};
/** Max citation sites listed per entry before folding into "+N more". */
@@ -379,8 +377,8 @@ The tables below are complete; these five show the transformation step by step.
training capacity, the cited compute dataset's picture of the US share; the index
point is the mapping, not a measurement of its own.
2. **Capability pacing (curves.capabilityPerRnd).** The cited analyses report training
- compute doubling roughly every six months and the 50% task horizon doubling every
- 188 days on the all-time fit. The curve compresses that regime onto the 0-1000
+ compute doubling roughly every six months and the 50% task horizon P50 doubling every
+ 196.5 days on the all-time fit (130.8 days for models from 2023 on). The curve compresses that regime onto the 0-1000
index so an all-in racer reaches threshold resolution around turn 10 of 16; the
compression ratio is the design decision, and it cites the constitution.
3. **Eval band floor (evalUncertainty.floorBandWidth 100).** The deception literature
diff --git a/scripts/validate-data.ts b/scripts/validate-data.ts
index faf4f9a..27eba52 100644
--- a/scripts/validate-data.ts
+++ b/scripts/validate-data.ts
@@ -482,6 +482,10 @@ if (sources) {
}
}
+// House typography, stated so nobody has to guess the motive: displayed text
+// follows a style guide (short sentences, commas over dashes), and since
+// displayed text is data here, the style rule lives with the data rules.
+// CONTRIBUTING.md documents it for writers.
// Displayed markdown prose obeys the same rules. SOURCES.md is exempt because
// quoted publication titles legitimately carry em dashes.
for (const relPath of [
diff --git a/src/engine/schemas.ts b/src/engine/schemas.ts
index 77c1f7d..bdf06ed 100644
--- a/src/engine/schemas.ts
+++ b/src/engine/schemas.ts
@@ -776,7 +776,7 @@ export const sourceEntrySchema = z
url: z.url().optional(),
type: z.enum(['paper', 'report', 'book', 'dataset', 'web', 'design']),
tags: z.array(z.string().regex(TAG_PATTERN)).optional(),
- status: z.enum(['verified', 'flagged', 'book', 'pending']),
+ status: z.enum(['verified', 'flagged']),
tier: sourceTierSchema,
evidenceClass: sourceEvidenceClassSchema,
/** load-bearing only: what the citations actually use this source for. */
diff --git a/src/ui/components/SettingsDialog.tsx b/src/ui/components/SettingsDialog.tsx
index 6789c89..63b53ca 100644
--- a/src/ui/components/SettingsDialog.tsx
+++ b/src/ui/components/SettingsDialog.tsx
@@ -85,7 +85,7 @@ export function SettingsDialog({ onClose }: SettingsDialogProps) {
>