ADR-127: Full-body embedding corpus for way matching

Context

Way matching (ADR-108) embeds a keyword-curated string per way — the description field plus the vocabulary list — rather than the way's body prose. When specific parent/child promotions failed (e.g. environment/deps dominating supplychain/depscan/node), the instinct was that a denser text signal would sharpen discrimination. That instinct carried an implicit premise: routing quality is limited by the text-matching dimension and can be improved by making the embedded representation smarter.

This ADR tests that premise by productizing full-body embedding in two shapes, and documents what the test revealed about where routing quality actually lives.

Considered

Two variants against the existing description + vocabulary baseline, evaluated on a 90-way global corpus with 18 prompts (16 actionable across three firing bands, 2 negative):

• full-truncated — embed the first ~256 tokens of each way's body
• full-chunked — embed overlapping chunks of the full body, max-pool to a single vector

Results

Variant	Rebuild	Cost ×	Clear 8	Cofire 5	Bound 3	Top-1
baseline	569 ms	1.00×	6/8	4/5	1/3	11/16
full-truncated	3918 ms	6.89×	7/8	3/5	1/3	11/16
full-chunked	5252 ms	9.23×	6/8	3/5	2/3	11/16

All three variants tie at 11/16 top-1 hits. Full-body variants fix some parent-promotion failures and introduce new ones at the same rate:

• P1 audit npm dependencies for known vulnerabilities — baseline promotes parent environment/deps over expected supplychain/depscan/node. Full-body variants correctly return the child.
• P7 decide between two approaches for the user schema — baseline promotes meta/knowledge over expected architecture/design. Full-body variants correctly return design.
• P9 catch me up on what happened in my inbox overnight — baseline correctly returns ea/briefing at 0.597; both full-body variants are pulled to ea/email at 0.297–0.341.
• P10 find a free 30-minute slot on my calendar — all three correct, but baseline confidence 0.671 vs full-body 0.414/0.420.

Chunked vs truncated is a wash on recall. Chunking costs 34% more than truncation, confirming the content swap is doing any work, not the chunking+pooling mechanism.

The softwaredev-only pilot's apparent win was sample homogeneity: technical prose is dense enough that full-body's broader signal dominates keyword curation. Once the corpus includes conversational (ea/), operational (itops/), and reflective (meta/) trees, full-body wins and losses cancel.

Decision

Rejected. Do not productize full-body embedding — neither truncated nor chunked.

The premise the experiment was testing — that text-matching density is the axis of improvement — is not supported by the data. Net-zero aggregate recall across a 7–9× rebuild-time penalty is the surface reading. The deeper reading is structural.

Siblings in the authored graph share vocabulary by design — they are authored into the same subgraph. No text-only representation, keyword-curated or full-body, can reliably disambiguate them. P9 is the cleanest demonstration: ea/briefing and ea/email are graph neighbors with overlapping surface language, and baseline only picks correctly through keyword luck. Full-body's denser signal reveals the underlying ambiguity rather than resolving it, which is why confidence regresses on correct hits (P10: 0.671 → 0.414).

The text-matching axis is exhausted.

Forward path

The real value in way routing is the authored graph itself (ADR-125): parent/child structure, progressive disclosure (ADR-105), firing history, and neighborhood context. Embeddings are one coordinate on each node, not the routing algorithm.

Tactical (bridging): - Targeted vocabulary adjustments when a specific sibling collision becomes painful. - Per-way embed_threshold tuning under ADR-125.

Strategic: graph-aware routing. Subsequent ADRs should explore how disclosure state, parent context, and recency break ties that text matching cannot. This ADR's contribution to that direction is negative evidence — the text-matching axis has been tested and yielded.

Provenance

Full results, per-prompt top-5 for all 18 prompts, and timing distribution were produced by the harness at experiments/chunked-embeddings/ on branch experiment/full-body-embeddings. Both were discarded with this ADR; the evidence above is the preserved record.