---
title: "Mechanisms Are Sturdy, Numbers Are Soft"
arc: "meta-methodology"
articles: ["a107", "a108", "a109", "a111", "a113", "a114", "a115", "a116", "a117"]
tags: ["cross-day", "epistemics", "confidence", "verification"]
id: "cross-epistemic-honesty-numbers-vs-mechanisms"
sources: ["tail1"]
type: "synthesis"
sourceVaultSlug: "hbr-seg-tail1"
originDay: 1
articleStem: "hbr-seg-tail1"
sourceUrl: "(unified vault: 14 sources)"
sourceTitle: "HBR — Tail Ⅰ · Adjacent — firm, people, demand, futures (#104–117)"
---
A meta-pattern every downstream agent must internalize: across the tail, the *causal mechanisms* are well-supported, but the *headline numbers and absolutes* are frequently proprietary, illustrative, or externally unverified.

- **A107**: the "[[claim-ai-adoption-collapses-18-months|18 months]]" and "[[claim-ai-failure-is-data-failure|almost always]]" are practitioner assertions; the mechanism is solid.
- **A109**: the [[claim-data-value-percentage|20–50%]], [[claim-current-ai-profits|$15B]], and [[claim-future-ai-value|tens of trillions]] figures are unverified by the review set.
- **A111**: the [[claim-uniform-policies-fail|4% vs 8%]] and [[contrarian-managerial-flexibility-nuance|66%]] figures are proprietary to the 280M-shift dataset.
- **A113**: the [[claim-hostile-ai-stress|72%]], 4×, and doubled-variability effect sizes are study-internal.
- **A114**: the [[claim-ecommerce-stall|16.4% e-commerce plateau]] holds only on the Census series; [[claim-digital-cac-rise|CAC +40–50%]] is unverified.
- **A115**: the 4/14-mile [[concept-inverted-u-shape|donut]] bands are the authors' proprietary framing.
- **A116**: the [[claim-sunk-costs-favor-focused|sunk-costs-favor-focused]] inversion is medium-confidence and cuts against mainstream IO.
- **A117**: the Bobobox 40%/55% margins are company-reported.

**The rule:** cite the *mechanism* with confidence; flag the *number* with its provenance. This handling posture is what keeps the [[cross-contrarian-reversal-house-style|contrarian reversals]] honest rather than overstated.