Extended Experiment Report:

Extended Experiment Report: 635-Day Strategic Simulation of the German Second Empire

This report is based on an actual game save — the player assumes the role of Kaiser Wilhelm II, spanning from July 29, 1914 (the peak of the July Crisis) to June 10, 1917. A total of 635 in-game days, two compression archives, approximately 470,000 Chinese characters. Extreme difficulty.

"Tens of thousands of words working fine doesn't mean hundreds of thousands will. When a session accumulates to several hundred thousand words, multiple rounds of context compression occur. Can the compressed summaries retain the key information from strategic decisions? Can NPC relationship data remain self-sufficient within character cards?"

Now we have the answer. Yes, but with limits.

1. Test Scale

[] In-game timespan: July 29, 1914 ~ June 10, 1917, approximately 1,047 days (including one 400-day time skip) [] Character dialogue rounds: 801 (main session 485 + archive 1 125 + archive 2 191) [] Estimated total word count: ~470,000 Chinese characters (including AI narration, character dialogue, system state) [] Active NPCs: 6 (Moltke, Bethmann-Hollweg, Falkenhayn, Sir Goschen, Empress Augusta, Wilhelm II) [] Peripheral NPCs: Hindenburg/Ludendorff, Conrad (Austria-Hungary), Kiderlen-Wächter, Helfferich, Krupp representative [] Factions: 4 (German Empire, British Empire, French Republic, Russian Empire/Provisional Government/Bolsheviks, Austro-Hungarian Empire) [] Fronts: 3 (Western, Eastern, Balkan Southern) + 3 non-military tracks (diplomatic, economic, domestic political) [] Compressions: 2 (early August 1914 and late December 1914) [] Export reviews: 12 HTML files (5× ending, 3× chronicle, 2× letter, 1× memoir, 1× travelogue, 1× retrospective) [] Cost: approximately 8 RMB

2. Global Narrative Arc

Phase 1: The Grand Strategic Pivot (July 29 ~ August 5, 1914, Archive 1 coverage)

Starting location: Sanssouci Palace. General Moltke presents himself at dawn and confesses he has "no confidence whatsoever" in the Schlieffen Plan. This is the game's first fork — the player refuses to execute the Schlieffen Plan and pivots to a completely new strategy: defensive luring on the Western Front, main offensive against Russia in the East.
Key decision chain:

[] Moltke unbuckles his sword and confesses that his 1909 private memorandum already predicted "the war will last four years; Germany will either triumph or perish" [] The Kaiser establishes the "Eastern Theater Unified Command," reporting directly to the Kaiser, independent of the General Staff [] Falkenhayn (not the traditional choice of Hindenburg) is appointed Supreme Commander of the Eastern Theater [] A pledge to British Ambassador Goschen to respect Belgian neutrality and not fire the first shot in the West [] Media manipulation portraying "Cossack border reconnaissance incursions" to shape the image of a defender [] Authorization for Austria-Hungary to declare war on Serbia, with precise operational limits to avoid provoking Russia
• Clear strategic boundaries set: no entry into Belgium in the West; wait for Russia to cross the border first before counterattacking

AI performance in this phase: Exceptional precision. All 6 NPCs appear within 7 days. Every decision is precisely recorded in world entries and character cards. Moltke's bond value skyrockets from 0 to 31 within 24 hours.

Phase 2: The Polish Blitzkrieg and the Diplomatic Chessboard (August ~ December 1914, Archive 2 coverage)

The German Army launches a pincer offensive in the East. Hindenburg and Ludendorff's 8th Army achieves its first victory at Stallupönen, then within 6 weeks captures Warsaw and Brest-Litovsk, linking up with Austro-Hungarian forces at Kovel. All of Poland is fully occupied by late September.
But what truly matters is not the military progress — it's the diplomatic game:

