# Model audit

The supplied reconstruction's arithmetic was reproduced. Its event frequencies are conditional outputs of supplied assumptions, not empirical probabilities. The original headline model is incomplete. Native Excel recalculation is not claimed.

## Findings that change interpretation

| Finding | Exact source | Result and required treatment |
|---|---|---|
| Original62% model missing | Read_Me!A10; V2_Reconciliation!B8:G10; deck25,28 | Weights/correlation/full event rules absent. The workbook explicitly reconstructs a different model; preserve62/40/12 as attributed claims only. |
| Reference price changes forecast | Forecast!F7:F8; Controls!B10:B11; V2_Reconciliation!E7 |14.6544% chance of≥35% drop from$0.68 versus93.604% from$1.15. Historical drawdown and forward decline differ. Blended price is not constant-quality price. |
| Economic event changed | Methods!B19:D21; Simulation!AK7:AN518 | Workbook requires capacity growth AND weak spend AND weak mix/utilization for two interpolated months; deck uses a more permissive OR definition. Finance adds two of target<80%,refi>14%,bridge<$250bn; thin bridge<$200bn. None is fitted legal default. |
| Bull sales exceed capacity | Consumption!D16; Production!D13:D17 |14.697× potential revenue,685.96% required utilization,11.719 baseline compute units unmet. Sales are not constrained by delivery in the revenue calculation. Show potential demand and capacity gap; do not publish these as realized sales. |
| Base utilization inconsistent | Production!C14:C16; Validation!D27 |40.746% implied versus72% separately assumed,31.254pp gap. A calibration warning, not actual idle-capacity evidence. |
| Price not physical compute | Consumption!C25:C31; Methods!B15:B18 | Frontier/cache energy corrections affect compute separately; net price already includes billing mix. Inputs8(review/error),12(assumed utilization),13(discount),15(capability) affect risk score or diagnostics, not the mechanical accepted-work demand equation. |
| Energy denominator mismatch repaired | Token_to_Power!J7:K9; deck20 | Deck facility/server ratio12.77/18.30/30.61%; corrected IT/server9.12/13.07/21.86%. Energy0.1–1.0kWh/M and cache/output energy weights are judgments, not fleet measurements. |
| Annual funding is not runway | CoreWeave!C31:C46; Quarterly_Bridge!B20:E22 | Blip simple need$14.58bn; all-in$15.2574bn; planned$14bn financing and$2bn opening cash still breach$1bn floorQ1 and turn negativeQ2. Timing is illustrative, not a six-to-nine-month default estimate. |
| Cash and interest perimeters | CoreWeave!B8:B10,C24,C35:C43 | Quarterly net interest is expense, not cash. OCF includes operating cash interest/prepayments. Only incremental finance costs are added to OCF-based need. |
| Hyperscaler funding is partly unidentified | Hyperscaler_Bridge!B12,D18:D21 |$612.7bn mixed gross assets×30%=$183.81bn; assumed$300bn bridge leaves$116.19bn unidentified. Neither gross assets nor other issuers' cash are automatically available to a borrower. |
| Valuation is an illustrative hurdle | Valuation!B7:B19 |$3.2563tn equity/15×sales=$217.0867bn annual revenue; comparator$117.2bn mixes run-rate and quarter annualization. Not consensus or known global end-user spend. |
| Workflow costs are illustrations | Workflow_Unit_Econ!B7:C21 |$14 vs$76 at100% acceptance reproduce deck stacks. At95% accepted outcomes,$76/95%=$80, eliminating its illustrated$4 saving. Integration overhead starts atzero only to reproduce deck. |
| Benchmark coverage≠failure rate | Sources!H9; Assumptions_18!row15; deck8 |90% unbenchmarked work does not mean90% unautomatable work or90% output failure. Blocked-share and success-rate ranges are judgments. |

## Probability and scenario reconciliation

| Quantity | Deck claim | Reconstructed250k conditional frequency |
|---|---:|---:|
| Price decline≥35% |92% |14.6544% from current;93.604% from peak |
| Risk-score threshold |called demand constraint62% |57.1544% |
| Stricter mechanical constraint |not the same event |41.094% |
| Score plus finance flags |joint shock40% |13.0444% |
| Mechanical plus finance flags |not the same event |7.948% |
| Score plus thin bridge |systemic claim12%, not equivalent |3.556% |
| Mechanical plus thin bridge |not a systemic default model |2.1816% |

The mechanical partition is58.906% no-constraint,33.146% constraint without joint finance flags,5.7664% joint excluding thin bridge,2.1816% thin-bridge joint. These sum to100%; they do not establish reality's state probabilities. The deck38/22/28/12 partition is a different supplied claim.

