Currency Delta · July 2026 · v3.2 release
GTM Currency Delta — July 2026
Model changes
| Change type | Details |
|---|---|
| State variables added | 10 — autonomy α, governance maturity G, review catch-rate γ(α,G), per-step error p, chain reliability R, agent spend C_a, HITL learning-rate λ, category noise-floor θ, category adoption share κ, buying-committee agent share B_agent (17 → 27) |
| Equations added | 9 — E-A1 (per-step error), E-A2 (reviewer-fatigue catch-rate), E-A3 (chain reliability), E-A4 (effective cost/task), E-A5 (agent spend in CAC), E-A6 (HITL as learning-rate), E-A7 (category noise floor), E-A8 (agentic-CAC holdout), E-A9 (buyer-state agent extension) (17 → 26) |
| Theses added | T8.1 (autonomy α and governance G are separable levers), T29 (governance is a precondition for autonomy to lower cost, not a safety tax), T30 (agent adoption is a category externality) (29 → 32) |
| Scope | Tier 1 (Execution / Behavioral) and the Operations & Systems substrate only. Tiers 0, 2, 3, 5 and the 17 base equations unchanged. |
| Measurement-gap register | 6 new agentic parameters added, all tagged unmeasurable_hypothesis: ΔCAC_agentic (highest priority), γ_max/ρ, θ/δ_θ, λ, B_agent, κ_gov |
The load-bearing result — T29
Governance maturity flips the cost-optimal autonomy level. Under low governance, a reviewer-fatigue term (ρ) makes human review a poor investment, and the cost-minimizing choice drifts toward near-full autonomy despite worse reliability. Under high governance, review pays for itself and the cost-minimizing choice flips to near-full human oversight — at roughly half the cost floor and materially higher reliability.
A worked 6-step outbound example produces α* = 0.93 (near-autonomous) at low governance vs α* = 0.09 (near-fully-reviewed) at high governance from the same equations — opposite strategies, not a difference of degree. The deciding GTM variable is not "how autonomous is our agent" but "what is our γ_max and ρ." All example parameters are illustrative, not vendor-calibrated, consistent with E-A4's estimator tag.
Epistemic honesty
Every added edge carries an epistemic type. The model refuses to assert the sign of the agentic-CAC comparison (E-A8) because no public holdout exists — it is written down precisely to name its own unmeasured inputs. Six of the new parameters remain unmeasurable_hypothesis: they may inform risk-awareness but must never be fed as a number into a forward agent action.