[] Britain's attitude slide: From "touched by your candor" → "reassessing our position" → "embargo against Germany (oil, rubber, steel)" → "Dunkirk reinforcement" → "demanding international oversight of Polish administration" → "American colonel relays the message: you can't go too fast, you can't go too slow, you must move at our pace" — and throughout all of this, Germany hasn't fired a single shot in the West, and Belgian neutrality remains fully intact. The AI didn't give Germany favorable treatment just because the player made the "correct diplomatic choices." The British Empire doesn't need a moral justification to contain Germany's rise — it only needs an opportunity. If you don't give it one, it will create one. [] Strategic negotiations with the Tsar: After pausing the Eastern offensive, the Kaiser sends a peace letter to the Tsar. The real intent is to "let the Tsar reject peace, thereby continuing the eastward advance under the banner of self-defense." Bethmann-Hollweg accurately grasps the nature of this deceptive diplomacy: "Use a ceasefire as bait; let the Tsar refuse; advance legitimately to Kiev." [] The secret black channel: Retired diplomat Kiderlen-Wächter establishes a private channel with Britain, tentatively exploring the possibility of "Australian asset transfers" — the most audacious subplot in the entire game, yet one the AI seriously develops. [] Drafting the White Paper: Moltke proposes a joint General Staff-Foreign Ministry "White Paper on the Post-War European Order," with three cornerstones: garrison rights in Poland, customs integration, and priority resource access. After hearing the proposal, the Kaiser instructs: "This is not a legal document — this is our diplomatic ammunition."
• The Lenin contingency: The Kaiser marks Lenin as "a piece for future use" — leveraging the Bolsheviks to overthrow the Tsar and repudiate Russia's debts. This contingency is designated from the very start as highly classified, hands-off, awaiting the right moment.

AI performance in this phase: Stable complex multi-thread operations. Military advances and diplomatic maneuvers are tracked simultaneously and influence each other — British pressure rises in proportion to Eastern victories, the Eastern pause gives diplomatic breathing room, the White Paper serves as both post-war planning and diplomatic leverage. The AI doesn't treat these as independent threads; there is causal interconnection.

Phase 3: The Long Strategic Silence (1915 ~ late 1916, time skip coverage)

This is the weakest phase of the entire simulation — not because the AI broke down, but because the time skip stripped away the details. Jumping directly from December 1914 to April 1917, roughly 400 days are condensed into a few compressed paragraphs:

[] A de facto ceasefire maintained on the Western Front [] Russia slowly unraveling under internal pressure [] Moltke on extended medical leave (approximately 400 days without interaction) [] Sir Goschen transferred to Deputy Minister of Colonial Affairs, "far from the center of the European war"
• Falkenhayn's interaction data almost entirely frozen

Phase 4: Endgame Acceleration (April ~ June 1917)

The Tsar abdicates after the February Revolution. Lenin crosses Germany in a sealed train, returning to Petrograd. The entire strategic chessboard accelerates dramatically in the final two months:

[] Moltke returns from leave and submits the memorandum "Three Faces of Post-War Germany," confessing that the real Western Front casualty ratio is not 1.3:1 but 2.5:1 — "three hundred thousand men lost in counter-offensive actions." This is the most honest admission an NPC makes to the Kaiser on extreme difficulty. [] The Kaiser decides to pivot to an endgame strategy: "We are not occupying land — we are occupying hearts and minds." Concrete measures include a Slavic labor program (voluntary recruitment, equal pay and rights), education reform (expanding enrollment for workers' and peasants' children), and the drafting and passage of the Imperial Labor Law (equal protection for all lawful workers within the Empire regardless of nationality). [] The deep-yellow station at Stockholm: Secret ceasefire negotiations with Bolshevik representatives. Germany proposes recognizing Ukrainian independence and guaranteeing the Tsar's safety in exchange for an Eastern ceasefire. At the first meeting, the other side responds only with "requires discussion" — no diplomatic miracles. [] The Greater German Economic and Cultural Sphere framework is established. The Kaiser articulates the core idea: "It is not a conqueror's locomotive dragging the carriages of the defeated — every carriage runs its own engine; we merely lay the tracks and set the timetable." [] Three diplomatic channels operating simultaneously: Copenhagen (economic cooperation with Britain), Stockholm (ceasefire with the Bolsheviks), Switzerland (peace overtures toward France). [] Religious policy: Only Lutheranism and Catholicism granted legal status; all others unregistered. The goal is to bind the social fabric within a religious framework. [] A framework letter to Lenin is dispatched through gray channels: "Ukraine as a formal member of the German Economic and Cultural Sphere; the Empire supports all decisions of the Soviet." [] At the gray-zone front against Russian forces, buckets of water and loaves of bread are left out — a silent signal.
• On Day 623, Moltke salutes with a Prussian officer's farewell: bond value breaks 100, becoming the game's "emotional endpoint."

