← Every exhibit

Figure 3From “What AI did to cost of goods sold is not visible in the line everyone quotes.”

Four tests, two instruments, no survivor

The first instrument alarms on noise and on bookkeeping. The rebuild is robust and blind: it detects a real 30% shift less often than it false-alarms. The marked rows are why nothing is prescribed.

Cite Embed

Figure 3Four tests, two instruments, no survivor

Of four simulation tests, the naive elasticity instrument fires on 49.8% of pure-noise runs and on 98.9% of pure accounting reclassifications; the first-difference rebuild fires on only 3.9% of reclassifications but detects a genuine 30% variable-cost shift in only 3.8% of runs, below its own false-positive rate.20%40%60%80%100%v1: fires on random noise49.8%v1, tested: fires on a reclassification98.9%v2: fires on a reclassification3.9%v2: detects a real 30% shift3.8%Share of simulation runs (%)

Source: diagnostic-v1-failed.py, 20,000 replications: fires on 49.8, 50.7 and 49.6% of runs across three noise levels, the first drawn here. diagnostic-v2-failed.py, 6,000 replications: a one-off 8% cost reclassification, and a ramped 30% variable-cost shift with a bootstrap interval. Both published as failures at isoglu.com/research/gross-margin-panel/.

PNG4:3, 1600 × 1200, with the title, the source line and the attribution set in the image isoglu-2026-what-ai-did-to-cost-of-goods-sold-fig03.png

From the essay

Cite or embed

APA 7

Isoglu, S. (2026). Four tests, two instruments, no survivor [Figure]. isoglu.com. https://isoglu.com/insights/exhibits/isoglu-2026-what-ai-did-to-cost-of-goods-sold-fig03/

BibTeX
@misc{isoglu-2026-what-ai-did-to-cost-of-goods-sold-fig03,
  author = {Isoglu, Sinan},
  title = {{Four tests, two instruments, no survivor}},
  year = {2026},
  howpublished = {Figure 3, isoglu.com},
  url = {https://isoglu.com/insights/exhibits/isoglu-2026-what-ai-did-to-cost-of-goods-sold-fig03/},
  note = {From the essay {What AI did to cost of goods sold is not visible in the line everyone quotes.}}
}
Embed this exhibit

Copy this HTML to use the PNG exhibit on your site with its source line and a link back to the exhibit page.

<figure>
  <a href="https://isoglu.com/insights/exhibits/isoglu-2026-what-ai-did-to-cost-of-goods-sold-fig03/">
    <img src="https://isoglu.com/social/isoglu-2026-what-ai-did-to-cost-of-goods-sold-fig03.png" alt="Figure 3: Four tests, two instruments, no survivor" width="3200" height="2400">
  </a>
  <figcaption>Figure 3: Four tests, two instruments, no survivor. Source: diagnostic-v1-failed.py, 20,000 replications: fires on 49.8, 50.7 and 49.6% of runs across three noise levels, the first drawn here. diagnostic-v2-failed.py, 6,000 replications: a one-off 8% cost reclassification, and a ramped 30% variable-cost shift with a bootstrap interval. Both published as failures at isoglu.com/research/gross-margin-panel/. <a href="https://isoglu.com/insights/exhibits/isoglu-2026-what-ai-did-to-cost-of-goods-sold-fig03/">View exhibit</a>.</figcaption>
</figure>

Related exhibits

Figure 4Figure

The three numbers off your own cloud bill

What AI did to cost of goods sold is not visible in the line everyone quotes.

Figure 1Figure

The panel summary: three weightings, one direction

What AI did to cost of goods sold is not visible in the line everyone quotes.

Figure 2Figure

The line everyone quotes, drawn over seventeen quarters

What AI did to cost of goods sold is not visible in the line everyone quotes.

Every exhibit

Contact

Bring the question you're working on.

Research, a guest lecture, or a commercial question.