¢ Cap forecast
04 · Architecture split

One pure forecasting core. One operational owner. One public explanation.

TokenEconomy should own deterministic estimation. Agent Studio should own observations, UI, and decisions. The website should explain the method and publish only scoped, timestamped aggregates.

Research plan only: every implementation boundary below still requires explicit operator GO.

Three ownership boundaries

TOKENECONOMY LIBRARY

Math and contracts

  • Normalized observation and scope contracts
  • Structural/statistical validity and estimability
  • Robust effective-capacity estimate + interval
  • Forecast v1 result for every candidate
  • Confidence/status/reason codes
  • Pure deterministic tests and fixtures
AGENT STUDIO

Evidence and decisions

  • Quota probes, parsers, start/end capture
  • Bus/run/config joins and persistence
  • Contextual parser/window/overlap/attribution gates
  • Calibration cohorts and scheduling
  • Card UI, headroom, reset, audit events
  • Shadow validation and AGT-2055 advisory input
WEBSITE

Method and aggregates

  • Public, human-readable method
  • Generated aggregate snapshot after validation
  • Scope, range, confidence, sample, as-of labels
  • Privacy-reviewed, no raw tasks/accounts
  • Self-contained publish-time HTML
  • No provider-official language

The handoff

Operational data moves inward as normalized contracts; decisions and publication stay with their owning applications.

Studio observations
tokens + snapshots + run facts
→
TokenEconomy
calibrate + forecast + explain status
→
Studio / website
decide locally · explain publicly
No I/O in the library: no quota probe, parser, persistence, current clock, account lookup, event bus read, UI, log side effect, or launch/wait policy.

How this fits TE-1 and TE-2

TE-1 · Pricing catalog ↗

Computes dollar cost from dated price data. Dollar price cannot supply the subscription quota denominator and must never proxy it.

TE-2 · SuggestModel ↗

Ranks capability/cost fit and leaves downshift/throttle/wait policy to Studio. The forecast adds a separate quantitative signal; it does not silently rewrite that ranking.

The policy boundary follows Studio's token-budget load-management concept ↗ and future AGT-2055 admission integration.

Boundary failures to prevent

FailureWhy it is wrongBoundary rule
Studio reimplements the estimatorTwo formulas drift and forecasts become unauditable.Studio constructs contracts; TokenEconomy computes.
Library reaches into Studio storagePure tests and reuse disappear; runtime concerns leak inward.All data arrives as explicit immutable inputs.
Forecast mutates SuggestModelOrdinal fit and quantitative quota evidence become inseparable.Return a distinct signal; Studio composes the decision.
Website calls operational storage liveLeaks data and breaks the self-contained static-site contract.Publish privacy-reviewed aggregates generated and inlined at release time.
Uncalibrated becomes zeroMissing evidence looks like free capacity.Typed unknown status propagates to UI and admission.

A second operator gate still exists

TokenEconomy 0.2.0 is published on nuget.org. Studio can consume the released package today with dotnet add package TokenEconomy; release operations are documented in the publishing guide ↗. This plan performs no release step.