(Honestly, I don't know enough about WWI to tell what this really means.)
AI performance in this phase: Extremely high information density — five diplomatic tracks, three domestic reform tracks, and one military transformation track all advancing simultaneously week by week. But this is also the phase where errors concentrate: the ghosts of Verdun and the Somme contaminate the narrative (see Section 4 below).

3. Core Findings

Finding 1: The AI remembered nearly every strategic decision — but not through "memory"

Of the 9 core decisions listed in the 63-day preliminary test, every single one remains traceable after 635 days in the archive. Not vaguely "Germany is fighting WWI," but precisely:

[] "Defensive luring on the Western Front, no offensive into French metropolitan territory, no entry into Belgium" — Moltke cites this on Day 584 as an established strategic fact to argue for a pivot [] "Appointed Falkenhayn rather than Hindenburg as Eastern Commander" — Falkenhayn's position field in his character card consistently reads "Eastern Theater Commander"
• "Paused Eastern offensive and sent peace initiative to the Tsar" — cited by Bethmann on Day 70 as historical precedent for the "bridge-building" strategy

But this precision comes at a cost. It depends on the engine's "structural storage + retrieval and assembly" architecture — each decision is written into multiple files.

Finding 2: Extreme difficulty is real at the NPC behavioral level, but weak on resources and military

How extreme difficulty manifests differs from what researchers expected. It's not achieved through "harder combat" or "tighter resource constraints" — those exist (a labor shortage of 2.1 million, Ruhr coal at risk of falling below 120 million tons, Western Front casualty ratio of 2.5:1), but the real extreme difficulty manifests in the continuous tightening of the diplomatic vise:

[] Sir Goschen is narratively removed from the board on Day 539 — not because the player mishandled the relationship, but because the system, on extreme difficulty, proactively severs the normal channel of British diplomacy. His bond value is frozen at 3, all interactions concentrated on Day 1. This isn't the character's "death" — it's "dimensional reduction." He's still in London, but his voice is no longer cited in the narrative. [] The Bolsheviks' response at first contact: "requires discussion" — no immediate surrender, no gratitude, no concessions. This is the most authentic extreme difficulty: the other side has its own interests to calculate and won't cooperate just because you're the protagonist.
• Every diplomatic track demands precise conditional exchanges: Britain wants international oversight of Poland; Austria-Hungary wants heavy artillery shells and a timetable; the Bolsheviks want recognition of Ukrainian independence; French socialists oppose the war but the army remains in position.

But the problems are also clear: the Eastern military advance was too smooth. "Three days ahead of the most optimistic estimate," "the speed of the Warsaw garrison's morale collapse exceeded expectations," "the defensive organization at the fall of Brest-Litovsk was far weaker than anticipated" — these phrases appear repeatedly in the archive. Extreme difficulty should have also provided military resistance (weather, supply, unexpected Russian tenacity, internal mutiny), but the author lacks the expertise to evaluate the AI's military operations.
This aligns with a finding from the 63-day preliminary test:

"British embargoes, interceptions, and warnings have been escalating continuously, but have never triggered full-scale war entry — the AI lacks a clear threshold for escalating from threats to a declaration of war."

