Scope
The initial incident application study contains 48 synthetic incidents in 12 organization clusters: three clear and one ambiguous incident per organization. Each incident appears under full flat context, grouped context, and an indexed target representation. All 144 requests succeeded.
A separate agent reviewed the labels without seeing model outcomes. This is not human annotation. The 36 clear narratives are the same across the three conditions; repeated representations do not create new semantic examples.
Observations
| Condition | Correct clear semantic labels | Correct owner lookup | Mean input tokens |
|---|---|---|---|
| Full flat context | 36 / 36 | 18 / 48 | 5,651.58 |
| Grouped context | 36 / 36 | 48 / 48 | 5,744.50 |
| Indexed target | 36 / 36 | 48 / 48 | 740.35 |
Indexed retrieval reduces mean input use by 86.9% relative to full flat context. The target-selection code receives the supplied ticket identifier and structured records, resolves ownership, and preserves the incident text. It does not receive the semantic labels.
Interpretation
The study demonstrates a concrete context-cost reduction with unchanged observed clear-case labels. It does not demonstrate a semantic accuracy improvement: the semantic task is already at ceiling under every condition. It also does not establish population equivalence or noninferiority.
Ownership is supplied explicitly in the indexed condition, so that result is a reading control. A production system with exact ownership available would ordinarily use it directly rather than ask a model to recover the computed value.
The primary statistical unit is the organization. Both primary semantic contrasts are zero, and the empirical bootstrap intervals collapse because all observed clear labels are correct. Such intervals do not prove population perfection. Ambiguous cases use permissive allowed-label sets and do not establish a unique semantic truth.
The historical source is round3/representation/application-pilot-report.md. The evidence download includes the frozen v2 plan, responses, protocol, and analysis under data/application/. It reproduces the non-latency endpoints and includes an independent count audit. This supporting study is outside the main three-study inventory.