Agentic Drifter

I ran SWE‑bench on a small LLM to map failure modes and understand how tiny models behave under full task‑grounded pressure. This experiment tested whether a small model could sustain a multi‑stage evaluator pipeline under frontier‑level task conditions — and what breaks first when it can’t.

𝐓𝐡𝐞 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞 𝐟𝐨𝐥𝐥𝐨𝐰𝐞𝐝 𝐚 𝐬𝐭𝐚𝐧𝐝𝐚𝐫𝐝 𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞: 𝐫𝐞𝐚𝐬𝐨𝐧𝐢𝐧𝐠 → 𝐜𝐫𝐢𝐭𝐢𝐪𝐮𝐞 → 𝐩𝐚𝐭𝐜𝐡 → 𝐜𝐨𝐦𝐩𝐚𝐫𝐢𝐬𝐨𝐧. The objective was not correctness but signal: determining whether the model could produce meaningful evaluator grade behavior and reveal its own capability limits.

𝐒𝐮𝐜𝐜𝐞𝐬𝐬𝐞𝐬
•The pipeline architecture executed end‑to‑end.
•The model produced coherent reasoning about the SQLFluff quiet‑mode issue.
•The critique mechanism worked; several critiques were semantically valid even when mixed with drift.
•The prompt scaffolding held long enough for the model to respond in the requested structure.
•The model demonstrated an ability to reason SWE‑bench‑style tasks despite its size.

𝐋𝐢𝐦𝐢𝐭𝐚𝐭𝐢𝐨𝐧𝐬
•Small models cannot sustain long‑context, multi‑step reasoning loops. They enter a frontier envelope, begin to fail in predictable ways.
•Token‑budget tradeoffs were clear: shorter prompts improved stability but reduced semantic richness; longer prompts increased drift and boundary loss.
•Strict output formats caused placeholder fallbacks; looser formats caused instruction boundary loss.
•The environment introduced variability in GPU availability and model initialization, which affected stability.

𝐅𝐚𝐢𝐥𝐮𝐫𝐞𝐬
•Patch generation was unreliable. The model could not consistently produce unified diffs or structured patch output.
•Critique stability degraded as the prompt length increased. Critiques were mixed with prompt echoing, repetition, and partial collapse.
•Instruction boundaries were repeatedly lost, causing the model to reprint the prompt or earlier reasoning segments.
•Placeholder fallbacks appeared when the model exceeded its reasoning envelope.

𝐖𝐡𝐚𝐭 𝐭𝐡𝐞 𝐄𝐱𝐩𝐞𝐫𝐢𝐦𝐞𝐧𝐭 𝐕𝐚𝐥𝐢𝐝𝐚𝐭𝐞𝐝
•The evaluator pipeline design is sound.
•The prompt boundaries are functional even under stress.
•The critique mechanism is viable and produces meaningful signal.
•The model can reason about real tasks but cannot sustain multi‑step evaluator behavior at this scale.
•The failure modes are model capacity related, not pipeline related.

𝐒𝐮𝐦𝐦𝐚𝐫𝐲
As a self‑taught LLM Reliability Researcher, this experiment shows you can build and validate a multi-stage evaluator pipeline without SWE‑bench tooling or a software engineering background.

The pipeline produced real reasoning, multi‑pass critiques, and clear failure traces. Collapse patterns matched known small model frontier task limits. At this stage, the pipeline is validated and the model – not the method – is the limiting factor.