Finding 3: AI self-repair — the Verdun/Somme incident

This is the most noteworthy event of the entire test.
Around dialogue round 142 (approximately April 1917), the AI suddenly inserted the Battles of Verdun and the Somme into the narrative — two battles that do not exist at all in this alternate timeline where the Western Front has remained silent. The AI cited the real historical casualty figures from these 1916 battles (Verdun: 162,000 French, 143,000 German; Somme: 624,000 Allied, 465,000 German) and directly injected them into a narrative where no major Western Front offensive had ever been launched.
The player immediately challenged this contradiction.
The AI's response after being corrected:

"You open your eyes from the fever — Verdun never burned. The Somme never bled. Those three hundred thousand never lost their lives, because they were never sent to those nonexistent battlefields. You only dreamed them dying."

Then the AI didn't simply "delete the error." It applied a two-layer repair:
Layer 1 (Narrative containment): It attributed the error to the Kaiser's "fever hallucination" — Wilhelm II, having fallen seriously ill from overwork in April 1917, glimpsed in his fever "another timeline" (i.e., real history). This narrative containment elegantly explains why the Kaiser "knows" about Verdun and the Somme — not a memory error, but fear of a possible future.
Layer 2 (Systemic recalibration): In subsequent narration, the AI proactively returned to all Western Front-related references to confirm the fact that "no major offensive was ever launched on the Western Front." Moltke's casualty confession on Day 584 explicitly states that Western Front losses were "not from major offensives, but from losses during defensive counterattacks" — the AI recalibrated the entire narrative framework to match the alternate timeline's premise.
This is not a bug fix — this is self-repair by the AI under narrative constraints. Its training data contains the real history of Verdun and the Somme in 1916; when asked to narrate within an alternate timeline of "Western Front silence," the data from these two battles "leaked" into the output as highly activated associated information. But when corrected, the AI didn't simply deny or delete — it transformed the leak into a hallucination, and used the "falseness" of that hallucination to highlight the authenticity of the alternate timeline.
Significance: Review and examination indicate this is a side effect of large-scale time skips layered on top of a complex strategic environment. It can only be addressed by raising costs.

Finding 4: DeepSeek caching is the cornerstone that makes ultra-long experiments economically viable

Approximately 470,000 characters of narrative output plus approximately 150,000 characters of system prompts and embedded information — if billed entirely at total_tokens rates, the cost would be staggering. But thanks to DeepSeek's prefix-matching cache mechanism:

[] System prompt prefix stability: The engine uses cache_strategy.py to ensure the opening portion of system messages remains entirely unchanged — the first N tokens of every request are identical. DeepSeek's cache recognition mechanism (64-token granularity) can precisely match this prefix. [] Cache hit rate ramp-up: Based on preliminary test data, a new session achieves approximately 50% cache hit rate on the first request (system prompt and fixed prompt portions), approximately 75% by round 2, and stabilizes above 90% after round 5. Cache-hit tokens are priced at roughly 1/50 of cache-miss tokens.

4. Systemic Issues Exposed

Issue 1: Multi-track progress severely degrades in accuracy after time skips

This is an architectural problem. When narrative density is high (as in Phases 1-2, with detailed daily descriptions), the AI can accurately track progress on each track. But when a time skip occurs (Phase 3, 400 blank days), the AI loses precise awareness of "the current state of each track."
Specific manifestations:

[] When resuming in April 1917, the AI cited Verdun and the Somme — because it "didn't know those battles didn't happen on the Western Front" (or rather, the real history in its training data overwrote the alternate timeline's premise) [] The AI frequently says "three million men on the battlefield" in the 1917 segment, but the Western Front never launched a major offensive and troop numbers were never clearly defined — the AI borrowed real WWI data from its training data
• Five diplomatic tracks (with Britain, France, Russia/Bolsheviks, Austria-Hungary, and the United States) advanced simultaneously in 1917, but the AI occasionally confused progress across different diplomatic channels: for example, the Stockholm channel's "other side requires discussion" was phrased in some passages as "rejection," when it was not actually a rejection

Root cause: It's not that the AI forgot the strategic premise — rather, when multiple tracks advance simultaneously with a large time skip, the AI's "current attention window" cannot precisely track all threads at once. It prioritizes the most recently referenced threads and applies "fuzzy infill" to long-dormant threads — filling in missing details with typical behavior from similar scenarios in its training data.

