Scoring rubric · Q3 2026

One hundred points
across seven criteria

Every criterion is scored 0 to 5 against the written anchors below, then multiplied by its weight. Two judges score each criterion independently. If they differ by two or more, a third judge scores and the median stands. The floors sit on top of the arithmetic and override it.

Weight distribution

Forty of the hundred points go to the product and the engineering behind it, forty-five to what you learned about Niha and whether you fixed it, and fifteen to the lifecycle artifacts that prove the design came before the code.

The anchors, in full

Anchors are what stop two judges scoring the same submission thirty points apart. A 1 or a 2 sits between the 0 and the 3; a 4 sits between the 3 and the 5.

Floors, bonuses and penalties

Floors

Applied to the weighted total before anything else. They cap, they do not subtract.

  • Fewer than eight verified findings across the programme caps the total at 60, however good the product is.
  • No findings at all from phases 0 to 3 caps it the same way. Findings that only appear at hardening time were reconstructed from memory, and reconstructed findings do not reproduce.

Bonuses

Added after the floor.

  • +5 for a finding class nobody else reported.
  • +5 for a killed or pivoted idea at Gate 2 with the evidence that killed it. That is a discipline we want and almost never see.

Penalties

Subtracted last. The total is then clamped to 0 and 100.

  • -10 per finding contradicted by its own excerpt.
  • -10 if the fresh-clone test fails, and criterion 1 cannot score above 2 of 5.
  • -5 per design artifact dated after the code it describes.

Score a submission yourself

The same arithmetic the judges panel runs, and the same the server computes. Try to make it produce a wrong answer.

Floors, applied before bonuses
Bonuses
Penalties
Weighted total
0/ 100

The reward bands

Level 1

Builder

90-100score out of 100
Reward: ₹25,000
Level 2

Contributor

70-89score out of 100
Reward: ₹15,000
Level 3

Starter

50-69score out of 100
Reward: ₹10,000

Bands are non-competitive: everyone who reaches one is rewarded at it, and several people can share a band. A band may go unawarded. Below 50 there is no reward, written feedback on every criterion, and an open invitation to the next round.

How judging runs

  1. Verification first. Before any judge opens the report, the CLI owner reproduces the findings and marks each one verified, not reproduced, or contradicted. Only verified findings count toward the floor.
  2. Fresh clone by a judge. Wipe, clone, follow the README, time it. Gate 6 is checked by someone who is not the author.
  3. Artifact dates against commits. Design documents are compared with the commit history of the code they describe.
  4. Two judges per criterion, scoring 0 to 5 independently against the anchors. Differ by two or more and a third judge scores; the median stands.
  5. Floors, bonuses, penalties applied to the weighted total in that order, then clamped.
  6. Written feedback per criterion within a week of Demo Day.

Deliberately not scored

  • Lines of code.
  • Number of stories closed.
  • Report length. A six-page field report with nine reproducible findings beats a twenty-five page one with three.
  • How clever the fix was. A merged fix is scored on whether someone else can trust it: before and after, a test, a demo and the documentation.

Full detail on the process, the findings format and the merge bar is on the challenge page.