Roadmap

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

This page shows the plan of the cognitive-map line, stage 2 of Psyche, at public granularity: its phases, the layers that keep the line standing if the main goal falls short, and the paper submission deadline. What each decision gate asks, and what happens if it is not passed, is on Decision gates. The line itself is introduced in the overview.

Phases

The work runs in phases, and most close at a decision gate. Each phase’s outputs are listed in the plan file. Where a phase or gate date depends on a date that is not yet known, such as the submission deadline, the plan records the rule that will fix it instead of a date. The last phase, the public release, follows submission, and its timing is set by the disclosure policy.

Phase	Title	Start	End	Gate	Status
P0	Novelty review and setup	2026-10-23	2026-11-15	A	active
P1	Environments, battery and baselines	2026-11-16	2026-12-31	B	planned
P2	Spiking model	2027-01-01	2027-02-28	C	planned
P3	Scaling, ablations and the energy estimate	2027-03-01	2027-04-15	—	planned
P4	Paper	2027-04-15	the submission deadline (not yet announced)	D	planned
P5	Public release	after submission	set by the disclosure policy	—	planned

The chart draws the same plan on a calendar. A phase or gate whose date is not fixed is drawn out to a placeholder and labelled “not fixed”. Its position on the chart is not a date. A deadline appears on the chart only once its date is fixed; until then it is listed only under Deadlines.

Items marked * in the chart have no fixed date yet; they are drawn up to a horizon derived from the plan, not to a date.

Layers

The plan is layered so that the line does not rest on its riskiest part.

Layer	Title	Summary
L1	Safe base	Environments, the battery, baseline comparisons, analysis and the report. The line stands on this even if the spiking model does not work.
L2	Main goal	The spiking model with local plasticity, evaluated on the battery, with an analytic estimate of its synaptic operations.
L3	Failure analysis	If the main goal falls short, a documented account of where and why, kept as a result in its own right.

The safe base exists so that the line stands even if the spiking model does not work. The environments, the test battery, the baseline comparisons, the analysis and the report are a complete piece of work on their own, and the plan completes the safe base before work on the spiking model begins. The main goal builds on that base rather than replacing it. If the main goal falls short, the failure analysis records where and why, and that account is kept as a result in its own right.

The safe base feeds the main goal and is part of the reported result on its own. If the main goal works, it joins the result; if it falls short, the dashed arrow applies, and the failure analysis of where and why becomes part of the result instead.

Deadlines

Deadline	Due	Status
Paper submission to a major peer-reviewed ML venue	not set; estimate: spring 2027 (unconfirmed)	upcoming

The paper is planned for submission to a major peer-reviewed ML venue (2027). Its deadline has not been announced. Until it is, the plan carries an estimate marked unconfirmed. The end of the paper phase and the Paper gate depend on that deadline, so the plan states them as rules relative to it rather than as dates.

How dates change

• One source. tools/research_plan.json is the only source of the line’s dates: phases, gates and deadlines. The plan tables and the timeline chart on this page are rendered from it, as are the plan blocks in the other documents that show the plan. The layers diagram is drawn by hand and carries no dates.
• Rendered, not typed. The plan tables and the chart are written by python3 tools/publish/render_plan.py --write and are never edited by hand. To change a date, change the plan and render again.
• Moves are recorded. Each dated gate keeps its originally planned date as baseline_date next to its current date. A gate set by a rule records its date when the rule resolves, and that change also gets a changelog entry. A gate date that moves away from its baseline needs an entry in the plan’s changelog, which records the item and field that changed, the old and new value, the date of the change and the reason. The plan’s validator rejects a moved gate date that has no such entry. The change log is part of the public plan file.
• Fixed status words. A phase is planned, active or done; a gate is upcoming or decided; a deadline is upcoming, met or missed. A decided gate shows only the date it was decided and the next step: the line continues to the next phase, extends, or stops (see Decision gates).
• Recording a decision. A decided gate is recorded in one edit to the plan, and the plan’s validator refuses a partial one. The edit sets the gate to decided; adds decided_on, the date it was decided, and next_step, one of continues, extends or stops; and moves as_of to a date on or after decided_on. When the next step is continues or stops, it also marks the phase the gate closes as done. When it is continues, the same edit marks the next phase active, because the site and the generated blocks read the phase status from the plan and not from the date. When it is extends, the phase stays active, so the same edit gives it a new end date or adds a gate that ends it. No other text about the decision is added to the plan.

Not on this page

These stay in a private companion until submission, under the disclosure policy:

• the weekly milestones;
• the experiment plan and the order in which the experiments run;
• the criteria behind each gate.

Material reaches this repository only as complete, working, older versions, never the current state of the work.