Decision gates

Public summary. Detail is withheld until submission under the disclosure policy.

The plan of the cognitive-map line, stage 2 of Psyche, has decision gates labelled A to D. A gate is a fixed point in the plan where a decision is taken against criteria written in advance. Each gate falls at or near the end of a phase of the roadmap, and the decision sets the route the next phase takes.

Why gates exist

The decisions that shape the line are taken at fixed points, not gradually and not after the convenient answer is known. Each gate has three parts:

Part	What it is	Where it lives
Question	What the gate decides	Public, in the plan
Criteria	What counts as a yes, written before the work they judge	Private until submission
Fallback	The route taken if the answer is no, also fixed in advance	Public, at summary level

Criteria are written in advance so that a result cannot reshape the test it is judged by. Changing a criterion after it has been written needs a recorded decision with its reason; it is never a silent edit. This is the same discipline the substrate, stage 1 of Psyche, applied to its pre-registrations (see Origins).

Missing a gate sends the work along the fallback fixed in advance for it; it does not by default end the line. The plan is layered so that the line stands on its safe base even if the spiking model does not work, and a documented account of where and why it fell short is kept as a result in its own right. The layers are described in the roadmap. After any gate, the next step is one of three: the line continues to the next phase, extends, or stops (see what the public status means).

The gates at a glance

The table and the flow below are generated from tools/research_plan.json, the only source of gate dates, and are not edited by hand.

Gate	Title	Date	Status	Question
A	Novelty	2026-11-15	upcoming	Is this line’s combination of aims new, or is its difference from the closest published work written down and defensible?
B	Comparisons	2026-12-31	upcoming	Do the environments, the battery and the baseline agents run, and do the baselines reproduce the behaviour their papers report?
C	Spiking model	2027-02-28	upcoming	Does the spiking model meet its pre-registered behavioural criteria on the battery?
D	Paper	one week before the submission deadline	upcoming	Are all figures, the multi-seed results and the ablations in, and have external readers’ comments been worked in?

Each arrow leads to the next gate in plan order. If a gate is missed:

• Gate A (Novelty): Reposition the contribution and record the decision privately before work continues.
• Gate B (Comparisons): Extend briefly, then narrow the baseline set to the strongest published reference.
• Gate C (Spiking model): Shift the submission target; the work proceeds on the safe base and the failure analysis.
• Gate D (Paper): A workshop paper and a preprint; a later main-track venue.

The flow shows the four gates in order, each with the route the work takes if that gate is missed. The sections below describe each gate in turn.

Gate A: Novelty

Gate A closes the phase “Novelty review and setup”. It asks whether this line’s combination of aims is new. If closely related work has been published, the gate asks instead whether the difference from it is written down and can be defended.

If the answer is no, the contribution is repositioned, and that decision is recorded privately before work continues. The public record of any gate decision is only its date and the next step (see what the public status means).

Gate B: Comparisons

Gate B closes the phase “Environments, battery and baselines”. It asks two things: whether the environments, the test battery and the baseline agents all run, and whether the baseline agents reproduce the behaviour that published work reports for them. A comparison means little if the reference agents do not behave as published, so the references are checked before anything is compared against them.

If the answer is no, the phase is extended briefly; if the baselines still fall short, the baseline set is narrowed to the strongest published reference.

Gate C: Spiking model

Gate C closes the phase “Spiking model”. It asks whether the spiking model meets its pre-registered behavioural criteria on the battery. Which tests and environments those criteria cover is withheld until submission.

If the answer is no, the submission target shifts, and the work proceeds on the safe base and the failure analysis.

Gate D: Paper

Gate D falls near the end of the paper phase, shortly before submission to a major peer-reviewed ML venue (2027). It asks whether the paper is complete in substance: all figures, the multi-seed results and the ablations are in, and the comments of external readers have been worked in. Its date is tied to the submission deadline, which has not been announced; the plan records the rule.

If the answer is no, the route becomes a workshop paper and a preprint, with a later main-track venue.

What the public status means

Each gate shows one of two public status words:

Status	Meaning
upcoming	The decision has not been taken yet.
decided	The decision has been taken against its criteria. The public record holds two facts and no others, listed below.

When a gate is decided, the public record states only:

• The date it was decided.
• The next step: the line continues to the next phase, extends, or stops.

The gate table shows both in the Status cell of a decided gate. Nothing else about a decision is public: no verdict detail, no evidence and no reasoning (D-034). The next step says whether the line goes on, not why. The full record of a decision stays in the private companion repository and moves here only under the disclosure policy. How the plan is edited when a gate is decided is described on the roadmap.

The site also reads each gate against the day a page is opened. A gate whose date has passed with no recorded decision shows as “decision pending”, with its date. After a decision that stops the line, the gates behind it show as “not reached”. Both are display wording only; the plan’s status words stay upcoming and decided.

A decision does not lift a claim rule. A term that the claims discipline forbids before a gate stays forbidden until release, whatever the outcome. A decided Gate A, in particular, does not allow a priority claim.

The criteria behind each gate are pre-registered privately. When a criteria file is frozen, a salted SHA-256 hash of it is added to the commitments list in the plan, with the date. The date is public from that moment; the content is published at submission, and anyone can then check it against the hash. The disclosure policy describes the procedure.

Gate dates can move. A moved date keeps its original baseline in the plan, and the move needs an entry in the change log of the public plan file; the roadmap explains how dates change.