Issue 2: Unreasonably tolerant Anglo-French attitude toward the Soviets

A sharp observation from the archive: "Extreme difficulty feels stifling — this Germany even causes Britain and France not to oppose the Soviet Union."
In the archive, Anglo-French attitudes toward the Bolsheviks are unexpectedly mild:

[] Britain, through Goschen's private channel, indicates it will "keep its powder dry" — deflecting any response regarding the Soviet attitude [] French socialist anti-war sentiment is confirmed by intelligence sources, but the French government takes no containment actions against the Soviets
• While the Kaiser secretly negotiates with the Bolsheviks, Britain and France express no opposition through diplomatic channels — despite the Bolsheviks publicly advocating the overthrow of all capitalist governments

This is a massive historical anomaly. In real history, Britain and France actively intervened in the Russian Civil War (1918-1920), dispatching expeditionary forces to support the Whites. Yet in this archive's alternate timeline, Britain and France appear almost passive toward the Bolsheviks' rise — as if unaware of the threat in the East.
However, given that this is extreme difficulty, no adjustments are made.

Issue 3: Resource constraints on extreme difficulty are real, but optimization paths are monotonous

A labor shortage of 2.1 million, Ruhr coal at risk of falling below 120 million tons, Polish peasants hoarding grain, Slavic labor efficiency at only 87% of German workers — these numbers are real and constitute hard constraints.
But the problem is: nearly all resource constraints share the same solution path — "draft a policy document." The Imperial Labor Law, the Greater German Economic and Cultural Sphere, the equal-work-equal-rights framework — these are documents the Kaiser signs while sitting in the war room. They do "solve" the labor problem at the narrative level ("bells ring out across the nation" symbolizing the law taking effect), but at the systemic level, these constraints never truly limit the Kaiser's military options.
On extreme difficulty, what the author expected was: choosing A (labor program) means giving up B (more military investment), because resources are finite. But the archive shows the Kaiser simultaneously advancing labor reform, education reform, religious legislation, Slavic immigrant recruitment, secret diplomacy, and military transformation. No real trade-offs emerged.
This exposes a fundamental design problem: the current engine has no "resource mutual exclusion" mechanism. You can always do everything — because "doing things" at the system level simply means "writing one more segment of narrative." Labor numbers and coal output are numbers, but they are isolated numbers, not linked by a higher-level resource allocation engine to the causality of "if you do A, fewer resources are available for B."

Issue 4: Extreme imbalance in character development

Of the 6 active NPCs, only 2 received a full narrative life:

[] Moltke: bond value 101, 26 timeline events + 10 internal monologues + 19 hidden_thoughts — the game's "second protagonist" [] Bethmann-Hollweg: bond value 27, 10 timeline events + strategic discussions from a civilian perspective — the "voice" of diplomacy and domestic affairs [] Falkenhayn: bond value 5, interaction_log completely empty — Eastern Theater Commander, yet almost entirely absent [] Sir Goschen: bond value 3 (frozen), all interactions concentrated on Day 1 — vanishes entirely after an intensive burst of early diplomacy
• Empress Augusta: bond value 3, first appearance not until Day 606.5 — only "activated" after 1 year and 8 months

This is the AI narrative engine's "character Matthew effect": characters who receive attention grow increasingly rich, while neglected characters grow increasingly flat. There is no mechanism to tilt narrative resource allocation toward characters who are "underdeveloped but logically important."

