05 · Slice proposal

Six small cards, ordered to learn before deciding.

Build the estimator and evidence adapter, then forecast, shadow-test, expose the card, and only later let the signal influence admission. Publish aggregate numbers last.

BLOCKED — OPERATOR GO REQUIRED. These are proposed inputs and outputs, not active cards. The parallel Studio start/end-capture card is an external dependency and is not re-authorised here.

Card-sized proposals

Identifiers are planning labels, not allocated tracker IDs. Every implementation row stays blocked until the operator explicitly authorises it.

Card / owner	Inputs	Outputs	Dependencies	Gate
1 · TE-F1 Calibration contracts + estimator TokenEconomy	Context-cleared normalized spans, scope, token components, meter resolution	Structural/statistical validation; slope, effective capacity, interval, confidence, diagnostics; tests	Agreed normalized contract	GO required
2 · AGT-F1 Observation adapter + store Agent Studio	Cached boundary events plus calibration-grade fresh/settled snapshots, reliably run-correlated bus tokens, lifecycle/config metadata	Versioned spans; contextual parser/window/overlap/attribution gates and quarantine reasons; TE input	Capture completion; TE-F1 contract	GO required
3 · TE-F2 Forecast v1 API TokenEconomy	Task features/attempt history, all candidates, calibration result	Median attempt tokens and cap %, prediction range, confidence/status/reasons; tests	TE-F1	GO required
4 · AGT-F2 Replay + shadow forecast Agent Studio	Eligible historical cohorts, TE-F2, later meter outcomes	Temporal MAE/bias/coverage and direct-delta baseline; frozen gates plus later untouched confirmation	AGT-F1; TE-F2	GO required
5 · AGT-F3 Card + AGT-2055 advisory input Agent Studio	All candidate rows, current 5h/weekly state, reset, SuggestModel rationale	Human comparison and auditable advisory field/event; no admission behavior change	AGT-F2 passes; NuGet publish; AGT-2055	GO required
6 · WEB-F1 Promote plan + aggregate snapshot Website	Delivered explainer pages and privacy-reviewed generated aggregate	Extended self-contained docs; timestamped “live-ish” figures with scope/sample/range/confidence	Validated export; privacy review	GO required

Suggested sequence after GO

PHASE 1

Calibrate

Design TE-F1 and the Studio adapter contract together. Keep probes and storage on the Studio side.

PHASE 2

Forecast

Implement TE-F2 against deterministic fixtures and explicit uncalibrated outcomes.

PHASE 3

Shadow

Replay history and observe future tasks without changing routing. Establish error and coverage.

PHASE 4

Expose advisory evidence

Show the UI and attach an audit signal to AGT-2055. Any admission influence is a separate later card and operator authorization.

Evidence gate before admission

Do not invent numeric pass thresholds in the plan. Use the first shadow distribution to choose and freeze them, then test those frozen thresholds on a later untouched temporal hold-out—the threshold-design report cannot also be the pass.

Shadow report must show

Temporal hold-out percentage-point MAE and signed bias
Empirical coverage of the advertised 80% prediction range
Scoring on later clean, eligible meter-delta intervals—not beta × realizedTokens
A direct empirical cap-delta baseline on those same held-out pairs
Breakdown by plan/model/thinking/task type
Terminal-outcome mix and quota-truncated/still-running censoring rate
Sparse/uncalibrated rate and error outside calibrated support
Parser, overlap, reset, rolling-expiry, and stale exclusions

Decision review must show

Examples where the upper range would change launch/wait
False-wait vs quota-wall trade-off
Interaction with weekly windows and SuggestModel
Stable audit reason codes
Explicit operator enable decision

Until this gate passes: the metric is advisory. It cannot launch, downshift, throttle, wait, or otherwise change AGT-2055 behavior.

Any future admission-enforcement change remains outside AGT-F3 and requires its own explicit operator authorization.

Dependencies that GO does not erase

Parallel Studio work

Calibration-grade quota capture

Cached boundary events now exist, but they must be upgraded or supplemented with fresh raw, same-run, same-window observations plus a documented settle policy. This plan consumes that completed contract.

Operator policy

First TokenEconomy NuGet publish

Studio package consumption remains blocked until the operator completes the one-time publishing step—even if implementation work receives GO.

← Architecture boundaries Back to overview →