README.md

---
language: en
license: apache-2.0
base_model: Qwen/Qwen3-14B
tags:
  - tool-calling
  - routing
  - code-generation
  - typescript
  - healthcare
  - aac
  - qwen3
  - gguf
---

# prism-coder:14b — Prism Memory Tool Router + Healthcare TypeScript Coder

Fine-tuned Qwen3-14B for the [Prism AAC](https://github.com/dcostenco/prism-aac) / Synalux healthcare platform.

## Current Production Model: S14 (eval_300 — 17-tool routing)

**299/300 = 99.7% strict** on eval_300 — 300 cases, 17 Prism Memory tools

Single remaining failure: `"Save."` — genuinely ambiguous between `session_save_ledger` and `session_save_experience`. All other categories at 100%.

| Category | Accuracy |
|----------|:--------:|
| session_save_ledger (ledger logging) | 100%* |
| session_load_context (context loading) | 100% |
| session_search_memory (memory recall) | 100% |
| session_save_handoff (agent handoff) | 100% |
| session_forget_memory | 100% |
| session_health_check | 100% |
| session_compact_ledger | 100% |
| session_export_memory | 100% |
| session_task_route | 100% |
| session_save_experience | 100%* |
| session_synthesize_edges | 100% |
| session_backfill_links | 100% |
| knowledge_search | 100% |
| knowledge_forget / upvote / downvote / set_retention | 100% |
| abstain (general questions, greetings, CS concepts) | 100% |
| multi-intent (compound tool calls) | 100% |
| natural phrasing | 100% |

\* One edge case (`"Save."`) scores as a failure on one tool; both are correct interpretations.

### eval_300 Details — S14
- **Base**: Qwen3-14B → surgical LoRA chain (S1→S14)
- **Eval**: 300 cases, strict scoring (exact tool match), 17 Prism Memory tools + abstain + multi-intent
- **Training**: MLX LoRA, rank=8, scale=20.0, 16 layers, 100 iters, LR=5e-6, mask_prompt=true
- **Corpus**: S14 — balanced natural-phrasing + tool-use SFT (100 train / 20 valid)
- **SYSTEM_PROMPT**: Synalux identity + 17 Prism Memory tools + 13 multimodal tool modules + `<tool_call>` JSON block format

### Tools (S14 routing model)
All 17 Prism Memory tools:
`session_save_ledger`, `session_load_context`, `session_search_memory`, `session_save_handoff`,
`session_forget_memory`, `session_health_check`, `session_compact_ledger`, `session_export_memory`,
`session_task_route`, `session_save_experience`, `session_synthesize_edges`, `session_backfill_links`,
`knowledge_search`, `knowledge_forget`, `knowledge_upvote`, `knowledge_downvote`, `knowledge_set_retention`

---

## Legacy: Coding Eval — v42

**22/22 (100%)** on the Synalux healthcare TypeScript eval.

Task: write a production Next.js API route for X12 835 ERA reconciliation against existing 837P claims.

<details>
<summary>22-check eval breakdown (click to expand)</summary>

| Check | Pass |
|-------|------|
| withAudit wrapper | ✓ |
| authenticateRequest | ✓ |
| supabaseAdmin (not client) | ✓ |
| cross-tenant guard (workspace_members + BILLING_ROLES) | ✓ |
| UUID_RX validation | ✓ |
| decryptPhi before PHI access | ✓ |
| HIPAA audit (hipaa_access_log) | ✓ |
| HIPAA non-blocking (.then) | ✓ |
| 409 already-reconciled guard | ✓ |
| 422 no CLP segments | ✓ |
| parse CLP segment | ✓ |
| parse SVC segment | ✓ |
| parse CAS CO (contractual) adjustment | ✓ |
| parse CAS PR (patient responsibility) | ✓ |
| GL cash_received entry | ✓ |
| GL contractual_adjustment entry | ✓ |
| GL patient_ar entry | ✓ |
| claim status map (1=paid) | ✓ |
| claim status map (4=denied) | ✓ |
| no postgres detail in 500 | ✓ |
| belt-and-suspenders workspace_id eq on update | ✓ |
| marks ERA file reconciled | ✓ |

</details>

---

## Legacy: BFCL Routing Benchmark — v36

**Mean: 100.0% PERFECT** (3-seed average, seeds 2027/2028/2029, 102 cases each) — 6-tool routing

---

## GGUF Files

| File | Use | Size |
|------|-----|------|
| `qwen3-14b-s14-q4km.gguf` | **Routing** — production Prism Memory (17 tools, 99.7%) | ~9 GB |
| `qwen3-14b-v42-q4km.gguf` | **Coding** — Synalux TypeScript (22/22, 100%) | ~9 GB |
| `prism-coder-14b-v36-q4km.gguf` | Routing legacy (6-tool BFCL, 100%) | ~9 GB |

## Version History

| Version | Eval | Type | Notes |
|---------|------|------|-------|
| **S14** | **299/300 = 99.7% (eval_300)** | **Router** | **Production — 17-tool Prism Memory routing** |
| v42 | 22/22 coding (100%) | Coder | Claim status patch; Synalux TypeScript |
| v36 | 100% BFCL (6-tool routing) | Router | Legacy 6-tool routing |
| v34 | 98.0% BFCL | Router | — |

## Usage

```bash
# Pull production routing model (S14 — 17-tool Prism Memory)
ollama pull dcostenco/prism-coder:14b

# Or pull GGUF directly from this repo and use with Ollama:
# FROM qwen3-14b-s14-q4km.gguf
# PARAMETER temperature 0
# PARAMETER num_ctx 8192
```

### System Prompt (S14)

```
You are Synalux, a memory-augmented coding and clinical reasoning assistant. You have access to 
Prism Memory tools (session_save_ledger, session_load_context, session_search_memory, 
session_save_handoff, session_forget_memory, session_health_check, session_compact_ledger, 
session_export_memory, session_task_route, session_save_experience, session_synthesize_edges, 
session_backfill_links, knowledge_search, knowledge_forget, knowledge_upvote, knowledge_downvote, 
knowledge_set_retention) and 13 multimodal tool modules (image_gen, office, web_scraper, browser, 
tts, ocr, git, terminal, deps_scanner, hipaa, data_graph, templates, pdf_parser). Think 
step-by-step before answering. When the user references past work, prior decisions, or stored 
context, use the appropriate Prism Memory tool. Format tool calls inside <tool_call>...</tool_call> 
JSON blocks with fields 'name' and 'arguments'. If no tool is needed, answer directly in plain 
text. ABSTAIN for general programming questions, CS concepts, greetings, and capability questions.
```

## Cascade

| Tier | Model | Role |
|------|-------|------|
| 1.7B | `dcostenco/prism-coder:1b7` | Fast verify / edge cases |
| 4B | `dcostenco/prism-coder:4b` | Mid-tier verify |
| **14B** | **`dcostenco/prism-coder:14b`** | **Production routing** |
| 32B | `dcostenco/prism-coder:32b` | Top-tier / complex reasoning |