The World Model

Track record

Five pooled walk-forward windows (2021-2025): each year predicted from data strictly truncated to the prior year, scored against what actually happened.

Walk-forward backtest: profiles truncated to each cutoff, predictions frozen before scoring, rule changes apply forward only. The predictor model was trained on post-2020 data, so pooled scores measure method consistency and calibration, NOT out-of-sample foresight. The only true out-of-sample test is the live forward ledger from 2026H2 onward.

The Registry — 97 nodes, 83 profiled

Every scored stakeholder, one row each — including the five pilots this model started with (US President, Chinese party-state, Altman/OpenAI, Jensen/Nvidia, salaried knowledge workers), now first-class rows among the full set. Click a row for its incentives, reaction functions, edges, and recent logged actions. By class: sovereign-political 14 · ai-compute 14 · sector 12 · population 12 · watchlist-domain 9 · infrastructure-chokepoint 8 · meaning-adversarial 8 · governance-rules 6 · field 5 · capital 4 · states-blocs 4 · robotics 1

Power model

evidence-card scoring over 82 profiled nodes, registry-normalized 0-100 per dimension, fixed blend — 83 nodes scored, as of 2026-07.

Kept the task's ASI-transition blend (authority .25, compute .22, capital .18, chokepoint .15, attention .10, talent .10). It is consistent with engine-rules-v6 Step C (feasibility-weighting by power) and the v0 doctrine that legal coercion + the physical compute substrate are the binding constraints of the transition. Only change vs the v0 pilot: compute raised .20->.22 and chokepoint lowered .20->.15, because across the 82-card corpus compute-capacity control (owned fleets, custom silicon, contracted GW) proved more discriminating and less double-counted than chokepoint, which overlaps heavily with authority for sovereign nodes and with compute for silicon nodes. Same weights applied to every node so relative power stays comparable.

Forecast evolution

The engine that produced this forecast has already revised itself seven times — v1-v6 rewrote the prediction rules after each backtest window; v7 expanded coverage from the 5 pilots to all 83 profiled nodes.

Cross-node dynamics

Horizon shown: H2 2026 → 2027.

