Release regression corpus · 2026-09-06.v1
What has actually been measured
A transparent technical check of source selection, refusal, provenance, and change impact. This is regression evidence, not a claim about nationwide accuracy.
Bharat Career ID
Source-aware resolver
- 30/30 supported cases answered with the correct source
- 25/25 cases without a usable source refused instead of guessed
- 5/5 conflicting cases sent to citizen review, not decided automatically
- 30/30 selected values carry credential, version and policy
- 20/20 source-change cases handled correctly (100%)
Flat-profile baseline
Ordinary autofill
- 27/30 supported cases correct
- 28/30 cases that should stop were auto-selected
- 0 provenance tracked
- Change impact: silent re-use of stale values
Corpus composition
60 labelled cases across three groups
Two denominators are in use, and they are not the same. The three groups above describe how the cases were written (20/20/20). The results above are grouped by the outcome each case expects: 30 expect a supported answer and 30 expect the system to stop. The change group contributes to both.
- Supported cases span ten receiving fields and multiple fictional citizens.
- Unresolved cases cover missing, unsuitable, unverified, and unknown evidence policies.
- Change cases cover deletion, replacement, version updates, and conflicting reissues.
- The resolver selected no value in any of the 30 cases that required refusal or review.
Honest boundary
Applicant validation is still pending
No usability or time-saving result is claimed. Real adult applicants must complete the prewritten protocol before this gate can pass.
The flat-profile comparison selected 27 of 30 supported values correctly, but also auto-selected 28 of 30 cases that should have stopped for evidence review. The corpus was written after the resolver existed, so it is a release regression suite—not an independent benchmark.
Open the anonymous study recorderRun the validation test to verify the case count, expected decisions, provenance, baseline comparison, and change-impact results.
2026-09-06.simulated.v1 covers English and Hindi, mobile and desktop, and both interface orders. It checks the scripted journey only. synthetic walkthroughs, not people (12). Human participant evidence remains 0/12.
Inspectable product proof
What this release demonstrates—and what it does not
One precise failure
A flat profile repeats one name everywhere. The replay demonstrates why a Matriculation-name field needs a different sourced value from a current-legal-name field.
Regression evidence
The labelled corpus matched 60/60 expected outcomes and made 0/30 unsupported automatic selections. It was authored after the resolver and is regression evidence.
A fair baseline
The same synthetic inputs run through a flat-profile baseline. It answers many easy fields, but it also answers cases where the reviewed policy requires the system to stop.
Human approval remains visible
The citizen sees each proposed value, source, policy, and recipient before consent. A captured local record preserves what this demo handed off; it is not an official receipt.
Honest about limits
We clearly document what works, what is simulated, and what still needs validation with real applicants. Nothing on this page is a claim about a citizen’s real application.
Handles credential change
Deleting, replacing, or versioning a synthetic credential recomputes dependent drafts while an earlier captured Application Passport remains unchanged.