[tnalbant]

Closes #28.

Adds examples/multilingual/ with six cases in Hindi and Tamil, following the structure of examples/extraction:

• hindi-sentiment-positive, hindi-sentiment-neutral and tamil-sentiment-positive: sentiment classification with the labels written in the prompt's own language.
• hindi-order-json: order extraction (id, vendor, amount, currency, city) with a JSON Schema that requires all five.
• tamil-event-json: event extraction (event, venue, date) with a date format check.
• hindi-exact-answer: one exact string answer.

Baseline and candidate outputs in recorded/ are hand-written replay fixtures, labelled as such in metadata.note and in the suite description. The candidate produces the intended three regressions and one passing change:

Case	Candidate	Result
hindi-sentiment-positive	positive label flipped to negative	regression
tamil-sentiment-positive	positive label flipped to neutral	regression
hindi-order-json	currency dropped	regression (schema)
tamil-event-json	notes added	passing output change

Verification (Python 3.14, repo venv):

• modeldiff validate examples/multilingual/suite.yaml -> exit 0, 6 cases (3 classification, 1 exact, 2 json).
• modeldiff compare out/multilingual-baseline.json out/multilingual-candidate.json -> exit 1, REGRESSIONS FOUND (3), one passing-with-changed-output entry.
• pytest: 184 passed, including two new tests in tests/test_cli.py that pin the regression ids and assert both fixtures and the suite description label the outputs as hand-written.
• ruff check clean, ruff format --check clean on the changed Python file.

One thing to review carefully: the Hindi and Tamil wording was drafted with an AI assistant and I am not a native speaker of either language. The sentences are short and everyday on purpose, and the suite description says the wording should be confirmed by a native speaker. Please route it to the Hindi and Tamil speakers you mentioned before merge; I will apply any wording corrections they suggest.

This change was prepared with an AI coding assistant.

[github-actions]

Thanks for your first pull request to ModelDiff! A maintainer will review it soon. CI for first-time contributors starts after a maintainer approves it.

While you wait, please check that:

• The issue was assigned to you before you started (see One issue at a time).
• The description links the issue (Closes #12) and says how you tested the change.
• pytest, ruff check src tests examples and ruff format --check src tests examples pass locally.
• The Files changed tab shows only changes for this issue.

[AdityaWagh06]

Thanks @tnalbant, this is really well done. The fixtures and the regression mix are exactly what #28 asked for. Two small README fixes and one question:

1. The README says "two sentiment classifications", but the suite has three (two Hindi, one Tamil). Please change it to "three".
2. Please remove the last sentence about Windows ("On Windows, redirecting this report..."). The UTF-8 fix (#32) is already merged, so it's no longer true.
3. Were the Hindi and Tamil prompts checked by a native speaker? If not, that's okay. Just say so in the PR, and I'll ask for a native review before merging.

[tnalbant]

Both README fixes are in, as d00cf80:

1. It says three sentiment classifications now, and I listed them where it helps (hindi-sentiment-positive, hindi-sentiment-neutral, tamil-sentiment-positive).
2. The Windows sentence is gone — you are right that fix: handle non-ASCII reports when stdout is redirected (cp1252 -> UTF-8) #32 made it untrue.

On the third question: no, they were not checked by a native speaker. Neither the Hindi nor the Tamil prompts. I do not have one to ask, so please do line up the native review you offered — I would much rather have someone read them than have the fixtures merged on my word alone.

Nothing else changed: the suite, the fixtures and the tests are untouched by this commit.