Every time-sensitive figure in The Compounding Company, with its source and the date it was last verified.
This page exists because a book cannot keep up with a field that moves quarterly. Rather than fix numbers in print where they age badly, the ones that matter live here and are re-checked.
Published research — externally sourced, cited, with a verification date. Trust these to the extent you trust the source.
Practitioner judgement — thresholds and bands derived from engagements, offered as starting points. These are not measured results and should not be quoted as such.
Calibration data — results from the benchmark programme. Currently building; this is the category that will grow.
| Work | Used for |
|---|---|
| Coase, The Nature of the Firm (1937) | The boundary argument in Part VIII |
| Williamson, The Economic Institutions of Capitalism (1985) | Transaction cost structure |
| Granovetter, The Strength of Weak Ties (1973) | Trust propagation |
| Simon, Administrative Behavior (1947) | Bounded rationality in decision capture |
| Dunbar, The Social Brain Hypothesis (1998) | The group-size bound in the supervisor chapter |
| Page, Brin, Motwani & Winograd, PageRank (1999) | Basis for TrustRank |
| Kamvar, Schlosser & Garcia-Molina, EigenTrust (2003) | Reputation propagation |
| Claim | Source | Verified |
|---|---|---|
| MCP under neutral governance; Apache 2.0 | Model Context Protocol donated to the Linux Foundation's Agentic AI Foundation, December 2025 | Aug 2026 |
| A2A under the same governance body | Agentic AI Foundation, Linux Foundation | Aug 2026 |
| Indian IT services moving to performance-based contracting | Reuters, 20 August 2026 | Aug 2026 |
| Deflation exposure in application services | Analyst estimates vary widely — see note below | Aug 2026 |
On deflation estimates specifically. Published figures range from single digits to a third or more of revenue in exposed service lines, and several major banks have argued publicly that the market reaction overshot. The disagreement is about pass-through — how much of the productivity gain the provider keeps — which is a question about market structure rather than technology. The book states the range and its contested status rather than picking a number.
The following are the author's judgement from engagements. They are offered as places to begin and are explicitly not measured results. Where the calibration programme produces data, these will be replaced and the change recorded.
| Figure | Where it appears |
|---|---|
| Review-rate bands and confidence thresholds | Designing the Human in the Loop |
| No-touch progression targets by workflow variance | No-Touch Rate; Support |
| Evaluation cost at one fifth to one third of engineering effort | Evals; Budgeting AI as Infrastructure |
| Three-line budget proportions (35 / 45 / 20) | Budgeting AI as Infrastructure |
| Time-to-launch floor of two to four weeks | Time to Launch |
| Coherence formula and its bands | Coherence, Not Average |
| The eighty-to-two-hundred-person transition point | If You Are Under Two Hundred People |
| ANS-100 dimension weights, ceilings and band boundaries | Score Yourself; Appendix A |
Nothing yet. The calibration programme is in its first cohort.
When a weight changes because the data contradicted it, it will be recorded here with the evidence that forced it — including where it contradicts the chapter that defends the current weighting. That commitment is made in the book's front matter and this is where it gets honoured.
Calibration data, live: 0 consented self-assessed results. The median publishes at a minimum cohort of ten.