What has been built
- A small, synthetic DuckDB demonstration.
- Versioned metric definitions and original-result replay.
- Source reconciliation and receipt-backed reporting evidence.
- A retained comparison with an independent BigQuery baseline.
tokenledger / By Chandler Klose
One activity stream.
Every reported
number reproducible.
A reference implementation for AI businesses that meter tokens and seats. Report a business metric, explain a change, and reproduce the original result.
Built from the SOX-controlled subscriber-metrics work I ran at SiriusXM for 19 years.
The demonstration, explained
Follow one number through a change.
The local demonstration uses synthetic data on DuckDB. This page describes that workflow; it is not a live connection to the database.
- 01 / Report
Start with July’s token yield.
Calculate recognized revenue per million tokens. Retain the inputs and definition needed to reproduce the result.
- 02 / Explain
A late invoice arrives.
Revenue changes while the token denominator stays fixed. A bridge explains the movement between the two results.
- 03 / Reproduce
Replay the original receipt.
Recover the earlier answer using its original evidence and definition, even though the system now knows more.
The model
Keep the events. Make the rules explicit.
A fourteen-column activity envelope carries event identity, timing, and JSON feature data. Typed query-time projections give those features meaning. New reporting questions can reuse the same stream while retaining explicit definitions and evidence.
- 01 / PersistActivity streamSource events + JSON features
- 02 / InterpretTyped projectionsQuery-time structures
- 03 / CalculateVersioned definitionsRules + reporting cutoffs
- 04 / ReportReported resultMetrics + explanatory bridges
- 05 / ReproduceReceiptInputs + calculation identity
Scope
Local execution. Cloud evidence. Explicit limits.
What still needs to be assessed
The retained BigQuery comparison measured 51.47% fewer billed bytes; 14.96 times median slot time. Native execution was slower in the first measured pair. These results do not establish a general speed or cost advantage. Comparative agent accuracy and token efficiency remain unmeasured.
Reproducibility alone does not prove source completeness, approval of an accounting policy, or SOX operating effectiveness.
For people and their agents
Make the work assessable.
The public reference offers three paths: run the small example, understand how the result is derived, and assess it through an open-book reproduction exercise.
Questions for an evaluator
- Can you reproduce the reported token yield?
- Can you explain the late-invoice bridge?
- Can a metric definition change without erasing the old answer?
- Does incomplete source evidence hold publication?
Release status
v0.2.0-rc.6 is published. Run the small demo, understand the model, or try the agent challenge.
The release includes the terminal recording, executable tasks, and retained measurement evidence.
What would your team ask next?
The interesting work begins with your source contracts, customer identities, reporting policy, and the people who own the number.