• US-China chokepoint depreciation race (uspres-rf1 x china-rf1/rf4 x usnatsec-rf2/rf6 x hw-rf1): the escalation-truce cycle keeps running — next iteration trigger-anchored to the ~Oct/Nov 2026 rare-earth truce expiry, the truce priced FRAYING (the May 2026 Beijing summit bought agriculture with no rare-earth commitment; the Busan contrarian's closest-ever near-miss is promoted to a rising-tail watch item) — but BOTH levers depreciate with use even as BOTH ratchets run DENSER each cycle (rc-2024-4): chip controls accelerate Chinese substitution (the 2026-01 MIIT boycott order completed jensen-rf1's structural zero; Huawei's ~50% domestic share is the depreciation made visible), mineral controls accelerate Western reshoring. The v7 addition from the newly profiled us-natsec node: the ratchet now has a CASH REGISTER — usnatsec-rf4 converts each control rung into tolls and equity stakes, and usnatsec-rf6's negative space guarantees the total-denial line is never held, so the race's terminal state is managed, monetized interdependence, not decoupling.
• State-capital fusion of the US AI stack — the gate is now an INSTITUTION, not an event: altman-rf2/rf7 x uspres-rf8 x usnatsec-rf4/rf7 x ssi-rf3 x jensen-rf8. The GPT-5.6 hold-and-clear cycle (Jun-Jul 2026) plus the Sol-variant cyber-capability hold give the security-review gate its first exercised instances, CAISI is the standing operator, and the sovereign-fund equity proposal (2026-07, ~5% of OpenAI) is the first formal equity instrument on the table; the chip side already runs on revenue-share tolls and the Intel stake. the bespoke deal terms least likely; the tolls in between. LL-2026H2-23 is this dynamic's live test.
• The compute-buildout debt cycle (oracle-stargate x neoclouds x private-credit-complex x softbank x sector-finance x us-federal-reserve) — NEW in v7, the horizon's largest financial-stability dynamic: OpenAI's compute commitments flow into Oracle's $300B contract and the neoclouds' contract-collateralized debt (CoreWeave's ladder), which private credit holds and banks underwrite, while SoftBank levers the equity layer ($64.6B stake + $40B bridge) and Nvidia backstops the neocloud middle (~$6.3B unsold-capacity commitment). Coverage ratios are deteriorating in Fed research (3.1x -> 1.6x), the Oracle CDS blowout (2026-02, spread tripled) is the first market-priced stress event, GPU depreciation schedules are publicly flagged as understated (~$176B), and the Fed's fed-rf7 negative space says it will NAME the exposure without ACTING on it. RF-collision logic: privatecredit-rf5 (never concede systemic risk) x neocloud-rf5 (never shorten depreciation) x oraclestargate-rf2 (reaffirm-and-refinance) form a mutually-reinforcing denial lattice that holds until a demand disappointment at the OpenAI layer tests it; fed-rf1's market-put is the designated circuit breaker. Named early-warning indicators, in order: the Oracle CDS spread, a neocloud depreciation restatement, a flagship perpetual-vehicle gate deployment, an ASML/TSMC order-book inflection.
• The state-security-review gate as a standing institution (us-natsec-export-control-state x standards-safety-institutes x lab-governance-layer x altman-openai x amodei-anthropic x frontier-eval-complex): rc-2026H1-4's mechanism now has a full institutional stack — CAISI operates the reviews using frontier-eval-complex methodology (fec-rf1's capability clock becomes the trigger metric), usnatsec-rf7 extends the chokepoint from chips into model weights, and lgl-rf5 shows lab release cadence bending to it while internal governance (LTBT, boards) stays hollow. The gate has held and cleared, never yet cancelled or rolled back — that distinction is load-bearing in every negative-space no-pause claim (LL-2026H2-08 carries it). The collision to watch: a gate exercise against ANTHROPIC (whose amodei-rf6 negative space predicts no self-imposed pause) would test whether the accommodate-the-sovereign pattern generalizes beyond OpenAI; a gate ending in CANCELLATION is the class's regime break and an immediate re-emission trigger.
• Circular-financing correlation, now four-layered (jensen-rf3 x altman-rf4/rf9 x neocloud-rf2 x softbank-rf2 x nonnvidia-rf3): Nvidia equity funds OpenAI's silicon purchases ($30B closed 2026-02, $30B more in the $110B round), AMD pays for OpenAI demand in warrants (~10% of the company), Nvidia backstops CoreWeave's unsold capacity, SoftBank bridges the equity gaps, and the $110B round (Amazon $50B + SoftBank $30B + Nvidia $30B at $840B post) closed the loop at record scale. The strongest private nodes have deliberately correlated their balance sheets. Band forms split per rc-2026H1-1: monotone legs (usage, revenue, round ladder) keep open-upper bands; reversible legs (Nvidia market cap; OpenAI's the day it lists) carry haircut lower edges with the down-leg twice-demonstrated. The correlated upside is unbounded in-band; the correlated markdown is priced, observed behavior.
