CHRIS HAY

IDEAS · SYSTEMS · OBJECTS / LONDON · 2026

What has to survive?

The next word. The rest of the conversation.

France’s residual replaces the Australia state; the layer readouts converge on Paris.
FROM THE FILM / 4:00

Swapping the residual stream. In the 370,000-token film, a donor state changes the answer to a capital-city question. This is the starting point for asking which state was transferred—and what survives into a later continuation.

FILM & SOURCE RECORD ↗

What state must a model retain to continue? This notebook distinguishes all-position residual state, a single token’s residual and persistent decode state. Full-sequence transplantation and MAP-5 test different boundaries; neither justifies compressing every conversation into one vector.

N-STATEOPENRECORDED 2026-09-06DRAFT · V0.1REFERENCE DRAFT

What can be rebuilt is different from what can be forgotten.

01 / KEEPING THE STATE

A cache keeps work we would otherwise repeat. If its contents can be reconstructed, we can choose another way to retain the conversation. But we still owe an account of what remains: which positions, which layer boundaries, which tokens and which metadata.

ONE OBJECT · TWO INTERPRETATIONS

Two different interventions

One positionThe full sequence
  • Replace the current token’s residual
  • Earlier positions remain available to attention
  • A changed next word need not persist

Two different interventions — as one position: replace the current token’s residual, earlier positions remain available to attention, a changed next word need not persist. As the full sequence: replace all positions at a defined boundary, transfer the tested donor computation, a different claim from persistent decode state.

EVIDENCE

Full-sequence transplant / three conditions

The recorded all-position transplant produces the donor’s next-token distribution with KL = 0 in three tested conditions at L14/L26 on Gemma 3 4B. An earlier register entry lacked an all-position patch tool; the later transplant records the completed test. The result concerns these full-sequence forward-pass interventions.

SUPPORTED

02 / AFTER THE FIRST WORD

MAP-5 asks whether a donor state can control a later branch. In its planets test, repeated patches work at L24 and L28 but fail at L20. That brackets the tested transition between L20 and L24. It does not establish that a universal thought appears at layer 24.

PERSISTENCE WAS NOT MEASURED HERE

MAP-5’s one-shot continuation used stateless, full recomputation between calls.

That harness does not retain the patched hidden state across decode steps.

Bounded hot memory is not constant total memory: the documented engine also retains a growing cold tier.

Name the state before calling it sufficient.

persistence was not measured heremap-5’s one-shot continuation used stateless, full recomputation between calls.; that harness does not retain the patched hidden state across decode steps.; bounded hot memory is not constant total memory: the documented engine also retains a growing cold tier.. Name the state before calling it sufficient.

Chris explains retained residual state and rebuilding the attention cache.
FROM THE FILM / 13:20

The residual as retained state; K/V as a hot cache. This passage discusses rebuilding attention state. The film’s memory and timing figures describe its original setup, not a new benchmark established by this notebook.

FILM & SOURCE RECORD ↗
OPEN

What changes when the patched state really persists?

The recorded follow-up compares residual-only, token-only and residual-plus-K/V interventions in incremental decoding. The K/V intervention work was explicitly deferred. This remains a research question, not a claim that the test has run.

SOURCES & PROVENANCE

  • Film / swapping the residual stream · 04:00

    Chris Hay, 24 March 2026. Main film for this note; chapter and selected automatic captions reviewed.

  • Film / We Don’t Need KV Cache Anymore? · 13:20

    Chris Hay, 10 March 2026. Selected automatic English captions reviewed.

  • Full-sequence cross-task state transplant

    Experiment EXP-20260718-002201-00277; run RUN-20260718-002202-00293. Registry ingestion date is not asserted as the experiment date.

  • MAP-5 / waypoint teleportation

    Experiment EXP-20260802-102422-00640; write-up v7. Distinguishes immediate-token control, later branch control and unmeasured persistence.

  • LARQL / MarkovResidualEngine specification

    Working specification reviewed 6 September 2026. Architectural preconditions and cold-tier growth matter; the live source may change.

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. What has to survive? [Unpublished draft, version 0.1]. https://chrishayuk.com/notebook/what-has-to-survive
DOWNLOAD