Public workflow demo. The figures below demonstrate accounting with scripted fixtures; no measured multi-step model results are published here. The historical single-prompt pilot remains a separate release.

Useful-work economics · v1

What does an accepted job actually cost?

Scripted accounting fixture — no model or agent was run

A separate workflow basket. Existing published pilot results retain their original scope and methods.

Final accepted jobs: 2 / 3
API + tools / accepted job: $0.055
Known recorded cost: $0.11
Total job wall time: 4.0 min

Every recorded expense stays in.

Known amounts in the trace, before any scenario additions. Empty expense categories total zero within this ledger; omitted real-world overhead remains outside its scope.

Cost category	USD	Provenance
Model	$0.09	Fixture
Tool	$0.02	Fixture
Review	$0.00	Fixture
Rework	$0.00	Fixture
Error	$0.00	Fixture

All recorded costs / accepted job: $0.055. Includes logged human and error costs.

Declared run policy

Retries are visible.

Attempts / job limit: 2
Semantic retries: Allowed
Job timeout: 5.0 min
Infrastructure retries: 1
Semantic retries used: 1
Failed infrastructure-attempt costs: $0.01
API + tools / accepted, excluding infra failures: $0.05

An infrastructure retry follows a timeout or service error. A semantic retry follows a completed, rejected answer. All remain charged in the primary metric. The alternate metric excludes only failed infrastructure attempts; a successful retry remains charged.

From attempts to final accepted jobs.

Acceptance is recomputed from the pinned task and its final artifact. There is no ranking or pooled cross-vertical Octane score here.

Job / task version	Final result	Attempts	API + tools	All recorded	Job wall time
`fixture-job-1` workflow-csv-quality · v1.0.0	Accepted Schema and all declared artifact values match	2	$0.05	$0.05	1.7 min
`fixture-job-2` workflow-invoice-reconciliation · v1.0.0	Accepted Schema and all declared artifact values match	2	$0.035	$0.035	1.7 min
`fixture-job-3` workflow-support-resolution · v1.0.0	Rejected Artifact values differ from the answer key	1	$0.025	$0.025	40.0 s

Modeled assumptions · additional overhead

Price the work around the AI.

These inputs add expenses that are absent from the trace. Avoid entering costs already logged above. Zero is an assumption too. Assumed human rework never converts a rejected AI job into an accepted one.

Scenario inputs

Additional review / attempted job: 0 min
Additional rework / detected failed job: 0 min
Human hourly cost: $0.00 / h
Failure detection rate: 100%
Undetected error cost: $0.00

Modeled additions and scenario

Additional review: $0.00
Additional rework: $0.00
Undetected errors: $0.00
Modeled additions: $0.00
Scenario cost / accepted job: $0.055

Scenario = all recorded costs + modeled additions, divided by accepted jobs. Assumptions do not replace missing recorded expenses.

Versioned tasks and grades

Original multi-step public samples.

Each task combines multiple source artifacts and a concrete deliverable. The public answer fixtures test acceptance contracts; they are not private benchmark tasks or evidence about model performance.

Original public sample

Audit a CSV pipeline against control totals

data_operations · workflow-csv-quality · v1.0.0

Read the processing contract and source files.
Deduplicate event IDs and apply exclusions in contract order.
Join accounts, aggregate valid posted events by region.
Compare amount and count controls and write the report.

Inputs and acceptance contract

inputs/events.csv
inputs/accounts.csv
inputs/control.json
inputs/contract.md

Output: quality-report.json
Grader: structured-exact/v1

Original public sample

Reconcile invoices, settlements and credits

finance · workflow-invoice-reconciliation · v1.0.0

Read the invoice ledger and policy.
Deduplicate payments, then apply status eligibility.
Join approved credits and compute every invoice balance.
Reconcile totals and write the final artifact.

Inputs and acceptance contract

inputs/invoices.csv
inputs/payments.csv
inputs/credits.csv
inputs/rules.md

Output: reconciliation.json
Grader: structured-exact/v1

Original public sample

Resolve a defect exception without a duplicate refund

customer_service · workflow-support-resolution · v1.0.0

Read the ticket and verify the order and evidence.
Apply ordinary return and defect-exception windows.
Check stock and subtract previously settled refunds.
Check escalation threshold and prepare the decision artifact.

Inputs and acceptance contract

inputs/order.json
inputs/ticket.json
inputs/inventory.csv
inputs/refunds.csv
inputs/policy.md

Output: decision.json
Grader: structured-exact/v1

How to read this

One accepted job, counted once.

Execution accounting

Every attempt stays in the cost ledger, including failed answers, tool charges and infrastructure retries. Acceptance applies to the final artifact of a job; a successful intermediate call cannot add another accepted job.

API + tool cost per accepted job
All model and tool costs across all attempts ÷ final accepted jobs

A zero accepted-job denominator is unavailable. An unknown cost remains unknown: a known subtotal is not a complete bill.

Evidence and assumptions

Trace costs retain their source and provenance. Scenario additions price extra review, rework and undetected errors using stated assumptions. They do not become measured expenses or change the acceptance verdict.

Job wall time is reported separately from money. Parallel tool durations are not summed as elapsed time.

Imported traces rely on the recorder to capture every operation. Source labels and file hashes establish consistency, not audited billing or a complete enterprise cost. Omitted costs cannot be detected.

These public samples test the workflow contract. They do not establish production reliability, general agent capability or a cross-vertical Octane index.

Road to measured workflows

The next evidence needs real runs.

Instrument a bounded pilot

Freeze tasks, grades and retry policy; add provider and tool recorders; approve a capped run budget. Reconcile every billed line, record timed-out jobs and audit task acceptance independently.

Broaden before publishing comparisons

Add representative tasks and a private holdout, study variability across jobs, and publish workflow evidence as its own dated release. Keep the existing single-prompt pilot and modeled calculator identifiable.

Recorded cost sources

Event	Kind	Provenance	USD	Source / evidence reference
`fixture-job-1-attempt-1-model`	model	fixture	$0.02	Scripted accounting examples only; not provider prices, bills or model measurements. scripted fixture-job-1-attempt-1; not a provider record
`fixture-job-1-attempt-1-tool`	tool	fixture	$0.005	Scripted accounting examples only; not provider prices, bills or model measurements. Simulated source-read event; task inputs are bundled.
`fixture-job-1-attempt-2-model`	model	fixture	$0.02	Scripted accounting examples only; not provider prices, bills or model measurements. scripted fixture-job-1-attempt-2; not a provider record
`fixture-job-1-attempt-2-tool`	tool	fixture	$0.005	Scripted accounting examples only; not provider prices, bills or model measurements. Simulated source-read event; task inputs are bundled.
`fixture-job-2-attempt-1-model`	model	fixture	$0.01	Scripted accounting examples only; not provider prices, bills or model measurements. scripted fixture-job-2-attempt-1; not a provider record
`fixture-job-2-attempt-2-model`	model	fixture	$0.02	Scripted accounting examples only; not provider prices, bills or model measurements. scripted fixture-job-2-attempt-2; not a provider record
`fixture-job-2-attempt-2-tool`	tool	fixture	$0.005	Scripted accounting examples only; not provider prices, bills or model measurements. Simulated source-read event; task inputs are bundled.
`fixture-job-3-attempt-1-model`	model	fixture	$0.02	Scripted accounting examples only; not provider prices, bills or model measurements. scripted fixture-job-3-attempt-1; not a provider record
`fixture-job-3-attempt-1-tool`	tool	fixture	$0.005	Scripted accounting examples only; not provider prices, bills or model measurements. Simulated source-read event; task inputs are bundled.