• The white-collar shock chain (labs' agentic capability x sector-professional-services x sector-finance x amazon/microsoft execution x salaried-knowledge-workers x students-new-grads x us-congress x us-president) — expanded in v7 with its full transmission path: capability release -> enterprise adoption (50%+ of firms) -> junior-first substitution (Big-Four intakes -6-29%, bank analyst classes cut two-thirds, McKinsey's largest-since-2008 cuts) -> the workers node absorbs (29+ months of contraction, 100K+ AI-cited layoffs in H1-2026 alone) -> students shelter in debt-financed credentialing (LSAC +33%) -> NO union or collective response (workers-rf9, labor-rf5, students-rf5, sci-rf6 — Congress's role is confirmatory: congress-rf2 blocks comprehensive frameworks, so the response arrives as a reconciliation-class fiscal program, not conduct regulation. The apprenticeship-pyramid break (juniors never trained) is the chain's slow-burning structural payload —
• The open-weights front realigned by Meta's defection (zuck-rf2 REVERSAL x ds-rf1 x ali-rf1 x hf-rf1/rf6 x china-rf7): Meta's closed turn (Muse Spark, 2026-04) removed the US anchor of the open ecosystem, handing the open-weights front to the China track — DeepSeek (research-grade openness on the founder's bar) and Alibaba's Qwen (distribution-grade: 700M+ downloads, past Llama as the most-downloaded family) now define it, with Hugging Face the neutral-but-China-weighted distribution artery (hf-rf6 keeps hosting without geo-fencing). Consequence chain: US open-weights leadership now depends on second-tier players and the one-off gpt-oss class, the commoditization pressure on US closed labs comes almost entirely from Beijing-aligned releases (feeding altman-rf9's burn and the fusion dynamic's state-cover price), and a US policy action against Chinese open-weights distribution — previously unthinkable — becomes a live option in any 2027-28 hawkish turn. LL-2026H2-20/24/25 test all three legs.
• Taiwan-strait as the system's binding tail (tsmc-rf5/rf6/rf8 x taiwan-rf1/rf8 x china-rf5/rf10 x uspres-rf9 x samsung/korea contingency): the kinetic tail stays in the shock budget, but the SHAPE of the deterrence is now fully priced across four nodes — TSMC keeps the leading edge in Taiwan BY CHOICE (the shield is its siting policy), Taiwan trades thin slivers for US lock-in while its legislature cuts the defense budget 40% below the executive ask, China's drills gain specificity each cycle without crossing (rf10 restraint held every cycle), and uspres-rf9's anti-decoupling restraint holds unless a Taiwan kinetic event or a smuggled-compute capability shock flips it to embargo. RF-collision logic: every actor's dominant strategy is to keep the ambiguity — the dynamic's stability is real but it rests on three restraint patterns (china-rf10, taiwan-rf8, uspres-rf9) that have never been tested simultaneously.
• The grid rate-limiter (grid-equipment-oligopoly x grid-operators-regulators x sector-energy x sector-construction x microsoft/hyperscalers) — NEW in v7 as a first-class dynamic: the physical ceiling on the entire buildout is set by four stacked constraints — Microsoft naming an ~$80B unfulfillable AI backlog on power constraints is the demand side made explicit. so the rate-limiter binds harder each year, capacity auctions clear at caps, and ratepayer politics (13-governor class, 2026-28 cycles) becomes the buildout's largest domestic political exposure outside displacement itself. A greenfield mega-plant wave (LL-2026H2-28's refutation) is the single clearest signal the constraint is breaking.
• The HBM/memory pacing constraint (samsung-sk-hynix x south-korea x huawei-smic-cxmt x jensen/nvidia x tsmc) — NEW in v7: memory paces BOTH blocs' accelerator roadmaps. West: every Nvidia/AMD/Broadcom-class part queues at the Samsung/Hynix/Micron HBM4 qualification gates (samhbm-rf1), giving Seoul alliance-bargaining currency (korea-rf4) and making Korean capacity discipline (samhbm-rf3/rf8) a quiet co-governor of accelerator supply. East: CXMT's HBM output caps Huawei's Ascend volume at 250-300K 910C-equivalents against 600K targets — the single binding constraint on China's compute self-sufficiency, worth more than any single export rule. RF-collision logic: a Chinese HBM3-class breakthrough at CXMT would relieve China's cap AND crash the scarcity pricing the Korean duopoly's discipline maintains — the one event that moves both blocs' compute curves at once;
