Retrieval Dimensions

• Retrieval is not one score. It is a set of independent signals — nine numbered dimensions plus a few auxiliary tiers — that each nominate and weigh candidates, and are then combined by a single fusion step.
• Every dimension runs on deterministic CPU math. There is no model call anywhere on this page's subject matter; the opt-in encoders (dense, SPLADE) are the only neural components, and both ship off.
• The list below is the inventory. Whether a dimension actually fires at shipped defaults is the "default state" column — and several do not.
• Each default-off state is receipt-gated: a measurement decided it, and the measurement is linked. The corresponding cost figures are on Benchmarks and Receipts.

The dimensions

Auxiliary tiers that fire but are not numbered dimensions:

• Path-key index (Tier 0) — compound (path_token, key) lookup. [retrieval] pki_enabled off since 2026-08-17 (#370).
• Filename anchor — a lexical boost on filename matches, on since 2026-04-22.
• Lexical IDF anchor — an inverse-document-frequency anchor over the tag index. It issues its own probes outside the D2 tag gate, so disabling the tag tiers does not stop the tag table being read.
• Stage-2 dense recall (BGE-M3) — [retrieval] dense_embedding_enabled off since 2026-08-15.

Fusion: how the signals become one ranking

• Reciprocal Rank Fusion is the default ranker and has been since 2026-07-06. Each signal contributes a rank, not a raw score, so no hand-picked exchange rate has to hold across tiers with incompatible scales.
• The receipt: RRF measured +12pp gold-document delivery over the legacy additive accumulator on the hardest internal bed — 0.74 vs 0.62. The additive path mis-scales dense against the FTS5 BM25 cap, which is what the rank-level fusion removes.
• fusion_mode = "additive" restores the pre-Stage-3 accumulator byte-identically. It is DEPRECATED-FOR-REMOVAL on a condition gate, not a calendar — v(N+2) at the earliest.
• The per-tier weights in [retrieval] bind in both modes: under "additive" they are the tier coefficients and caps themselves; under "rrf" they are rank post-multipliers.
• Under RRF the abstain gates run ratio-only — the absolute score floors were calibrated against additive fusion and do not transfer.

Wave-1 ranking defaults (0.9.1, #407)

• rrf_k 60 → 20. A smaller k steepens the rank discount, so a candidate that any one signal ranks highly survives fusion instead of being averaged away.
• rerank_combinator_by_class extends to all five classifier classes (arithmetic, factual, procedural, multi_hop, default), each mapped to eps_band — post-fusion rerank signals break ties only inside a relative band of the leader's fused score.
• Measured together on the 829k bed at n=470: gold-document delivery 0.555 → 0.630, recall@12 0.651 → 0.681, median gold rank 3 → 2, zero question-type regressions.
• The cross-shard merge constant deliberately stays at 60 — that surface was not measured, and is annotated as unmeasured in the code rather than flipped on faith.

Cymatics, honestly

• Each term is hashed (MD5) into one of 256 bins and given a small Gaussian spread; query and candidate are then compared as the resulting 256-dimensional vectors. It is a cheap, deterministic, model-free term transform.
• "Cymatics" is a mnemonic for that binned spectrum, not a claim of signal-processing semantics. Nearby bins are hash placement, not related meaning. There is no wave physics in the scoring.
• It is a candidate-reordering signal that has not yet been isolated against hashed bag-of-words or random-bin controls. Treat it as an experimental cheap feature, not a proven one.
• What has been measured, once: the 2026-08-06 ledger's re-run of the cymatics ablation on a 141-needle 100k carve found recall@12 and recall@5 identical across all six arms — shipped, off, collapsed-spread, and three random-bin seeds. The entire measurable effect was confined to sub-window ordering, and the shipped arm was not distinguishable from random binning — it beat the best random seed by 0.0012 MRR against a random-control band of width 0.0035, and tied that seed exactly on recall@1. The verdict was KEEP-CHEAP, not KEEP-PROVEN — one bed, one needle family, so the standing "not isolated" claim above still holds for the project as a whole.
• The binned vector is exposed directly by POST /fingerprint if you want to look at it. See HTTP API.

What the retrieval-layer ledger measured

The 2026-08-06 retrieval-layer ledger is the cost-versus-impact audit of these layers: ablation ladder arms that self-validate by signal absence, a storage audit that maps GB to layers, and an adversarial council pass that re-probed the headline claims order-reversed. It is a dated document — these are its findings on its beds, not a statement of current shipped state, and several defaults have moved since.

Standing scope limits on all of the above: 141 carve needles at 100k and 30 of 469 blob needles at 829k, one config profile, one needle family (needle-in-haystack). At blob scale one needle moves recall@12 by 0.033, so several of those verdicts rest on one-needle movements and say so.

Two reading rules the ledger's own receipts forced, worth carrying into any comparison you run:

• Check delivered_count before crediting a delivered-gold gain. A small class of queries discards a rank-1 gold document with an empty delivered window; any perturbation that flips one back on looks like a retrieval win and is not one.
• Rank-based recall cannot see delivery. Signals that pull documents forward after ranking, and gates that trim the window before assembly, are both invisible to recall@k by construction — which is why the delivered basis is the metric the 0.9.x work reports.

Where to change these

• Every knob named on this page lives in cymatix.toml under [retrieval], [ingestion], [context], [cymatics], or [synonyms]. Defaults, flip dates, and env-var precedence: Configuration.
• If a query returns nothing, the usual cause is D2, not D1: the query's keywords never map to the tags assigned at ingest. Add entries to [synonyms] — and re-read the ledger row above before adding many. Symptom-first version: Troubleshooting.

Go deeper

• docs/architecture/DIMENSIONS.md — the formal dimension inventory, per-lane implementation status, and the decision gates each lane must clear
• docs/architecture/PIPELINE_LANES.md — which table each signal reads at query time
• docs/benchmarks/2026-08-06-retrieval-layer-ledger.md — the full cost/impact ledger, its addenda, and its scope limits
• docs/benchmarks/2026-08-14-encoder-isolation-scale-curve.md — the four-scale receipts behind the dense and SPLADE default-off flips
• docs/benchmarks/2026-07-06-sike-run2-fts-depth-fusion.md — the RRF-versus-additive receipt
• docs/benchmarks/BASELINES.md — the comparability ledger every bed row is pinned against
• Next: Pipeline · Configuration · Benchmarks and Receipts · Architecture Map