Portable decision record
A reviewable bundle for one AI spending decision, with its supporting evidence, manifest and an offline verifier.
Enterprise AI controlARC1
Control for the autonomous enterprise.
Connect AI spending, quality comparisons and approved policy changes in one decision record.
Explore the interactive exampleCONTROL / ARC1
AI spending decisions
Understand the money and the decision behind it.
A provider bill, a gateway measurement and a coding agent's estimate describe different parts of AI spending. ARC1 preserves those differences while connecting the proposed change to its business case, comparison evidence and approved policy. The retained record gives engineering and finance a common basis for review.
The decision behind the spendExample
Usage contextComparisonPolicy record
Source records
Gateway usageOperating metadata
Provider recordsSource retained
Coding agentsMeasured usage
Workload contextOne decision record
QualityLatencyCost
Registered limitsSpending picture
Reconcile the sources without flattening their meaning.
Inspect provider-reported amounts alongside gateway measurements and agent estimates. Source coverage, attribution and the accounting period explain what a total represents and where a reconciliation needs further investigation.
- Usage counts and monetary amounts retain their own measurement basis.
- Credits, currency and the difference between usage and invoice periods stay with the source record.
- An unexplained difference remains an investigation item, rather than becoming an assumed saving or ungoverned spend.
Interactive example
Compare a model change.
Choose an alternative, inspect the paired sample and see what the declared limits support.
Synthetic inputsComputed in this browser
ARC1Interactive workspace
Choose a starting pointThen change any input to explore the difference.
01 / Set the context
Example workload · document classification
Three authored candidates use the same four example tasks.
Each pair contains baseline and candidate readings for the same task.
Technical evidence and change authority have separate standing.
Authored examples for exploring the workflow. Your choices update only this demonstration.
02 / Inspect the basisComparison record
One review item remains
4 of 5 example checks met. Follow the evidence questions before carrying the example forward.
4 / 5 checks met
- Quality difference
- −1.0 pts
- Candidate latency
- 1,650 ms
- Task cost difference
- −50.0%
Candidate minus baseline · mean rubric score / 100
Arithmetic mean over the selected paired tasks
$0.030 candidate mean / task · sample difference
Review checks4 / 5 met
Quality loss ≤ 2 pointsMet
Baseline mean 97.0; candidate mean 96.0 on the authored rubric.Mean latency ≤ 2,400 msMet
1,650 ms candidate mean across 4 selected tasks.Mean task cost ≤ $0.045Met
$0.030 candidate mean compared with $0.060 baseline mean.Four paired tasks requiredMet
4 of 4 authored task pairs are included in this comparison.Separate configuration reviewerReview
The example change awaits a separate authorized reviewer.
Next review step
Ask a separate authorized reviewer to examine the comparison and proposed configuration.
Inspect the selected evidence
- Proposed alternative
- Balanced candidate. A lower-cost candidate with similar rubric scores.
- Task pair 01
- Baseline → candidate: quality 97 → 96 / 100; latency 2100 → 1500 ms; cost $0.060 → $0.030.
- Task pair 02
- Baseline → candidate: quality 96 → 95 / 100; latency 2400 → 1800 ms; cost $0.065 → $0.035.
- Task pair 03
- Baseline → candidate: quality 98 → 97 / 100; latency 2200 → 1700 ms; cost $0.055 → $0.025.
- Task pair 04
- Baseline → candidate: quality 97 → 96 / 100; latency 2300 → 1600 ms; cost $0.060 → $0.030.
- Paired observations
- Four paired task observations.
- Configuration approval
- The example change awaits a separate authorized reviewer.
Illustrative review record · human decisions and subsequent actions keep their own standing.
The product
Control autonomous work.
ARC1 gives enterprise AI decisions an inspectable record. Teams can examine usage and cost, compare an alternative against agreed limits, authorize a policy change and retain the evidence of what followed. Authority, technical observations and financial readings stay connected to the decision they support.
Enterprise AI journeyCONTROL
Inside ARC1
Give each AI change a decision record.
Bring usage, cost, quality comparisons and policy decisions into the same review. AI platform teams can inspect why a change was proposed, which limits it must meet, who approved it and what the follow-up evidence establishes.
- AI platform teams
- Technology leaders
- Finance and control owners
How the work moves
Establish the spending pictureUsage and cost context
Read gateway measurements, provider cost records and coding-agent telemetry. Keep each source's measurement basis visible as teams examine usage by team, project and workload.
Set the comparison limitsRegistered comparison
Register the proposed alternative and its quality, latency and cost limits before collecting comparison observations. Give reviewers an explicit basis for accepting or rejecting the change.
Review the observed differenceEvidence for the decision
Import paired observations and inspect how the alternative performs against the registered limits. Keep a forecast, an observed comparison and finance-confirmed value distinct.
Approve the policy changeApproval and target readback
Review the configuration plan for a supported target. A separate authorized approver accepts the exact change before publication; readback records whether the target matches the plan.
Keep the follow-up attachedContinuing decision record
Retain the approval, acceptance limits and later observations with the decision. Inspect configuration drift and review financial evidence against the approved baseline.
What the work produces.
Policy publication receipt
The approved configuration change, signed audit record and target readback give reviewers a trace from authority to observed configuration.
Value record
An approved baseline with forecast, provisional and finance-confirmed readings kept on their own basis.
Comparisons establish results for the observations supplied. ARC1 reads operating metadata and publishes approved policy to supported targets; test execution and request routing remain with the organization's existing systems.
In practice · illustrative scenario
Review a model change for a recurring workload.
A platform team is considering a different model for a recurring document task. Finance wants a clear spending basis, and the workload owner needs acceptable quality and response time.
- 01
Name the comparison
Record the proposed alternative and its acceptance limits, then bring in paired observations from the team's evaluation system.
- 02
Examine the decision
Review quality, latency and cost together. Keep a forecast separate from the observed result and any later financial confirmation.
- 03
Retain the follow-up
For a supported target, obtain separate approval for the configuration change and attach target readback and later evidence to the decision.
The useful outcome
The illustrative outcome is a reviewable change record: what was compared, which limits applied, who authorized the policy and what the follow-up supports.
A closer look
Know what you can do with ARC1.
Does ARC1 replace our gateway or evaluation tools?
ARC1 works alongside them. It reads operating metadata and evaluation observations, publishes approved policy to supported targets and records readback. Your existing systems execute tests and route model requests.
What can finance take from the record?
The approved business-case baseline, source amounts and periods, recorded comparisons, and the evidence attached to a finance review. Forecasts and provisional readings stay identifiable so they can be reconciled on the appropriate basis.
Can a recipient check an exported decision independently?
The decision bundle includes a manifest and offline verification material. A recipient checks it against a trusted key obtained independently. Those checks establish the integrity of the retained record; the recipient still evaluates the underlying evidence and decision.
