C/PAdvanced-compute evidence ledgerWhen Controls Raise the Cost

Methodology / version 0.1.0

Separate progress from proof of policy failure.

The method asks what capability returned, what it cost, what it still depends on, and whether it can be repeated—not whether a dramatic result exists.

Interpretive rule

Why progress does not equal failure.

Export controls can matter without stopping every model release. A Chinese developer may restore a capability by using more chips, older chips, a larger network, more electricity, greater engineering effort, a fragile foreign route, or public money. That outcome is evidence of adaptation. It becomes evidence that a control weakened only when the relevant capability is repeatable at useful scale and dependence falls at the control point being tested.

The distinction is causal. Observing progress after a rule does not reveal what would have happened without the rule, how much the response cost, or whether the same method supports the next generation. This project therefore avoids binary “worked / failed” scoring.

Current finding: One response weakens a narrow control point; none proves system-wide independence.

That finding can change as stronger, independently auditable evidence arrives.

The common instrument

The same six questions for every response.

No case receives a bespoke standard. Answers must remain tied to evidence cards and major unknowns stay visible.

  1. Q1

    What AI capability was restored, and on which tasks?

  2. Q2

    Can it work repeatedly at useful scale?

  3. Q3

    What extra hardware, electricity, money, engineering, or time does it require?

  4. Q4

    Does it reduce dependence on technology controlled by the United States or its allies?

  5. Q5

    Could a realistic enforcement change break the workaround?

  6. Q6

    Can it support the next generation of development, or only current needs?

Decision rule

A judgment states the mechanism.

The label is narrower than the case headline. It describes whether the relevant control point was displaced, paid around, bypassed, or not yet demonstrated.

01
Genuine weakening

The relevant capability is restored at useful scale while dependence falls at the targeted control point.

02
Adaptation + cost penalty

The capability is real, but added hardware, energy, time, engineering, or outside dependence remains.

03
Circumvention, not independence

Access bypasses enforcement but remains tied to controlled foreign technology and a disruptable route.

04
Insufficient evidence

The public record cannot connect the response to restored capability, scale, or a particular rule.

Source hierarchy

Start with the record closest to the claim.

  1. 01Rules, court records, regulatory filings, and official policy documents
  2. 02Technical disclosures, model cards, repositories, and reproducible benchmarks
  3. 03Peer-reviewed research and clearly labelled preprints
  4. 04Independent technical evaluation and specialist analysis
  5. 05Careful reporting used only where primary evidence is unavailable

A company preprint is primary evidence of what the company reported. It is not independent proof that the result is reproducible or the comparison is like-for-like.

Evidence-card contract

Claim-sized data, reciprocal references.

01Claim
02Source and type
03Publication + access dates
04Quotation or data
05Exact source location
06Counterevidence
07Confidence
08Remaining uncertainty

Automated checks validate the shape and citation graph. Human review still determines whether a claim accurately represents its source.

Known limits

What this MVP cannot establish.

  • L1Exact chip inventories, acquisition dates, manufacturing yields, energy use, subsidies, and total engineering costs are usually private.
  • L2The Pangu and CloudMatrix performance evidence is vendor-authored and has not been independently reproduced at full scale.
  • L3Benchmark gains mix the effects of algorithms, data, hardware, and implementation; they do not isolate the causal effect of export controls.
  • L4Enforcement cases reveal detected schemes, not the prevalence of undetected diversion or the capability ultimately produced.
  • L5Rules changed repeatedly. Each response must be matched to the rule and license policy in force at that time.
  • L6This MVP is selective, not exhaustive. It favors cases with inspectable public evidence and labels gaps instead of estimating hidden quantities.

Falsifiers

What would change my mind.

  1. F1Repeated, independently audited frontier-scale training on domestically fabricated accelerators with disclosed HBM, yield, power, cost, and volume would move the domestic-substitute finding toward broad control failure.
  2. F2Metered CloudMatrix deployments showing cost and power parity—not only per-TFLOPS efficiency—would weaken the current cost-penalty judgment.
  3. F3Evidence that diversion or offshore access reliably supplies next-generation compute after realistic customer, ownership, and data-center checks would show a durable enforcement failure.
  4. F4Audited firm-level data showing no material delay, redesign, inventory drawdown, or cost increase after a matched rule change would challenge the claim that controls still impose friction.
  5. F5Conversely, documented project delays, unmet accelerator demand, falling training scale, or sustained upstream shortages would strengthen the cost-imposition finding.