Source symmetry and complete revision ledger
The final baseline records both sides of load-bearing sources and every material correction made before the first observation cycle.
Source-symmetry rule
A load-bearing source must be read for supportive, neutral, and disconfirming observations. Selection is allowed only for scope or relevance, and the reason must be recorded. The ledger below is the July 24, 2026 baseline; future reviews append rows rather than silently replacing them.
| Source | Supportive evidence used | Counterevidence or qualification used | Treatment |
|---|---|---|---|
| JPMorgan U.S.-China report | U.S. capacity, AI spending and VC; lower Chinese power costs; seven-dimension framework. | U.S. national average about $0.18/kWh with regional variation; DeepSeek V4 claim of 3-6 month frontier lag; Chinese idle-capacity estimate is not an official census. | Both sides appear in Chapter 8; figures remain source-qualified. |
| NVIDIA FY2026 10-K | Direct customers at 22% and 14%; platform and supplier exposure. | Indirect-customer concentration is estimated; some customers may be >=10%; one AI research company contributed meaningfully through cloud purchases. | Direct and indirect concentration now appear together. |
| Amazon-OpenAI partnership | $50B investment and broader distribution/compute relationship. | $35B is conditional; AWS gains exclusive third-party Frontier distribution; OpenAI commits to 2 GW Trainium and a larger infrastructure agreement. | Investment, distribution and compute are recorded as separate flows. |
| State Street SPY holdings | NVIDIA is a large index constituent. | Weight changes daily; fund page rounds the dated workbook. | Workbook file, as-of date, retrieval date and rounded page are all named. |
| SpaceX amended S-1 | Anthropic and Google agreements disclose large payments and GPU counts. | Payment-per-GPU arithmetic is not a comparable unit price because scope, generation, ramp and services differ. | Chapter 4 shows the ratios and the denominator caveat. |
| Cloud Security Alliance note | Reported GPU, HBM, packaging and security concentration. | Market shares draw partly on public and secondary estimates; organization-specific exposure is not audited. | Used as reported estimates, never as a calibrated systemic measure. |
| Open-weights coalition statement | Broad coalition preference for diffusion, competition and control. | Weights are hard to recall; signatories benefit commercially from more compute, cloud, security and application demand. | Used as evidence of policy and commercial alignment, not realized outcomes. |
| Stanford AI Index 2026 | Narrow U.S.-China capability gap in the cited comparison. | Closed-open gap widened from 0.5% to 3.3% over the cited interval; metrics and dates differ. | No blanket convergence claim is permitted. |
Table D.1. Source-symmetry ledger for load-bearing evidence at the final baseline.
Complete revision ledger
The ledger distinguishes factual corrections, newly available events, methodological changes, scope expansions, and production fixes. A correction does not imply that the earlier claim was fabricated; it records why the baseline changed.
| Version | Type | Material change | Reason and effect |
|---|---|---|---|
| v1.0 | Factual | Restored SpaceX revenue/loss, proceeds-use estimate, independent valuation, actual index weight and $25B bond. | Removed directional selection and created a real equity-to-credit observation. |
| v1.0 | Factual | Corrected Stanford closed/open gap direction and reconciled IPO share arithmetic. | Prevented a directional misstatement and kept offer value separate from reported proceeds. |
| v1.0 | Method | Retired composite systemic scoring; separated financial, macro, operational, strategic and resolvability tests. | Made disaggregation the core contribution and moved rigor into channel thresholds. |
| v1.0 | Method | Added numeric hypothesis thresholds, affirmative nulls, evidence states and baseline dashboard. | Converted broad claims into a prospective monitoring protocol. |
| v1.0 | Structure | Made Chapter 5 the canonical SpaceX-Cursor treatment; retained blocked/abandoned contingency. | Preserved the boundary case without counting repeated references as evidence. |
| v1.0 | Structure | Repaired scenario typology, formula units, open-weight ecosystem, index literature and multipolar cases. | Resolved internal dimensional and coverage gaps. |
| v1.0 | Production | Added full source URLs and citation consistency sweep. | Made the evidence trail reproducible. |
| v1.1 | Scope | Added NVIDIA, Huawei, TSMC/Taiwan, HBM, packaging, data-center commissioning and hardware portability. | Extended the thesis to the physical substrate and build-out ladder. |
| v1.2 | Factual | Updated TSMC U.S. announced investment to $265B; replaced 100k GPU baseline; added Anthropic/Google contracts. | Brought hardware facts to the July 24 cutoff and separated historical from current capacity. |
| v1.2 | Taxonomy | Added NVIDIA to issuer testing and split H1 into issuer-specific and common-substrate shocks. | Closed the NVIDIA exemption and separated idiosyncratic from shared shocks. |
| v1.2 | Method | Added six-state adjudication, frozen Boolean rules, support-pattern simulation, related work and observability flags. | Made the protocol independently reproducible while retaining explicit limits. |
| v1.2.1 | Method | Corrected endpoint precedence; overlapping support/null is mixed and missing mandatory evidence is not adjudicable. | Removed the internal support-priority contradiction. |
| v1.2.1 | Method | Added immutable event ledger and hypothesis-level replication rules. | Separated one-case results from general mechanism conclusions. |
| v1.2.1 | Method | Separated support-pattern simulation from event occurrence and evidence availability. | Prevented a clean simulation from implying that private evidence will be observable. |
| v1.2.1 | Method | Completed H1-H7 units, denominators, sample rules, redeployment, demand, RTO and task-cost variables. | Closed remaining executable and dimensional gaps. |
| v1.2.1 | Method | Renamed historical exercise as applicability/rule coverage and added backtest-readiness table. | Removed an unsupported predictive-performance claim. |
| v1.2.1 | Factual | Named dated SPY workbook and exact NVIDIA weight; added indirect-customer context and Amazon-OpenAI terms. | Improved traceability and source symmetry. |
| v1.2.1 | Factual | Displayed Anthropic/Google payment-per-GPU arithmetic with non-comparability caveat. | Kept denominators with headline contract figures. |
| v1.2.1 | Taxonomy | Separated primary frontier firms from material AI-system issuers. | Kept the founding cases coherent while testing infrastructure issuers consistently. |
| v1.2.1 | Structure | Added migrating constraint regimes and source-symmetry ledger. | Reconciled hardware scarcity with grid stress and made evidence selection auditable. |
| v1.2.1 | Governance | Declared final baseline freeze before the October 2026 observation cycle. | Future changes must be prospective and versioned. |
Table D.2. Complete factual, methodological, structural and production revision ledger through v1.2.1.
REFERENCES