Bug: Resume evaluation crashes with ValidationError when LLM returns negative deductions total

[ManikanteshwaraReddy]

When using smaller LLMs (e.g., qwen3.5:4b) for resume evaluation, the model sometimes returns deductions.total as a negative value (e.g., -4 instead of 4). While semantically correct (deductions are subtractive), this violates the ge=0 constraint on the Deductions Pydantic model, causing a ValidationError that crashes the entire evaluation pipeline.

The Deductions model defines total with ge=0 and describes it as "stored as positive, applied as negative", but smaller LLMs don't reliably follow this convention.

Expected Behavior

The evaluation should complete successfully regardless of whether the LLM returns a positive or negative deductions value. A deduction of -4 and 4 should be treated as equivalent (4 points deducted).

---

Environment

Component	Version / Details
OS	Windows 11
Python	3.11.3
hiring-agent	Commit 4db8655 (Merge PR #200)
pydantic	2.11.7
ollama (Python)	0.5.1
Ollama server	Latest (with qwen3.5:4b model)
LLM model	qwen3.5:4b via Ollama

---

Steps to Reproduce

1. Configure .env to use Ollama with a smaller model:

   LLM_PROVIDER=ollama
   DEFAULT_MODEL=qwen3.5:4b

2. Ensure Ollama is running with the qwen3.5:4b model pulled:

   ollama pull qwen3.5:4b

3. Run the scoring script on any resume PDF:

   python score.py <path_to_resume.pdf>

4. The evaluation completes the LLM calls successfully but crashes during Pydantic validation of the response.

---

Relevant Logs / Stack Trace

2026-06-29 18:53:45,579 - evaluator -    90 -  evaluate_resume - ERROR - Error evaluating resume: 1 validation error for EvaluationData
deductions.total
  Input should be greater than or equal to 0 [type=greater_than_equal, input_value=-4, input_type=int]
    For further information visit https://errors.pydantic.dev/2.11/v/greater_than_equal

Traceback (most recent call last):
  File "C:\Projects\hiring-agent\score.py", line 377, in <module>
    main(pdf_path)
  File "C:\Projects\hiring-agent\score.py", line 326, in main
    score = _evaluate_resume(resume_data, github_data)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Projects\hiring-agent\score.py", line 184, in _evaluate_resume
    evaluation_result = evaluator.evaluate_resume(resume_text)
                        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Projects\hiring-agent\evaluator.py", line 85, in evaluate_resume
    evaluation_data = EvaluationData(**evaluation_dict)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
pydantic_core._pydantic_core.ValidationError: 1 validation error for EvaluationData
deductions.total
  Input should be greater than or equal to 0 [type=greater_than_equal, input_value=-4, input_type=int]
    For further information visit https://errors.pydantic.dev/2.11/v/greater_than_equal

The LLM returned valid JSON with "deductions": {"total": -4, "reasons": "..."}, but Pydantic rejected the negative value.

---

Root Cause

In evaluator.py, the raw LLM JSON response is parsed and passed directly to EvaluationData() without sanitizing the deductions value:

evaluation_dict = json.loads(response_text)
evaluation_data = EvaluationData(**evaluation_dict)  # crashes if total < 0

[ManikanteshwaraReddy]

I am Working on it