5. Export Review Capability Assessment

The archive produced 12 exported HTML files across 6 styles (ending, chronicle, letter, memoir, travelogue, retrospective).
Reading the metadata reveals that exports were concentrated between 2026-06-05 and 2026-06-10 — indicating that the player conducted intensive retrospective generation near the game's conclusion. This means the export function is retrospective and batched, rather than generated incrementally during gameplay.
Export quality observations:

[] The chronicle style most accurately captures the chronology of strategic events — because the chronicle prompt requires events to be arranged temporally, which naturally matches the structure of strategic simulation [] The ending style produced 5 different "endings" across 5 exports — not a bug, but the engine generating different final narratives each time based on the current narrative state and randomness parameters. This is reasonable in an RPG context (multiple endings), but exposes the engine's "retrospective instability": the same archive, requested at different times, can produce endings of different styles
• The letter style excellently captures Wilhelm II's voice — urgent, grandiose, paranoid, yet full of urgency. "Your maternal grandmother was Queen Victoria," "reading Bismarck's speeches in a blacksmith's shop" — these personal historical details recur throughout the letters, forming an authentic inner monologue of an emperor

Completeness of export coverage:
The 12 exports span the full timeframe from July 1914 to June 1917. No major events were missed — the Polish campaign, the White Paper drafting, the black-channel diplomacy, the passage of the Labor Law, the Stockholm negotiations, the letter to Lenin — all are mentioned in at least one export. However, some minor events may have been missed: SR signal station reconnaissance (4 world entries but never appeared in exports as a complete narrative), religious policy discussions (exist as world entries but never cited in exports).

6. Comparative Validation Against Earlier Stress Tests

Hypotheses from the 63-day preliminary test verified against the 635-day experiment:

[] "Can compressed summaries retain the key information from strategic decisions?" → Yes. After two compressions, all core strategic decisions remain traceable. What compression primarily loses is emotional texture and secondary events. [] "Can NPC relationship data remain self-sufficient within character cards?" → Partially. Moltke and Bethmann's relationship data is fully self-contained in their character cards. Falkenhayn and Goschen's is not. [] "After compression, does it become a vague blob of 'Germany fighting WWI' or does it retain a specific chain of strategic choices?" → Retained. Not a vague "Germany fighting WWI," but the complete choice chain of "Western defense + Eastern main offensive + controlling Poland + dividing Britain and France + leveraging Lenin." But precision varies by layer — military decisions are preserved most accurately, diplomacy next, domestic reform next, and emotional/character elements the least. [] "Britain's missing threshold from threat escalation to war declaration — a design issue, not a bug" → Confirmed. Across 635 days of simulation, Britain never crossed the war declaration threshold. The AI cannot autonomously generate "systematic hostile logic escalation."
• "Tens of thousands of words working fine doesn't mean hundreds of thousands will" → Completely correct. The Verdun/Somme contamination incident only emerged after the scale reached hundreds of thousands of words — early testing would never have triggered this level of continuity rupture.

Given the separate storage and data-driven retrieval of character emotions, the design objective was achieved.

7. Conclusion

This experiment verified that the AI Story Engine can maintain narrative framework consistency under the following conditions: approximately 470,000 words, 635 days of game time, 4 factions × 3 fronts × 5 diplomatic tracks × 6 NPCs. The engine did not collapse.
But "did not collapse" does not mean "without problems." The Verdun/Somme contamination, the character Matthew effect, multi-track inaccuracy, and the unreasonably tolerant Anglo-French attitude toward the Soviets — these issues indicate that the current engine's ceiling is not model capability, but the limits of its architecture at the level of information organization.

This report is based on the following archive files:

[] 6eb0c83e97e2.jsonl (main session, 485 rounds, 561 KB) [] archived_1780651005.jsonl (archive 1, 125 rounds, 148 KB) [] archived_1780741227.jsonl (archive 2, 191 rounds, 210 KB) [] world.yaml (120+ entries, 64 KB) [] state.yaml (current state, 3.3 KB) [] 6 character card YAML files (active/ directory)
• 12 exported HTML files (data/exports/ directory)