Triangular bounds/modes are judgmental. Common Gaussian shocks impose0.10 cross-block and0.45 within-block latent correlation. Block weights50/30/20 and equal within-block weights are newly chosen. Numerical Wilson intervals shrink with more draws even if the model is wrong. Predictive quantiles condition on those assumptions. The128 specification variants vary score weights/threshold only, not every economic equation or causal relationship. Deck claims1,000 variants; companion uses128. Sensitivity tab shifts a variable while holding others' realized draws, breaking the original conditional dependence; interpret it as a score perturbation, not a causal marginal effect.

## Sources, actuals and gaps

All source register statuses are **supplied provenance claims**, not a new source check by this audit. `model_data.json.source_register` retains names, dates,locations,statuses and caveats; current evidence team must adjudicate external claims.

Stronger claimed provenance: S03 Mercor benchmark study; S04 Ramp transaction aggregate; S06 OpenRouter platform telemetry; S07–S09 Anthropic rate cards; S12 CoreWeave filing excerpt. Each has scope limits: benchmark coverage, platform sampling, token denominator, list versus contracted pricing, or vintage/borrower perimeter. Their numbers are not universal market estimates.

Explicit inherited/weak points include S05 September Ramp price/mix page unresolved, S10 GPT-6 rate card inherited, S11 DeepSeek discounted rate inherited, mixed-vintage S27 liquidity balances, and energy coefficients. Paid/task/reliability/blocked-work/training growth ranges and financing willingness remain assumptions even when a source is cited beside them.

The separate research paper mentioned on slides4/33 and workbookS02 is **not present among these three supplied originals**. Claims about its78.35% and `unknown_url` are visible only as inherited descriptions; this audit does not independently review the missing paper. Source URLs alone are not evidence of retrieval or entailment.

Actual AI incremental profit, actual accepted jobs, actual token workload/cost, actual provider fleet utilization, actual cash runway, actual contract termination rights and realized financing availability remain unestablished. Unknown must not become zero. Existing-company consolidated OCF cannot be called AI contribution.

## Slide story extraction

`slides.json` and `slides.md` preserve all33 slides in order. Sequence:1–6 thesis/event definitions/master drivers;7–12 accepted work/verification/code economics;13–16 spend/mix/cache/always-on demand;17–20 lab targets/valuation/tokens/power;21–23 effective supply/power delays/CoreWeave financing;24–29 probabilistic scenarios/assumptions/falsification;30–31 exposures/conclusion;32–33 sources/conventions. Slide3's+27pp evidence bridge is explicitly judgmental rather than statistical decomposition.

For the webpage, use the user’s exact ten-point narrative as the governing sequence. The slide order is supporting material, not a replacement. Place the strongest quantified caveats at the relevant argument: capability measurement, dollars per accepted job, cash timing, and macro value capture. Keep full probability audit available under a concise disclosure.

## Acceptance limits and follow-up

VERIFIED: source binary identity, formula extraction, key arithmetic, reproduced companion frequencies,10 meaningful boundary tests. NOT VERIFIED: source truth/currentness by this lane, complete original62% simulation, native Excel render/recalc, legal default probabilities, actual6–9month liquidity deadline, measured GDP contribution. Root and independent reviewer own final integrated acceptance.

The highest-value missing data are a same-scope paid workload/price/accepted-output panel; serving utilization and fleet capacity by region/hardware; monthly borrower cash, capex and maturity schedules; executed financing and backlog conditions; and end-customer retained value distinct from supplier receipts. These would test the thesis. More random draws would mostly print the same assumptions with more decimal places.