i/eIndia-native evaluations Development preview

RESEARCH WORKBENCH / RECON-72

Test the work.
Not just the answer.

Explore how an agent compares invoices, checks arithmetic and dates, and decides when to ask for help.

—development scenarios
8conceptual roots
0statutory-validity claims
0recorded live attempts
Read this as a methods preview. These cases use an invented workflow policy, not current GST law. They are exposed development examples with shared roots—not a holdout, a model ranking, or professional advice.
Loading

Loading case library…

Interactive demo · you are the agent · no model calls or spend
0 / 6 stepsAwaiting inspection
The evidence is not pre-digested.
Start by reading the book, statement and supplier record.

Submit your decision

Derive the fields from the evidence and fictional policy below.

Case-specific fictional policy

Measured results, without invented wins.

Every prespecified attempt counts. Transport incidents are separate from evaluated failures.

01 / INPUTS

Primitives, not conclusions

Invoice references, independent entity aliases, line items, decimal amounts and dates. The revised campaign hides these human-facing case titles and mechanism labels from the model, along with the private evaluator's expected answer.

02 / EXECUTION

One bounded episode

Up to six actions, the same evidence tools and policy for every model, and one terminal submission. Paid campaigns run through an operator CLI using direct Azure Foundry. The browser cannot start one.

03 / EVALUATION

Check more than the label

Decision, candidate count, matched row, line-rounded tax, expected period and evidence inspection are checked independently. A plausible conclusion with wrong arithmetic does not pass.

04 / LIMITS

Development evidence only

Seventeen scenarios share eight conceptual roots. English only. No representative GST coverage, legal correctness, held-out generalization, training benefit or regulatory assurance is established.

Read the exact fictional policy