• Gulf compute-for-alignment trade (uae-rf1/rf4 x mbs-rf3/rf5 x uspres-rf8 x jensen-rf8 x altman/amodei raise cadence): the Gulf sovereigns convert hydrocarbon capital into US-conditioned compute and cross-lab equity — UAE executes (Stargate UAE phases, the ~$49B MGX fund, stakes in OpenAI+xAI+Anthropic simultaneously), Saudi announces bigger and delivers slower (mbs-rf5's announcement-layer vs Humain's real 100MW-class delivery layer), and both bought in via personally-channeled access to the administration (the World Liberty Financial entanglement is now publicly exposed). RF-collision logic: the trade is triple-conditioned — making Gulf compute the single most politically-contingent tranche of global capacity;
• The embodiment wave's supply-demand geography (china-rf6 x robotics-complex x sector-manufacturing x japan-rf6 x korea demographics x amazon-rf3 x musk/Optimus) — DEMAND concentrates where demographics force it — Japan (japan-rf6: robots INSTEAD of immigration, the ¥500bn Physical-AI strategy) and Korea (world's highest robot density, Population Ministry) are the US-bloc's structural early adopters. RF-collision logic: US export controls extending to Chinese humanoids (the DJI precedent, roboticscx-rf6's negative space still holding) is the policy fork that would split the wave into two hardware ecosystems exactly as chips split;
• The discontinuity cadence as a system clock (rc-2022-4, verified): roughly one ChatGPT-magnitude mass-adoption or capability-salience discontinuity per ~2 years (GPT-3 2020, ChatGPT 2022, DeepSeek R1 2025), sourced from the US closed-lab track (consumer-agentic mainstream crossing by end-2027; the io device ~2027-28), the China open-weights track (a second R1-magnitude shock is base-rate-expected), or — new in v7 — Each event moves the whole board: deepens state-capital fusion (a discontinuity-scale release is the likeliest next gate trigger), step-functions worker adoption AND displacement, delivers Nvidia a demand surge or an absorbed efficiency scare, and — if China-origin — arms uspres-rf9's embargo flip. Magnitude bands asymmetric on monotone metrics only (rc-2025-1/rc-2026H1-1). Plan the horizon around 1-2 such events; every node read degrades gracefully to its directional call if a specific event is mistimed.
• The governance-rupture cadence as a second system clock (rc-2023-3, verified both directions across four windows) — now priced across 83 nodes: internal instability is a recurring class with forecastable frequency and un-forecastable specifics. The stable floors that keep the enumeration honest: Nvidia (thrice-validated 0.15), TSMC (rf7 pre-announced transitions), Singapore (sg-rf8 stage-managed succession), Korea (korea-rf5 — ruptures resolve THROUGH institutions). The three sovereign apex tails (Xi succession, MBS kingship, Tahnoun concentration) remain unmodelable and outside the table. ALL clocks are annual-rate constructs: Poisson-scaled per rc-2026H1-2 in every sub-annual window — quarterly silence is expected behavior, a zero-event YEAR is the break signal.

Live ledger

MANDATORY FRAMING (rc-2026H1-6): the pooled five-window backtest (2021-2025) is METHOD CALIBRATION — the predictor's training overlaps those windows, so backtest scores validate the machinery, not out-of-sample skill. The 2026H1 window was a knowledge-horizon SMOKE TEST (rc-2025-5): its mechanics findings inform this file's structural rules but its outcomes never entered the pooled record and re-fit no confidence value. The liveLedger below is the ONLY true out-of-sample test: every claim postdates the Jan-2026 training cutoff, is frozen once posted, and will be publicly scored. The first genuinely-blind scored window in this framework's history is 2026H2. Coverage bound (v7): the five-node pilot bound is RESOLVED — this file prices all 83 profiled registry nodes (33 Tier-1 full treatment; 39 Tier-2 and 11 flow-population nodes compact), on the same frozen engine (engine-rules-v6) and power-scores-v1; the residual unpriced surface is the registry's unprofiled remainder and genuinely novel mechanisms. The 15 v6 liveLedger claims are carried byte-identical and remain frozen for blind scoring; the 18 v7-appended claims (addedIn:v7) freeze on posting and extend the same out-of-sample ledger. 15 frozen claims + 19 added in v7.