Submodular Evidence Packing (token-budgeted, diverse evidence)

File: src/search/packing.rs

The problem

When an agent’s context window is finite, recall must choose which of many candidate chunks to surface. Naive top-k over a single score over-selects the same story and wastes tokens on near-duplicates. You want a set of evidence that is jointly relevant, novel, and representative under a hard token budget.

The reference

arXiv:2607.00725 — budgeted monotone submodular maximization with lazy greedy, achieving the classic (1 − 1/e) optimality bound, shown to gain +5.1 F1 on HotpotQA. The objective rewards coverage and penalizes redundancy; a diversity gate keeps the set from collapsing onto one cluster.

The implementation

src/search/packing.rs::pack is a deterministic lazy-greedy under a knapsack:

• DEFAULT_MAX_CONTEXT_TOKENS = 160, MAX_CANDIDATES = 64 cap the work.
• Objective = relevance + coverage + representativeness (the Weights config, tunable via env), gated by an MMR-style diversity bound: DEDUP_SIMILARITY = 0.85 — a candidate whose best overlap to an already- chosen chunk exceeds 0.85 is dropped.
• est_tokens(text) estimates tokens at CHARS_PER_TOKEN = 4 — a cheap, deterministic proxy (no tokenizer in the hot path).
• /recall?max_context_tokens= triggers packing; the response reports packed_tokens and (with a gold_answer) the answer_in_context diagnostic — is the answer actually inside the chosen evidence?

Measured ceiling

• Diversity is lexical Jaccard, not embedding cosine (a cheap, deterministic proxy; cosine would pull the model into the packer).
• The weights are corpus-independent defaults; weights_from_env() lets an operator calibrate without a rebuild.
• Greedy is near-optimal, not optimal — the honest (1 − 1/e) claim is stated plainly, not exceeded.

The answer_in_context diagnostic is the bridge to a judged-corpus recall floor (brain eval).