PRIVATE FINANCIAL-SERVICES DEMO
Forge Sandbox
Before you trust an AI change, see what it gets right—and what it misses.
Your sandbox access
Enter the private code supplied by Forge. You do not need an account or technical knowledge.
One-hour session · 60 runs per session · 10 runs per minute · Shared daily limits apply
One payment story, four views. Discussing a model upgrade? Start with the Evaluate tab.
1 · SET UP YOUR EXAMPLE
Tasks means payment reviews, not agents. This is an illustration, not a capacity limit.
Fictional workload assumptions, not measured usage
This is task volume, not the number of agents. Each task assumes 10,000 input and 1,000 output tokens, with 20% of input reused. When a comparison adds an execution step, the new-version estimate adds 3,000 input and 400 output tokens per affected task. These are fixed illustrative assumptions, not measured customer usage or an annual savings claim.
Ready for your first run.
2 · UNDERSTAND THE RESULT
Your answer appears here.
Start with the example as it is. Then change one thing and run it again.
- See what Forge found.
- Read the reason in plain English.
- Compare what changed.
Previous result and submitted inputs
Start here: one example, then one change.
Govern checks supplied evidence. Evaluate compares recorded answers and steps. Optimize explores model-usage cost, alongside those checks. Record reruns the payment evidence and shows its accountability record; it does not sign the Evaluate or Optimize comparisons. Start with Govern; you do not need to open the technical section.
- Choose an example and run it as supplied.
- Change the highlighted evidence or comparison, then run again.
- Read the new result and what would need human review.
This sandbox is not monitoring your systems. It calculates only when you press Run. It cannot interpret an arbitrary document or connect to a live agent.
Hold versus block: a hold pauses an action for missing evidence or review. A block rejects that request; a retry is a new request. In a customer integration, the customer’s protected system enforces these outcomes. This example demonstrates a hold and readiness for human review—not an executed payment or a live block.
Want to get more technical?
This is a separate Govern evidence request, not an editor for the Evaluate or Optimize examples. JSON is a structured form: field names on the left, values on the right. Start with a sample, change one value, and run it. Never paste confidential information.
What the numbers mean
- Confidence · 0 to 1
- Your supplied confidence in an evidence item. For example, 0.95 is a demo input—not a measured 95% accuracy rate or a guarantee that the evidence is true.
- Support weights · −1 to +1
- How strongly this item argues for or against each possible outcome. +1 strongly supports it; 0 is neutral; −1 strongly opposes it. The three weights do not need to add up to 1. They are not probabilities. A failed mandatory check cannot be outweighed by these scores.
- The three possible outcomes in this evidence example
- Send to a human reviewer: evidence supports review, not execution. Wait for more evidence: do not proceed on the evidence supplied. Escalate for closer review: a person needs to examine the issue. None authorizes a payment or trade.
The engine combines the structured evidence and applies its assessment checks. These controls are demo assumptions; in a customer integration, approved evidence mappings and policies must control them. Editing a title or narrative does not cause Forge to infer new scores.
JSON field guide
- subject / decision_question
- The label and question for the reviewer; changing them does not prove a fact.
- evidence
- The supplied items, each with an ID, source label, description, confidence and support weights.
- decision_options / abstention_action
- The available outcomes and the fallback when the assessment cannot complete.
- tenant_id / client_id
- Keep the sample values. They identify the fictional demo tenant, not your company.
- claim_support_level
- A supplied evidence classification, not independent verification of a real source.
Use decimal scores: 0.97, not 97. Confidence must be 0–1; support must be −1–1. Use valid JSON with double quotes and no duplicate field names. Maximum 24 KB and 20 evidence items. Manual edits disable the scoring controls so an old control cannot overwrite your text. Restore the sample to use those controls again.
The result appears in the result panel above. Open “Previous result and submitted inputs” to check exactly what was submitted. No engine source, keys or raw engine response is returned.
How this grows into governance
Accountability checks an individual proposed action. Evaluation checks whether a new version behaves better. Optimization asks whether a lower-cost option deserves testing. Governance brings those checks into your approval process: your evidence sources, your rules, named reviewers and a verifiable history of what changed. This sandbox does not create that production integration or approve a model release.