School Evidence Coach

What problem does this solve?

Students often have a claim and a source, but no simple way to separate supported, contradicted, and insufficient evidence without turning an AI tool into a truth oracle.

School Evidence Coach is a bounded Assignment Mode for that exact school-life problem. A student enters an assignment claim and one evidence excerpt. The tool returns one of three relationships and gives a concrete next evidence step.

The goal is not to tell students what is true. The goal is to make evidence reasoning visible, limited, and explainable.

Judge quick test

Use these three exact cases from the committed test suite:

SUPPORTED

Claim: School garden reduced cafeteria food waste Evidence: School garden project reduced cafeteria food waste by 18 percent

CONTRADICTED

Claim: School garden reduced cafeteria food waste Evidence: The school garden did not reduce cafeteria food waste

INSUFFICIENT

Claim: School garden reduced cafeteria food waste Evidence: Students reported liking outdoor classes

The tool also fails closed to INSUFFICIENT when the claim or evidence input is missing.

What was built during the CSC event window

School Evidence Coach is a CSC 2026 event-window prototype for students working with conflicting or incomplete sources.

The dedicated CSC implementation includes:

standalone Assignment Mode UI;
student-facing SUPPORTED / CONTRADICTED / INSUFFICIENT explanations;
visible claim ceiling and next-evidence-step guidance;
mobile/readability styling;
deterministic event-specific tests;
runtime/CI evidence;
CSC-specific provenance and AI-use disclosure.

Event-window regression fix

A fresh review found a real directional-language bug in the compact classifier: words such as "decreased" or "lower" had been treated as generic negation. That could incorrectly contradict a claim about reduction, while opposite-direction language such as "increased" could pass as support.

The fix was developed test-first during the CSC event window. Two regression cases were added before the implementation change:

REDUCED claim + DECREASED evidence -> SUPPORTED
REDUCED claim + INCREASED evidence -> CONTRADICTED

The first test commit intentionally failed in GitHub Actions, proving the regression tests caught the existing behavior. The classifier was then changed to separate explicit negation from directional polarity. Downward terms such as reduced/decreased/lower/less are treated as directionally aligned, while upward terms such as increased/higher/more are treated as the opposite direction. Explicit negation such as no/not/never/did not/cannot remains contradiction logic.

Verification

Original tested implementation commit: 28f51507e2fa470ede0e6138a6483c499b652df7
Original CI trigger/readback commit: 98f2321a6f702c7d7eb1b5e219d9f3669c6a4d33
Regression RED test commit: 6288180d65ab25a665f8c78008dbece3a4a2e58d
Directional-polarity fix commit: a9d2298ba2f820c1a1d31baca3c085a5b1676e65
Deterministic committed tests: 6/6 PASS
Fresh GitHub Actions run: SUCCESS
Fresh CI run: https://github.com/lbkapcsolat-dotcom/ProofPath/actions/runs/34680171501
Source / event-window implementation: https://github.com/lbkapcsolat-dotcom/ProofPath/tree/main/csc-2026
Runtime receipt: https://github.com/lbkapcsolat-dotcom/ProofPath/blob/main/csc-2026/RUNTIME_RECEIPT_2026-09-05.md

A fresh native Chromium evidence run also rendered the three expected runtime labels from the earlier frozen demo HTML and coach.js bytes: SUPPORTED, CONTRADICTED, and INSUFFICIENT. That earlier runtime receipt predates the directional-polarity regression fix, so it is not claimed as runtime evidence for the new fix.

Pre-existing work disclosure

The broader ProofPath repository predates the hackathon and is used only as repository context. Its earlier evidence-classification concepts, prior interfaces, prior tests, and prior claim-ceiling work are not represented as CSC-created work.

The dedicated csc-2026 implementation, tests, runtime evidence, CI work, and submission materials are CSC event-window work.

AI-use disclosure

AI assistance was used for planning, coding support, test design, documentation, wording, and review. The entrant retains responsibility for the project decisions, understands the implementation, and can explain how it works.

The project does not collect private or sensitive student information.

Claim ceiling

TESTED IMPLEMENTATION ONLY

EDUCATIONAL_EVIDENCE_ASSESSMENT_ONLY

School Evidence Coach is a reasoning aid. It does not certify truth, replace teachers, establish source authority, claim universal correctness, or verify scientific truth.