SEA evidence selection

What it can help with

Prototype project recall or document assistance without placing every available record in a prompt.

These are potential pilot applications, not claims of customer outcomes.

What the source implements

SEA combines semantic, keyword, metadata, recency, authority, graph, contradiction and task-fit signals. Task-aware scoring adjusts selected task weights before bounded selection. The learning loop separately persists adaptive persona-router weights; these are not evidence that the eight retrieval weights were trained on a quality benchmark. Recall skips audit-rejected completions and retries within its bounded candidate ceiling, and the learning loop does not adapt router weights from an audit-rejected turn.

Available today

Eight-signal selection, task-specific scoring and bounded prompt assembly are implemented.

How to evaluate it

Use a documented corpus with relevant, irrelevant, conflicting and missing evidence. Review selected context and bounded behavior alongside answer quality; a successful selection test alone is not a hallucination-reduction result.

Supporting research and guides

Read the accompanying archive entries: Sparse Evidence Attention for Bounded Context Assembly; Backtrace Confidence Signals for Hallucination Reduction; Chunk Provenance in Long-Context Retrieval Systems. Each distinguishes implemented behavior from proposed applications and includes pinned implementation evidence.