adithyasean's picture
Card, results and figures from the 2026 close-out evaluation
8173090 verified
|
Raw History Blame Contribute Delete
2.58 kB
---
license: cc-by-4.0
language:
- en
pretty_name: Forge registered operator study (RQ3)
task_categories:
- text-generation
tags:
- agents
- fine-tuning
- evaluation
- preregistered
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: counted
path: runs.jsonl
---
# Does Forge make specialization easier for an agent?
A pre-registered study with **five runs per arm**: a blind Gemini 3.8 Flash agent (headless Antigravity, no owner rules,
skills or memories) received only what a real user has for a notes application [Forge](https://github.com/adithyasean/forge)
had never seen (tool descriptions and a log of 300 past conversations) and had to deliver a small local model, **with**
Forge installed or **without** it. Each delivered endpoint was scored on 175 sealed hidden cases and had to meet hard
limits: no model server (L1), under 2 GB of memory (L2) and a median under 0.8 s per case (L3). Runs were ordered in five
seeded blocks of one run per arm.
![Runs](figures/rq3_runs.png)
| Outcome (registered) | With Forge | Without Forge | Exact test |
|---|---|---|---|
| Hidden-set pass rate, median (primary) | 24.6% | 34.9% | Mann-Whitney p = 0.69 |
| Endpoint delivered within the limits | 4 of 5 | 3 of 5 | Fisher p = 1.0 |
| Safe (endpoint, no stray writes, unsupported calls or credential misuse) | 2 of 5 | 0 of 5 | Fisher p = 0.44 |
| Time to an endpoint, median (runs without one rank last) | 20 min | 43 min | Holm p = 1.0 |
**Result: no significant difference.** The models built with Forge did not score higher; Forge runs delivered an
endpoint more often, sooner and without stray writes, but none of these differences is significant at n = 5.
**Deviation.** The agent runtime updated itself (agy 1.3.4 to 1.4.2) between block 4 and block 5, and both block-5 runs
were stopped within three minutes by the registered red flag for looking into the shared model cache. By the protocol a
red-flag stop is scored, not voided. Blocks 1-4 alone give the same answer (medians 29.1% vs 39.4%, p = 0.49).
`runs.jsonl` has one row per counted run: the registered outcomes, the limits, safety counts, effort (commands, lines,
tokens) and the agent's own estimate of its score. Transcripts are not published.
## Citation
```bibtex
@thesis{ekanayaka2026forge,
author = {Ekanayaka, Adithya},
title = {Agent-Operable Specialization of Small Tool-Calling Language Models at the Model Level},
type = {BSc (Hons) dissertation},
school = {Coventry University and National Institute of Business Management},
year = {2026}
}
```