The address is built through depth.
Relation is there early. Entity arrives late. The coordinates change in between.
01 / THE DEPTH OF A QUESTION · RECORDED MEASUREMENTS
The relation is already there.
| Residual entering | |||||||
|---|---|---|---|---|---|---|---|
| RELATION0–1 accuracy | 0.90 | 0.94 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 |
| ENTITY0–1 accuracy | 0.17 | 0.14 | 0.19 | 0.16 | 0.30 | 1.00 | 1.00 |
| BINDING0–1 accuracy | 0.18 | 0.18 | 0.23 | 0.19 | 0.45 | 1.00 | 1.00 |
↻ Basis change L8 → 12L20 → L24
↻ Basis change L20 → 24
L8–24 / mechanistic evidenceL28 → L30
Answer-token endpoint
Capital or language is readable, and changing the relation is already causally consequential. Entity and binding readers remain weak.
- L28: Perfect readability may decode the answer token; this endpoint is not clean address geometry.
- L30: Perfect readability may decode the answer token; this endpoint is not clean address geometry.
ABOUT THIS NOTE +
For capital versus language on Gemma 3 4B IT, two instruments agree: relation information is readable and causally consequential by L8; entity binding becomes decisive in the final-position residual between L24 and L28. Cross-layer reader alignment recovers relation information across changing coordinates. The L28–30 endpoint may carry the answer token, so the mechanistic evidence window is L8–24.
What does the state
actually control?
Same intervention.
Different depth.
France’s state. Still Japan’s answer.
The entity has not taken control here.
At L20, the final-position state carries a usable relation frame but does not override the recipient’s entity. By L28, the wrong-entity donor controls the answer.
The relation survives.
The coordinates change.
Not a projection of measured states.
CHANCE .50 SAME-LAYER 0.94
CHANCE .50 SAME-LAYER 1.00
Across depth, the same reader falls below chance.
Relation first.
Entity binding later.
The evidence supports staged construction in the final-position residual, through changing intermediate representations.
ADDRESS-BUILD-1
Gemma 3 4B IT · bf16 · 2026-09-06
Causal transplant + cross-layer reader
Methodology & frozen gates +
Causal half. Replace only the residual entering layer L at the final prompt position. Continue the remaining layers normally and read next-token logits; no generation and no weight editing. Four wordings, 48 retained prompts, 1,169 transplants. The different-relation arm has 23 eligible pairs per depth; the other arms have 48.
The primary gap compares full-vocabulary KL under wrong-binding and same-binding transplants. Recipient retention alone is insufficient: donor-answer rate distinguishes incompatibility from transporting a value. Self-splicing must reproduce the baseline bit-identically.
Reader Part 2b. 194 retained prompts, 20 bindings, 14 entities, capital and language, 12 wordings. Three folds hold out 4 wordings and train on 8. PCA dimension 128; ridge 1.0. Component-specific same-layer gates were frozen before measurement: relation ≥ .85, entity ≥ .40, binding ≥ .30. All pass, but entity and binding pass only in the confounded endpoint band.
The original hypothesis needed reframing. Wrong-entity swaps were as harmless as same-binding swaps through L20, so the early convergence is not binding-specific. The original causal gate condition 2 failed. Coordinate-change support is for relation only.
Bank boundary. Currency and continent could not be made reachable with the same wording bank. Relation-specific wordings would entangle relation with phrasing. This note makes no claim beyond capital versus language.
Complete measurements +
| Depth | relation | entity | binding |
|---|---|---|---|
| L8 | 0.90 | 0.17 | 0.18 |
| L12 | 0.94 | 0.14 | 0.18 |
| L16 | 1.00 | 0.19 | 0.23 |
| L20 | 1.00 | 0.16 | 0.19 |
| L24 | 1.00 | 0.30 | 0.45 |
| L28Answer-token endpoint | 1.00 | 1.00 | 1.00 |
| L30Answer-token endpoint | 1.00 | 1.00 | 1.00 |
* L28–30 may be answer decoding. Chance floors: relation .50; entity .071; binding .05. Weak early entity/binding readability is not evidence of complete absence.
