MEASUREMENT NOTES & DISCLAIMER

A ledger,
with its workings shown.

This is an effort portrait of Asterwyn, not a measure of quality, intelligence, hardware consumption or human labour. It contains recorded counts and explicitly labelled estimates. No financial metrics are displayed.

What is included

The local usage history of “Assess NWN EE module creation”, beginning 5 September 2026, plus locally discoverable historical worker conversations explicitly linked through parent metadata. Unrelated chats and this dashboard-building side conversation are excluded. Missing provider records, external services and undiscovered worker logs are not invented. These are the available recorded totals, not a claim of universal coverage.

Tokens, counted once

We sum per-response token_usage_record.usage entries belonging to each selected thread, deduplicated by response ID. Mirrored token-count events and cumulative thread totals are not added again. The total is input plus output; cached input is already inside input, and reasoning is already inside output. “Visible output” is output minus recorded reasoning. It includes generated code and tool instructions, not only chat replies.

Cached context is still token traffic, but is reused context rather than freshly generated work. The chart’s “New input + output” view removes cached input. Percentages in the account allowance are not used to infer tokens. OpenAI prompt-caching documentation.

Hours and activities are estimates

Agent-hours sum the gaps between consecutive usage records within each conversation when the gap is no more than five minutes. Longer gaps are excluded. This approximates active working periods and includes short tool waits; it can omit long uninterrupted tasks. It is not a GPU timer, human timesheet or the time elapsed since the project began. Parallel worker periods add together, so agent-hours may exceed wall-clock hours.

Activity categories are heuristic labels based on nearby tool names and operations: research and design, building and scripting, art/film/sound, validation and recovery, publishing and operations, and planning and coordination. A response inherits the most recent tool category. Categories can overlap conceptually; each response is assigned once. Tool counts are recorded invocation entries, not a count of every nested action or successful operation.

The Seedance film

Joel supplied two Together AI Video Inference entries for 7 September 2026, model bytedance/seedance-2.5: 16.799 and 25.467, totalling 42.266 reported units. The supplied unit label does not establish seconds, frames or millions of tokens. We preserve those numbers without pretending that a reported unit is a text token.

The separate visual-token estimate uses actual duration, width, height and frame rate measured from each unique retained source MP4 in the film’s source folder. Byte-identical clips are deduplicated. The proxy is seconds × width × height × frames per second ÷ 1024, based on the Seedance-family video-token estimation formula. It measures the retained visual volume, not confirmed Together inference usage. It excludes discarded generations not retained in that folder, input references, additional internal processing and audio. Transcoding, retiming or editing can also change the proxy. It must not be added to the recorded text-token total as equivalent computational work.

Sources: Together AI Seedance 2.5 model, Together video API documentation, BytePlus video-token estimation reference. Provider documentation was checked on 13 September 2026. The source entries are retained as supplied; currency columns are omitted throughout this site.

Freshness and privacy

A local collector reads new usage records every two minutes while the publishing computer is running and connected. The page fetches its aggregate every thirty seconds and displays the last successful collection time. If collection stops, the timestamp ages visibly; no synthetic activity is generated. An active model response may not appear until its usage event is written.