ADR-192: Late-interaction matching: the flow

A visual companion to ADR-160 (chunked late-interaction matching with softmax-share gating). It diagrams two things this session settled by watching the matcher run: the per-surface evidence pipeline that decides a semantic fire, and the multi-channel fire model that explains why most fires you see are not semantic.

The evidence pipeline (one surface → fire / no-fire)

The pipeline is lenient where it ranks (a way is admitted on its strongest single-chunk evidence) and strict where it confirms (the winning way's own body must corroborate the chunk it won). Everything downstream of Chunk is the late-interaction path; a surface too sparse to chunk, or a session with no embedder, takes the single-vector fail-safe instead.

flowchart TD
    S(["Surface<br>prompt · task · command lookbehind"])
    KW{"Keyword lane<br>pattern: regex match?"}
    FIREKW["FIRE — keyword channel<br>(fires before semantic)"]
    CHUNKABLE{"≥ 2 chunks<br>and embedder present?"}
    FALLBACK["Single-vector fail-safe<br>ADR-156 calibrated gate"]
    CTX["Contextualize<br>enrich sparse surface with intent"]
    CHUNK["Chunk + embed<br>each sub-unit → vector"]
    MATCH["Late interaction<br>each chunk vs every way alias"]
    PEAK["Rank by peak<br>score = max per-chunk similarity"]
    ADMIT{"Admission co-gate<br>share ≥ 0.15  OR  peak ≥ 0.50"}
    CONFIRM{"Body-confirm<br>best winning-way body chunk<br>vs the won chunk ≥ 0.35"}
    FIRE(["FIRE — semantic channel<br>carries score · surface · won chunk"])
    NOPE(["no fire"])

    S --> KW
    KW -- yes --> FIREKW
    KW -- no --> CHUNKABLE
    CHUNKABLE -- no --> FALLBACK
    FALLBACK --> NOPE
    CHUNKABLE -- yes --> CTX
    CTX --> CHUNK --> MATCH --> PEAK --> ADMIT
    ADMIT -- "neither" --> NOPE
    ADMIT -- "admitted" --> CONFIRM
    CONFIRM -- "< 0.35" --> NOPE
    CONFIRM -- "≥ 0.35" --> FIRE

    classDef surface fill:#475569,color:#ffffff,stroke:#94a3b8;
    classDef proc fill:#2d7d9a,color:#ffffff,stroke:#4a5568;
    classDef gate fill:#fbbf24,color:#1a1a1a,stroke:#d97706;
    classDef fire fill:#2d8e5e,color:#ffffff,stroke:#4a5568;
    classDef fallback fill:#f6821f,color:#1a1a1a,stroke:#d97706;
    classDef dead fill:#475569,color:#ffffff,stroke:#94a3b8;

    class S surface;
    class CTX,CHUNK,MATCH,PEAK proc;
    class KW,CHUNKABLE,ADMIT,CONFIRM gate;
    class FIRE,FIREKW fire;
    class FALLBACK fallback;
    class NOPE dead;

Why the co-gate. share = Σ softmax-mass / n_chunks caps a way that owns one of N topics at ≈1/N, so on a topic-diverse prompt a specific single-chunk match dilutes below any fixed share gate. Admitting on a decisive peak ≥ 0.50 recovers that match; the strict body-confirm ≥ 0.35 then carries precision so the looser admission does not leak noise. (SOFTMAX_TAU = 0.08, TOP_K = 8.)

The multi-channel fire model

A way can fire through several channels, and the channel decides how to judge the fire — not one rubric fits all. Keyword fires before semantic; a session-start trigger fires on the session lifecycle, not on intent at all.

flowchart LR
    subgraph intent["judge on intent-relevance"]
        SEM["semantic<br>late-interaction alias match"]
    end
    subgraph pattern["judge on pattern-appropriateness"]
        KWD["keyword<br>pattern: on prompt"]
        BSH["bash<br>command lookbehind"]
        FIL["file<br>glob trigger"]
    end
    subgraph lifecycle["judge on condition-correctness"]
        STA["state<br>session-start / signal"]
    end
    subgraph ride["judge on parent correctness"]
        CHK["check-pull<br>rides its parent way"]
    end

    classDef sem fill:#7c3aed,color:#ffffff,stroke:#8b5cf6;
    classDef pat fill:#2d7d9a,color:#ffffff,stroke:#4a5568;
    classDef life fill:#fbbf24,color:#1a1a1a,stroke:#d97706;
    classDef rid fill:#475569,color:#ffffff,stroke:#94a3b8;

    class SEM sem;
    class KWD,BSH,FIL pat;
    class STA life;
    class CHK rid;

The trap a naive relevance classifier falls into: judging a state fire (meta/localize, softwaredev/freshness at session start) by task-relevance. Those fire every session start to run a gated check; the right question is "did the gated macro surface — or correctly stay silent?", not "is this relevant to what we're doing." An automated fire-evaluator has to route each fire to its channel's rubric before scoring it.

Reading it live

• ways introspect fires --session <id> — semantic fires only, score · surface · way, lowest-score first. The suspect tail is --max-score 0.55.
• ways introspect dump --session <id> — every fire with its trigger_channel, the input to the channel-aware evaluation above.
• ways match "<multi-sentence query>" — authoring diagnostic: peak / share / confirm / outcome / won-chunk per candidate, without writing to the events log.