| Depth / donor | ntransplants | Recipient retained | Donor answer | Median KL |
|---|---|---|---|---|
| L8 · Same binding / new wording | 48 | 56% | Not applicableSame-binding donor | 1.518 |
| L8 · Different entity / same relation | 48 | 54% | 0% | 1.512 |
| L8 · Same entity / different relation | 23 | 22% | 30% | 6.109 |
| L8 · Unrelated donor | 48 | 17% | 2% | 6.680 |
| L12 · Same binding / new wording | 48 | 58% | Not applicableSame-binding donor | 1.096 |
| L12 · Different entity / same relation | 48 | 60% | 0% | 1.121 |
| L12 · Same entity / different relation | 23 | 22% | 39% | 6.484 |
| L12 · Unrelated donor | 48 | 15% | 4% | 7.162 |
| L16 · Same binding / new wording | 48 | 96% | Not applicableSame-binding donor | 0.727 |
| L16 · Different entity / same relation | 48 | 96% | 0% | 0.755 |
| L16 · Same entity / different relation | 23 | 22% | 61% | 7.327 |
| L16 · Unrelated donor | 48 | 15% | 4% | 8.066 |
| L20 · Same binding / new wording | 48 | 100% | Not applicableSame-binding donor | 0.647 |
| L20 · Different entity / same relation | 48 | 96% | 0% | 0.692 |
| L20 · Same entity / different relation | 23 | 0% | 83% | 8.402 |
| L20 · Unrelated donor | 48 | 0% | 4% | 8.403 |
| L24 · Same binding / new wording | 48 | 96% | Not applicableSame-binding donor | 0.643 |
| L24 · Different entity / same relation | 48 | 69% | 13% | 1.171 |
| L24 · Same entity / different relation | 23 | 0% | 100% | 10.088 |
| L24 · Unrelated donor | 48 | 0% | 6% | 10.067 |
| L28 · Same binding / new wording | 48 | 98% | Not applicableSame-binding donor | 0.675 |
| L28 · Different entity / same relation | 48 | 0% | 100% | 9.412 |
| L28 · Same entity / different relation | 23 | 0% | 96% | 10.467 |
| L28 · Unrelated donor | 48 | 0% | 100% | 16.715 |
| L30 · Same binding / new wording | 48 | 96% | Not applicableSame-binding donor | 0.693 |
| L30 · Different entity / same relation | 48 | 0% | 100% | 10.008 |
| L30 · Same entity / different relation | 23 | 0% | 96% | 10.693 |
| L30 · Unrelated donor | 48 | 0% | 100% | 17.056 |
| Transition | Raw | Norm only | Procrustes | Same layer |
|---|---|---|---|---|
| L8 → L12 | 0.39 | 0.39 | 0.90 | 0.94 |
| L12 → L16 | 0.90 | 0.90 | 0.93 | 1.00 |
| L16 → L20 | 0.88 | 0.88 | 0.99 | 1.00 |
| L20 → L24 | 0.46 | 0.46 | 1.00 | 1.00 |
| L24 → L28 | 0.97 | 0.97 | 1.00 | 1.00 |
| L28 → L30 | 1.00 | 1.00 | 1.00 | 1.00 |
| Transition | Raw | Norm only | Procrustes | Same layer |
|---|---|---|---|---|
| L8 → L12 | 0.08 | 0.08 | 0.15 | 0.14 |
| L12 → L16 | 0.16 | 0.16 | 0.13 | 0.19 |
| L16 → L20 | 0.15 | 0.15 | 0.18 | 0.16 |
| L20 → L24 | 0.08 | 0.08 | 0.21 | 0.30 |
| L24 → L28 | 0.48 | 0.48 | 0.48 | 1.00 |
| L28 → L30 | 1.00 | 1.00 | 1.00 | 1.00 |
| Transition | Raw | Norm only | Procrustes | Same layer |
|---|---|---|---|---|
| L8 → L12 | 0.06 | 0.06 | 0.19 | 0.18 |
| L12 → L16 | 0.16 | 0.16 | 0.20 | 0.23 |
| L16 → L20 | 0.17 | 0.17 | 0.22 | 0.19 |
| L20 → L24 | 0.08 | 0.08 | 0.29 | 0.45 |
| L24 → L28 | 0.58 | 0.58 | 0.64 | 1.00 |
| L28 → L30 | 1.00 | 1.00 | 1.00 | 1.00 |
All displayed tables round values for reading. The JSON files retain full recorded precision, gate values and the causal KL gap at each depth.
The failed reader stays in the record +
The first Part 2 had about 1.5 training examples per binding class. Entity accuracy stayed around .14 in all 21 cells. It is an instrument failure, not evidence about the model. Part 2b repairs coverage and uses component-specific gates; the relation threshold was raised to .85 before readability was inspected.
Download preserved Part 2 result ↓Registry, hashes & source provenance +
EXP-20260906-181253-00733
CONSOLIDATED WRITE-UP / V7
SOURCE / f97a470644df1c38498030896f8f6cfffe7689b1
The registry snapshot comes from chuk-experiments. Result JSON was extracted from the registered commit in the local LARQL repository. The public GitHub API did not resolve that commit at verification time; the frozen downloads below remain available with this note.
- Causal rows SHA-256 (reported by harness)
- 9e039cc98dda26749773df55b468ced2bc522d592b8eb72a73dae9e29924a7ba
- Reader Part 2b result SHA-256 (reported by harness)
- 516bd1139a73622080890958b5488261cad3e9e889642364ed65b2b450fdcebf
These are the harness’s payload hashes, not hashes of the formatted download files. The NPZ state captures were excluded by the source repository’s *.npz rule. Recorded aggregates support this recreation; they are not a complete state-capture archive.
What is the map
actually a map of? ↗
SOURCES & PROVENANCE
- ADDRESS-BUILD-1 — frozen experiment registry record ↗
Experiment EXP-20260906-181253-00733; consolidated write-up version 7. Retrieved from chuk-experiments on 2026-09-06.
- Causal transplant — complete recorded aggregates ↗
Extracted from local git object f97a470. Rates aggregate the eligible prompts in each arm.
- Reader Part 2b — complete results ↗
Component-specific readability, all transfer controls and frozen gates.
- Reader Part 2 — preserved instrument failure ↗
Insufficient examples per binding class; no mechanistic inference is drawn from this failed reader.
AUTHOR / CHRIS HAY · VERSION / 0.1
REFERENCE THIS DRAFT
An unpublished working record. These references identify the draft and omit a publication date. They become version-specific publication citations when the record is released.
Chris Hay. The address is built through depth. [Unpublished draft, version 0.1. First publicly recorded 2026-09-06]. https://chrishayuk.com/notebook/the-address-is-built-through-depth
FOLLOW THE WORK
New notebook entries and recorded work, as they appear. Point a feed reader — or an agent of your own — at an address below. No account, no email address, nothing for this site to keep.
The notebook
New ideas, experiments and essays, as they are recorded. Includes labelled working drafts.
OPEN FEEDhttps://chrishayuk.com/notebook/feed.xmlThe record
Everything published to the Chris Hay record.
OPEN FEEDhttps://chrishayuk.com/record/feed.xml
FOR PROGRAMS · follow.json · JSON Feed