India-native EvaluationsIndian GST Evaluation Workbench
How models handle
GST decisions
Follow a model as it reads a record, applies a rule and decides whether to resolve, inquire or escalate. Compare each step with the expected outcome.
01 / MODEL COMPARISON
Results across the same scenarios
| Model | Core pass | Challenge pass | Passed / scheduled | Scored attempts | Unavailable | Accounted USD* |
|---|
Scores count strict passes across all scheduled attempts. See scoring, availability and case notes for interpretation.
02 / TRACE INSPECTOR
Inspect a model’s decision
03 / SOURCE EVIDENCE
Sources and expected outcomes
Each synthetic case pins its own date and tested provision; no result claims to cover later amendments. Expected values below are added after model execution and were not sent as answers. Challenge rows test only the named statutory condition; a row not barred under that clause is not thereby eligible for credit. Credit-note timing does not decide issuance or output-tax adjustment.
04 / INTERACTIVE WORKFLOW
Try a case yourself
Read the records and rules, choose an action and compare your answer with the evaluator.
05 / METHODOLOGY
Methods and case notes
Each public scenario pairs synthetic records with supplied statutory rules. Related scenarios can share a rule family. The evaluator checks required fields and applies authored workflow rules for inquiry and escalation. Source citations are checked; case-specific practitioner review is ongoing.
Results describe this public, supplied-rule workflow. Technical limits and availability are recorded separately. A step limit is a strict failure. An unavailable attempt has no pass/fail assignment. The model table follows protocol order. Azure costs are token-price estimates; OpenRouter costs use provider reports when available.
This run uses strict structured-action JSON across providers, not native provider tool calling. Exact versions, routes, output bounds, settings, retry limits and the cohort digest are below. No answer-quality retry, best-of-k selection or hidden reasoning retention. Model aliases on routers remain provider-reported version bindings, not cryptographic weight identities.
Frozen campaign configuration
RESEARCH MATERIALS
Inputs published before evaluation
The disclosed synthetic inputs, supplied rules and official-source references remain visible. Download the candidate case JSON →
Previous studies · Repeated-read study · Core GST workflows