Solar Voice · Readership Response — Fidelity vs Outcome

Reading the proof — validation lens2 🟦 Graph · real node/edge 🟪 Hybrid · grounded + rendered 🟨 LLM · voice, no fabrication 🟩 Data · own, not inflated ⬜ Fixed · static 🟥 Care · two-tier holds readership is read in the same ink — leading vs lagging

Readership Response

— telemetry is the value function

FIDELITY (LEADING) vs OUTCOME (LAGGING)

Two readings of the same press run.

Every proof is judged on how readers actually responded. We read it twice — once ahead (the live edition's mechanism-fidelity, the sim's advisory score) and once after (real version-stamped production telemetry, per (prompt × corpus) arm). Held together, they are the calibration loop.

Advisory · view-only

The mechanism-fidelity score is a lens, never a ranker — it never auto-ranks or auto-promotes a candidate (I8 / I10). Pick on outcome first, then sim. A human grants the imprimatur; the score only informs the eye. Nothing on this screen mutates production.3

The Calibration Loop

— does the sim track the readership?

OUTCOME · PER (PROMPT × CORPUS × VARIANT) ARM

Each row is one P×C arm of real, version-stamped traffic. Activation = a 2nd distinct active day within 72h, joined days → run → arm. The live arm is crowned. This is the lagging ground truth — pick on it first.

SIDE BY SIDE · SAME BEAT-EVENT AXES

● Live edition reading config/live…

Mechanism-fidelity · leading · sim Outcome · lagging · real telemetry

Pick on outcome first

The leading bar comes before anything ships — the cheap signal. The lagging bar is ground truth. When they sit close, the sim is earning its keep. Committed-30d and graduation depth accrue over time — shown honestly, never fabricated.8

The Funnel

— real event counts + distinct users, all arms

collectionGroup('events') · LIFETIME COUNTS

The raw signal behind the readings above — every governed event, counted across all arms. These are volumes, not rates; the rate-of-warm-users they imply is what Committed and Graduation become as the cohort matures. Each cell also shows the distinct users behind those events, so a high count from a few heavy accounts is never misread as many people.

What Each Reading Means

— operational anchors, I13

Activation weekly · measured now

Returns for a 2nd distinct active day within 72h. Computed live per arm by joining days → run → arm. The leading indicator of graduation — never a substitute for it.

days/{YYYY-MM-DD} → run → arm

Committed monthly · accruing

The 30-day "would-miss-it" reliance habit. Tracked by event volume now; the per-arm rate firms up only once the cohort is 30 days deep.

attribution_recorded · outcome_checkin

Graduation quarterly ★ · accruing

The business North Star: a committed listener makes any paid purchase → B!G. No-gate; over-deliver free. Depth accrues across quarters.

graduation_intent

Listen-depth per-session

Played-through vs tapped-and-left, read from endReason. The felt dose; surfaced to the pioneer only, never to the user.

listen_ended · area_stacked

Mechanism-fidelity leading · advisory

The live edition's sim-score — cause-moments present in the transcript. Not a metric forecast; uncalibrated until validated against the outcome table. The fidelity field of config/live.simScore.

scoreSim() · fidelity{} · advisory

Share · viral no-gate

A felt win mints a public capability-URL — the single no-gate viral vector; carries no private dossier (I16).

share_initiated

An Activation Lift Is Not a Win

— how to read these two signals

THE READING RULE

Pick on outcome first, then sim. When fidelity (leading) and outcome (lagging) sit close, the sim is calibrated — trust it to tune the next setting fast. When they diverge — high fidelity, weak readership — the simulated-user or the scoring is wrong; fix that, not the prompt. And a lift that degrades graduation or trips an invariant is a REJECT, not a win — activation is the leading indicator of graduation, never a substitute. The invariant floor settles eligibility on the Proof Spike before these readings even begin (a flag, not a ranker). Outcome judgment never overrides the floor.11

View-only — this is the Readership room

Nothing here writes to config/live. To act on a reading: order a print run from the Print Run room (A/B a candidate), or grant the Imprimatur when a pair has earned it.

Marginal Apparatus · Operational Anchors

1: Imprimatur. The pioneer's lock — a write to config/live (runtime-config flip, app-cached, no deploy). The readership readings on this screen are the telemetry half of what justifies it; one live (prompt × corpus) pair at a time.
2: Validation lens. Each surfaced component is coloured by what proves it true: 🟦 graph · 🟪 hybrid · 🟨 LLM · 🟩 data (the user's own, never inflated) · ⬜ fixed · 🟥 care. Here the lead bar carries the 🟪 hybrid pigment (sim render) and the lag bar the 🟩 data pigment (real readership) — the same ink key the Compositor's Bench uses.
3: Advisory, never automatic. The mechanism-fidelity score is view-only — a human-initiated read is fine; an automatic ranking or promotion is not (no auto-mutating loop, I8 / I10). Humans pick; the score is a lens. Pick on outcome first, then sim.
4: The coupling. Prompt × corpus are not independent — what goes live is a pinned pair, and every reading on this screen is attributed to a specific promptVersionId × corpusVersionId arm. There is no "prompt-only" telemetry; every turn stamps both IDs plus its variant.
5: The calibration loop. Fidelity (leading) and outcome (lagging) read together. When fidelity tracks outcome, the sim can be trusted to tune fast; when they diverge, the simulated-user or the scoring is wrong — fix that, not the prompt. The sim's value is only as good as its calibration.
6: Mechanism-fidelity = cause-moments, not a forecast. Computed by scoreSim() at sim time from the transcript: did day-1 plant the return-hook, were turns corpus-grounded, did the reframe sit on a concept, did attribution fire, did the arc progress, were turns concise. It is uncalibrated until validated against outcome — never read as a prediction of the metric. Surfaced here as the live edition's simScore; overall is the fidelity, eligible the floor verdict.
7: Activation. Returns for a 2nd distinct active day within 72h; warm-data retention elbow ≈ +20.6pp backtest, recalibrated on the first cold cohort. Day-markers at profiles/{uid}/days/{YYYY-MM-DD}, joined to the run's arm. The one outcome metric computed live per arm.
8: Committed & Graduation accrue. Committed (the 30-day "would-miss-it" habit) and Graduation (the quarterly paid → B!G North Star) are lagging by construction — a per-arm rate can't exist until the arm's cohort is 30 / 90 days deep. The console tracks their event volume now (the funnel below) and surfaces the rate as accruing rather than fabricating a number. This is a real console: where there is no data yet, it says so.
9: Graduation — the North Star. A committed (would-miss-it) listener makes any paid purchase along the funnel → B!G. Quarterly. No-gate (I9): over-deliver free, expand the horizon, invite up — never gate or ration to chase it. Event: graduation_intent.
10: Listen-depth. Played-through (ran to the end / fell asleep — the intended default) vs tapped-and-left vs user-partial vs interrupted, read from the endReason on listen_ended. Never surfaced to the user — magical, not tracked (no surveillance, R6 / I16).
11: The reject-rule. A candidate that lifts activation but degrades graduation or trips an invariant is a REJECT, not a win — activation is the leading indicator of graduation, never a substitute. The invariant floor (care I15 · honesty I17 · no-gate I9 · no-fabrication I7 · concise) is a flag, not a ranker, settled on the Proof Spike; it hard-flags ineligible regardless of the readership numbers. Everything measured (I12).