# Reasoner line registration

## Current source update: 2026-09-05

The current upstream reference is `9e147326c9cd1e6dcc26591ff5d73e16bff9e17e`.
The complete 5.0–5.4 series records two passes and three no-go results. The 5.5
public-family work and speed comparison provide the current search baseline;
5.6 records a development search no-go and channel-readiness failures.

The [shared research report](../experiments/research-step-1/REPORT.md) and
[bound diagnostic result](../experiments/research-step-1/RESULT.json) record
the current evidence and next fresh-family comparison. The records below
preserve the earlier 3.x registration and its then-proposed 4.0 boundary.

## Historical 3.x registration

Reasoner is the lab's active scientific model line. ZERO.5 remains preserved
as a frozen training lineage, and ZERO.4 remains the promoted quantity model.

This registration changes research priority. It does not convert earlier
Reasoner runs into ilXyr-native evidence and does not authorize a new
experiment, training job, cloud launch, sealed evaluation, or model promotion.

## Source state

The complete Reasoner 3.x source and evidence history is consolidated by
[zero-grounded-literary-lm PR #180](https://github.com/atimics/zero-grounded-literary-lm/pull/180)
at source revision `3b917b6f54d151a43dd65e45f094606272d166c5`.
That integration keeps the successful line and its final no-go branches in one
reviewable history.

The imported status is `external_public`. ilXyr has not replayed or settled
the historical runs. Their original source commits, frozen contracts, result
digests, and chronology remain authoritative until separate evidence imports
are completed.

## Evidence boundary

The strongest result at this registration was Reasoner (3,9). Across 744 fresh sealed
episodes in dimensions nine through twelve, it identified every registered
law, made all 2,844 active queries exactly, completed every action and commit,
and passed all registered controls.

This supports one narrow claim: active compositional law induction inside a
fixed typed integer program language. It does not establish open-ended
mathematical reasoning, invention of operators outside that language, natural
language reasoning, or transfer from an arbitrary external representation.

The line also preserves the failures that define this boundary:

- Reasoner (3,3,1) solved every relational graph but used eight non-minimal
  traces, so that arm is no-go.
- Reasoner (3,7) missed 336 of 482,304 raw-observation decisions, so ordinary
  approximate fitting is no-go at the exact transfer gate.
- Earlier Cartan proposal and repair experiments failed their exact held-out
  or causal-use gates before the line changed the task and method.

## Next question

The proposed next boundary at this registration was Reasoner 4.0 representation transfer:

> Can the frozen Reasoner (3,9) core infer a structured input adapter from
> examples, then solve a fresh sealed representation without a task label,
> hand-coded domain switch, or core change?

A future contract should test an unfamiliar representation of familiar laws
before combining unseen representation and unseen law composition. It should
require exact raw-input-to-IR replay, exact reasoning traces and sealed
answers, and lookup, fixed-adapter, task-label, and scrambled-example controls.

No such contract exists in this registration. Governance therefore records
the Reasoner line as active with `active_experiment: null`.

The prepared
[`Reasoner 4.0 representation-transfer preflight`](../examples/diagnostics/reasoner-4-representation-audit.json)
turns the first boundary into a reviewable draft. It tests typed-role signal on
held-out surface forms with shuffled-example, header-only, random-projection,
and role-swap controls. A stable result opens an exact adapter contract. The
future sealed law-composition panel stays outside this preflight.