# AEON Bench — ThinkingCap-Qwen3.6-27B-TQ3_4S (75 cases @6144)

**Date:** 2026-07-09 20:09 EDT
**Framework:** AEON-Bench-Pod v1 (self-reported mode)
**Hardware:** NVIDIA RTX 5060 Ti 16 GB (turbo-tan/llama.cpp-tq3 77dd77473, TQ3_4S)
**Budget:** `--max-tokens 6144` blanket
**Run shape:** `--limit 75 --fast` (no agentic harness)

## Results

**Overall mean: 0.480** across 75 scored / 75 cases

| Category | Mean | N |
|---|---|---|
| math | 0.400 | 30 |
| instruction | 0.300 | 30 |
| reasoning | 1.000 | 15 |

| Difficulty | Mean | N |
|---|---|---|
| easy | 1.000 | 6 |
| medium | 0.778 | 9 |
| hard | 0.733 | 15 |
| expert | 0.429 | 21 |
| frontier | 0.125 | 24 |

## Comparison vs sibling Tess-4-27B-TQ3_4S (same harness)

| Model | math | instruction | reasoning | mean |
|---|---|---|---|---|
| thinkingcap-27b-tq3-4s.gguf (this) | 0.400 | 0.300 | 1.000 | 0.480 |
| tess-4-27b-tq3-4s.gguf | 0.500 | 0.333 | 0.933 | 0.560 |

**ThinkingCap vs Tess profile (both 27B-TQ3_4S on node-d):**
- Both struggle on instruction (ThinkingCap 0.300, Tess 0.333)
- Both dominate reasoning — ThinkingCap scores a perfect 1.000, Tess 0.933
- Math is the differentiator: Tess wins at 0.500 vs ThinkingCap 0.400

## Notes

- 30 prose cases (not in this 75-case run) require a frontier judge endpoint — excluded.
- Bench run in self-reported mode. Mothership returned HTTP 403 (expected).