🔧 Solution C: Change to more deterministic model

[GMNAPI]

Context

Alternative/complementary solution if GPT-4-turbo-preview proves too non-deterministic (issue #16).

Problem

gpt-4-turbo-preview is a preview/rolling model that may:

• Change behavior over time
• Have inconsistent inference across regions/servers
• Be less deterministic than stable versions

Solution

Switch to more deterministic and stable model.

Options

1. GPT-4o (Recommended)

• Model ID: gpt-4o
• Latest stable omni model
• Better instruction following
• More deterministic than turbo-preview
• Similar performance, lower cost

2. GPT-4 Stable

• Model ID: gpt-4-0613
• Frozen snapshot (no updates)
• Maximum determinism
• Slightly slower than turbo

Implementation

Update environment variables:

# .env (local)
CHAT_MODEL=gpt-4o

# Railway (production)
# Settings → Variables → Edit CHAT_MODEL
CHAT_MODEL=gpt-4o

Verify in /debug:

curl https://llarjove-production-754a.up.railway.app/debug | jq '.config.chatModel'
# Should return: "gpt-4o"

Comparison

Model	Determinism	Performance	Cost	Notes
gpt-4-turbo-preview	⚠️ Low	⭐⭐⭐⭐⭐	$$	Rolling updates
gpt-4o	✅ High	⭐⭐⭐⭐⭐	$	Recommended
gpt-4-0613	✅ Very High	⭐⭐⭐⭐	$$$	Frozen, expensive

Acceptance Criteria

• Model changed in both local and Railway
• /debug confirms new model in use
• Tested locally with consistent responses
• Deployed to production
• Production behavior matches local

When to Use

Use this solution if:

1. Solution B (few-shot) doesn't fully resolve inconsistency
2. Diagnostic shows embeddings are identical
3. Need maximum determinism for production reliability

Related

• Issue 🔍 Prod vs Local Inconsistency - Diagnosis #16: Main investigation
• Issue 🔧 Solution B: Improve prompt robustness with few-shot examples #19: Solution B (prompt improvement)
• PR feat(api): add debug endpoints for prod/local diagnosis #17: Debug endpoints