# Experiments and results

Read the question, comparison, findings, and limits before applying a result.

Author: Murali Sid (https://linkedin.com/in/muralisid)
Source: https://www.agenticarchitectureskills.com/research/experiments
Published: 2026-09-17
Updated: 2026-10-15
Licence: CC BY-SA 4.0, https://creativecommons.org/licenses/by-sa/4.0/
Cite as: Murali Sid. Experiments and results. Agentic Architecture Skills. https://www.agenticarchitectureskills.com/research/experiments

---


These six reports describe the author's own experiments. Each one answers one question. Read a result with the workload and comparison in its original report.

## Retrieval and context [#retrieval-and-context]

<Cards>
  <Card title="When one representation loses detail" href="/patterns/aspect-dilution">
    How document complexity affected retrieval in the reported tests.
  </Card>

  <Card title="What happened on real prose" href="/patterns/real-prose">
    Where matched chunks were competitive or better.
  </Card>

  <Card title="Conditioning on an objective" href="/patterns/why-not-to-condition-on-the-objective">
    A proposed improvement that did not survive the comparisons.
  </Card>

  <Card title="Selection and economics" href="/patterns/gate-and-economics">
    Why an additional selection step must prove both quality and cost.
  </Card>

  <Card title="Diversity and the consuming model" href="/patterns/diversity-and-consumer">
    The retrieved set and the model using it both affect the outcome.
  </Card>

  <Card title="Memory measurements" href="/memory/what-we-have-measured">
    The existing measured system, mechanisms, and limitations.
  </Card>
</Cards>

## Use a result carefully [#use-a-result-carefully]

Look for the dataset, questions, baseline, evaluation measure, model configuration, and uncertainty. Do not infer an industrial deployment result from a text-retrieval benchmark. New modality and agent-loop experiments need their own evidence.

See [how to test a context design](/patterns/how-to-test-a-context-design), [what changed](/research/what-changed), and [proposed investigations](/research/questions).