Workload Catalog for AI Media and Human Work

1) The real problem: “1,000 jobs/day” is not a capacity unit

Adding AI requests, image generations, video rendering, and human review produces a simple number but hides the bottleneck. Each class has different units, service time, concurrency, cost, queueing, and staffing constraints. This lesson builds Cross-Modal Workload Catalog v1—not a forecast and not an admission decision.

2) Bound the workload by business value

AWS Well-Architected describes a workload as a set of components that together deliver business value. That is useful boundary evidence, but AWS does not define the MAKNOW AI/media/human taxonomy. Each catalog row still needs a local business-value and operational boundary.

3) Preserve the native work unit

Measure AI as requests, images as images, video as rendered video minutes, and human work as review minutes. Renaming all of them “jobs” before capacity modeling destroys information. Any later normalization must be explicit and dimensionally meaningful.

4) Keep resource planes separate

AI inference, GPU/media generation, media rendering, and human operations do not consume interchangeable capacity. The catalog exposes the plane so a blended average cannot hide the real constraint.

5) Arrival pattern and burst are not daily volume

Two workloads can both report 1,000 units/day while one arrives evenly and another has a 4x peak hour. Daily totals alone cannot explain concurrency or queue pressure.

6) Concurrency, service time, and latency target are different fields

Concurrency is simultaneous work; service time is resource occupancy per unit; latency target is an experience or service objective. Conflating them produces bad capacity reasoning.

7) Keep media size and compute/token units only where they mean something

Tokens/calls can drive AI cost; rendered minutes, compute-minutes, and media size can drive video cost. n/a is better than invented precision. The catalog should expose heterogeneity rather than erase it.

8) Human minutes are a workload

Do not smear human review into a vague percentage on technical work. If only a high-risk subset requires review, represent that subset and its review minutes explicitly because staffing windows and backlog behavior differ from GPU or API queues.

9) Retries and failures can amplify demand

If 10% of calls retry, effective demand is not equal to initial arrivals. Retry policy, failure mode, and human rework must stay visible or downstream forecasting starts from understated demand.

10) Baseline: four rows, four native units

The catalog reaches CATALOG_READY only when native units, arrival/burst, concurrency/service, cost, queue, observability, and provenance are visible. Summing these rows into one total job count is dimensionally invalid.

11) Wrong shortcut: call everything a job

Relabeling request, image, rendered video minute, and review minute as job makes the table look uniform but destroys capacity evidence. The bounded lab state is DISPUTED, not success.

12) What states do negative cases produce?

Missing critical evidence → HOLD. Expired evidence → STALE. A material change that has not been recomputed → PARTIAL. Generic-job collapse → DISPUTED. These states prevent a decision stronger than the evidence.

13) Lab: eight fixtures and 64 assertions

The lab is offline and requires no API key. Eight fixtures cover the baseline, generic collapse, doubled media without recomputation, missing evidence, higher burst, hidden retry amplification, stale evidence, and final transfer. The oracle uses visible inputs and explicit rules only.

14) Transfer: double video duration, 20% fewer AI requests, review only high-risk work

The transfer reduces AI demand from 1,000 to 800 requests/day, recomputes video-native demand after duration doubles, and keeps human review as a separate workload for only the high-risk subset. The correct result is a rebuilt catalog—not a new blended job total.

15) Delivery checklist

Business-value boundaries are explicit; native units are preserved; arrival/burst, concurrency, service/latency, cost/human/retry/failure, queue/observability, and evidence references are visible; missing evidence is not coerced into success; transfer is recomputed; and percentile forecasting remains owned by ARC-014.