# Reprice from measured work

*Keeping 90-minute exports affordable · Measured task durations + cost model. From the portfolio of Guilherme Rohden. Excerpts are quoted verbatim; internal paths and identifiers are redacted.*

Measured rendering time (4.2 s per output second against 2.78 modelled), the cost comparison, and the 1.5 to 2.25 repricing decision.

## The decision

An export had to pay for itself on its own, because an outside assistant can skip the paid planning steps and only export. The trade is fewer export minutes per included credit grant (credits are HYVIRAL’s prepaid units).

## From the source

**Capture measured on run 5.** One run of 45 tasks; an earlier pooling error is corrected:

> **Capture measured on run 5** (45 `job_task_complete`, same release and Job shape as run 4): 443–658 s (median 503 s = 4.2 s per output second), Σ task wall 23,203 s; per-chunk sizes and peaks match run 4 within 2 %. The two runs are separate datasets of 45 tasks each; the Slack note of 9 Oct 21:26 that pooled them as "90 tasks" used run 4's distribution only.

**Cost comparison and pricing decision.** Amounts model infrastructure rates and measured task durations:

> Per delivered 90-minute export at the measured capture speed (4.2 s per output second against the ladder's modelled 2.78): capture **$4.24** + finalize **$0.135** + the model's one-download + ten-day-retention **$0.81** = **≈ $5.19**, against **$4.07 (Pro) / $4.15 (Hype)** allocated revenue at 2,575 credits — a deficit of about **$1.0–1.1** per export; the 1.5× rail does not recover the export. This is an **owner decision before step 2**: raise the long-form multiplier (the 1.5 floor allows it), speed the capture (the 90-minute chunks ran 448–639 s against a 120 s span), or both. **Owner decision, 10 October 2026: the default long-form multiplier is 2.25 (the 1.5 floor stays).** A 90-minute export now quotes **3,850 credits** (≈ $6.1 realised on Pro against the ≈ $5.19 measured basis) and a 3-hour one **7,675** (was 5,125 at 1.5×); the capture-speed investigation is step-2 work and the rail is re-priced after it. Production is unaffected until the beta flag is on there.

**The retained floor:**

> 2. **Long-form render multiplier: 1.5× above 60 min**, as a **floor** — raised to a **2.25 default on 10 October 2026** after the step-1 measurement (§3.1). Final pricing is settled from measured export-only / direct-agent costs after the 90 and 120 min steps, not from these models.

**Model versus measurement:**

| Input or choice | Value |
|---|---|
| Capture task seconds per output second | 2.78 assumed → about 4.2 measured |
| Estimated 90-minute export cost | About $5.19 |
| Pro revenue allocated at full included-credit use | About $4.07 at 2,575 credits |
| Default multiplier above 60 minutes | 1.5 → 2.25 on 10 October 2026 |
| Revised quote for that 90-minute export | 3,850 credits |

Parallelism shortened elapsed time, not billable task work.

## What this shows

A measurement changed a pricing decision: the modeled deficit justified raising the default, and the configurable 1.5 floor remained. This is a cost model built on one download and ten days of retention, not invoice data.