# The ablation record — what each deleted stage measured

This is the evidence behind *A Forward Pass on the Graph*, collected in one file so the claim can be
checked without reading the whole engine. Labels follow the piece: **[MEASURED]** is a number the
engine's own audit produced, **[FACT]** is a file or command you can inspect.

The ablation question was always the same: **remove this stage, hold the store and the query set
fixed, and see whether hit@10 moves.** A stage that could not move the number went.

## The baselines

| Configuration | known-entry hit@10 | Note |
|---|---|---|
| Live weighted walk (the Epoch-1 configuration) | **0.750** | every additive term, the cross-node multiplier and the synthesizer re-rank switched on |
| Raw `querySim` ordering, no bonuses | **0.995–1.000** | all additive terms zeroed; the walk's own stages removed |

[MEASURED — the receipts the engine cites are `data/audit/forward/` and the constraints ledger's
forward-pass section; the same numbers are restated in `architecture.md` and in the header of
`forward-pass.ts`.]

Hit@10 is the share of queries whose known correct entry lands in the top ten. A lower score is
better code doing worse work; the live configuration was 0.750 while doing strictly more computation.

## The removed stages

From `architecture.md`, "Removed from this pipeline on 2026-09-06" — every row lowered known-entry
hit@10 relative to the raw ordering:

| Removed stage | What it did | Verdict as recorded |
|---|---|---|
| Contrast sharpening | `amp ^ CONTRAST_SHARPENING_POWER` | exponent was 1.0 — inert |
| Ond/onde dual resonance | split-vector routing | removed earlier; full 768d throughout |
| Phase path walk | followed `K_ij · cos Δθ` to choose "walked" nodes | fed the partition below |
| Cascade weight | `0.4 + 0.6·(amp/maxAmp)` on walked nodes | 24–37% of top-10 membership; harmful |
| Cross-node multiplier | **×0.3** on everything not walked | a hard partition; ×1.0 lifted hit@10 by +0.15 to +0.25 on every arm |
| Additive bonuses | affinity, confidence, type, lanceBoost, length, age | tiebreakers over dead calibration; three were byte-identical dead |
| Synth re-rank | `querySim × (1 + α·proximity)` | the live partition; α → 0 lifted hit@10 from **0.579 to 1.000**, moved blind-judge useful retrieval **30/43 → 37/43** (seven gains, no losses), and rescued **96 entries** with 0 losses |
| Phase-displacement sim | 5-step forced Kuramoto → `phaseWinner` | computed every pass, never read |
| Coherence composite | 0.3 amp + 0.3 queryFit + 0.3 phaseDiscrim + 0.1 ctx | no control reader; now returns 0 |
| Density `capMap` | `dict.length / tierCap` | the caps are falsified |
| Phase handoff | `gatedKuramotoStep` / `scoreModulatedStep` | the phase engine is gone |

[FACT — `architecture.md` §"Phase Engine — REMOVED"; the same removals are stated in the header
comment of `forward-pass.ts`.]

## The three probes that killed the phase engine

The engine ran thirteen coupled oscillators and called the result learning. Three pre-registered
probes falsified it [MEASURED, recorded in `architecture.md`]:

1. **Locking capacity.** Offline simulation: the coupling cannot lock *any* edge at any strength
   within 10–20× the configured maximum. There is no regime in which the mechanism does what it was
   named for.
2. **Persistence.** Phase state is erased within roughly five passes by the natural-frequency
   rotation. Nothing it wrote survived to be read.
3. **Ablation.** A frozen rotor, a static `K_ij`, resting phases and *random* phases all produced
   retrieval indistinguishable from the live engine — and the no-walk arm scored higher
   (hit@10 0.995–1.000 vs 0.750).

## The two rules the survivor obeys

Stated in the piece and enforced in the code:

1. **A precision stage can only reorder what recall surfaced.** The cross-encoder re-ranker is
   bounded to the similarity band where the bi-encoder cannot discriminate (top similarity 0.72–0.80
   on the hosted instance, top eight scored) and returns its input unchanged on timeout or failure.
   It never changes the candidate set.
2. **An admission stage may only add.** The second-key path (the embedding of an entry's opening 200
   characters) can only put a candidate back into the list; the candidate is then scored on its own
   document vector exactly like any other. It is a union, not a max — taking the better of two scores
   would let a key evict a document match.

[FACT — `forward-pass.ts`, `retrieve.ts`, `reranker.ts`; the rules are stated in the piece and in
the comments around the admission path.]

## The measurements the piece publishes

These are the numbers a reader should be able to re-find in the pack; they are not new claims:

| Measurement | Value | Where |
|---|---|---|
| Raw no-bonus hit@10 | 0.995–1.000 | this file, `architecture.md` |
| Live weighted hit@10 | 0.750 | this file, `architecture.md` |
| Cross-node multiplier ×1.0 lift | +0.15 to +0.25 on every arm | `architecture.md` |
| Synth re-rank α → 0 | 96 entries rescued, 0 losses | `architecture.md` |
| Retrieval audit, 2026-10-04, in-corpus | hit@10 0.875, 80% grounded, 40+ probes | the engine's `retrieval-audit` run; method described in `architecture.md`; the raw output is not in this pack |
| Out-of-corpus decline rate | 1.000, every audit since 2026-09-28, 50 out-of-corpus probes | the same read-only retrieval audit; re-run it to reproduce |
| Pure retrieval time on the box | 75–160 ms | a production-box measurement; not in this pack |
| HNSW recall at a million vectors | ~99.8% of the true top ten | the engine's HNSW benchmark (whitepaper); not in this pack |

The retrieval-audit figures come from the engine auditing itself against its own store; the
ablation figures come from the harness described above. They are different instruments and should
not be averaged together.

## What this record does not prove

- It does not prove the surviving pipeline is the best possible one; it proves the deleted stages
  were not paying rent.
- It does not prove the numbers reproduce on your store. Your embedder, your chunking and your
  queries are different. Re-run the harness on your own data before trusting a delta.
- The store changed while the audit ran. The decline rate held at 1.000 across a changing dictionary,
  which is the strongest claim here, but the ablation table is a snapshot of one tree at one commit.

The code is the authority. This file is the index to it.