research-document
Domain Leakage Report
Domain Leakage Report
| Packet | Provider | Introduced Terms | Classification | Notes |
|---|---|---|---|---|
| ECR-000003-EXP003-P001A | gpt | diagnosis; evidence | likely_prior_knowledge_leakage | low; recognized artifact recorded only, with no additional procedural structure inferred from prior knowledge. |
| ECR-000003-EXP003-P001A | claude | diagnosis; hypothesis testing; evidence | likely_prior_knowledge_leakage | moderate — the recognized shape is highly conventional, which creates pressure to import a canonical elimination/confirmation structure and a candidate-exhaustion stopping rule. This was resisted: the loop boundary was left underdetermined rather than defaulted to the canonical 'regenerate candidates each pass' reading, and no acceptance threshold or elimination semantics were asserted. Residual risk is concentrated in the primitive_sequence, where the sequential ordering of enumerate/collect/evaluate is partly a convention of the recognized shape rather than an explicit packet claim. |
| ECR-000003-EXP003-P001A | gemini | diagnosis | likely_prior_knowledge_leakage | Risk of assuming specific diagnostic frameworks (like Bayesian inference, decision trees, or hardware line-replaceable unit isolation procedures) which are not explicitly detailed in the text. |
| ECR-000003-EXP003-P001B | gpt | none | generic_procedural_language | low; extraction is limited to generic identification without adding unsupported structure. |
| ECR-000003-EXP003-P001B | claude | evidence | possible_structural_inference | low-to-moderate. Risk concentrates at exactly the points where the packet is silent — the loop-back target, the confidence update rule, and the relation between sufficiency and acceptance. Familiar instances of this pattern supply confident defaults for all three (e.g., 'loop back to information gathering; confidence is a posterior updated by evidence; sufficiency is acceptance'). Those defaults were deliberately withheld and recorded as ambiguities instead. Residual risk: the sequential reading of the six lines, and the assumption that assessment consumes the gathered information, are both mildly pattern-driven, though both are also reasonable implications of the packet's ordering. |
| ECR-000003-EXP003-P001B | gemini | none | generic_procedural_language | Risk of imposing specific algorithmic steps (such as multi-criteria decision analysis, Dijkstra's algorithm, or heuristic searches) on what is presented as a highly abstracted procedural template. |
| ECR-000003-EXP003-P001C | gpt | none | generic_procedural_language | low |
| ECR-000003-EXP003-P001C | claude | evidence | possible_structural_inference | moderate — the main risk is over-specifying the loop-back target to the familiar 'regenerate candidates' pattern, and asserting that assessment consumes the gathered inputs. Both were explicitly held open as ambiguities rather than resolved. |
| ECR-000003-EXP003-P001C | gemini | none | generic_procedural_language | Low. The extracted structure is maintained strictly within the bounds of the provided 6 steps without borrowing external governance framework details. |
Leakage classifications are observational and may still reflect structural inference rather than memorized framework recall.