Questions to investigate

Proposed work that could improve the guide, with an explicit comparison and decision to inform.

The questions below are proposals. They are not described as running experiments or completed findings.

Do specialized representations improve industrial investigations?

Compare exact records, keyword search, ordinary text chunks, and representations designed for media or sequences. Use realistic questions with accepted evidence. Measure useful retrieval, false matches, review effort, and full cost. The decision is whether the added representation earns a place in the knowledge design.

When does an agent loop outperform a fixed workflow?

Use cases with both routine and ambiguous evidence. Compare the current process, a fixed tool sequence, and a bounded adaptive loop. Measure completed outcomes, unnecessary tool calls, cost, and recovery. The decision is where adaptive control adds value.

What temporal information must remain explicit?

Test questions about order, duration, validity, and exact dates alongside questions about seasonal or operational patterns. Compare learned representations with explicit timestamps and source records. The decision is what metadata and query operations are required.

Does evidence collection reduce total audit effort?

Measure collection time and the reviewer's work together. Include invalid, stale, conflicting, and wrong-scope evidence. The decision is whether the agent reduces the full effort of reaching a defensible conclusion.

How does human supervision change the economics?

Measure the distribution of exceptions and the effort required to resolve them, including peak demand. Compare the cost per resolved outcome and service delay. The decision is whether a larger scope is operationally supportable.

Each investigation needs an agreed dataset, rights to use it, a baseline, and a publishable result before it can appear under experiments.