Claims and scope

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

This page sets what the cognitive-map line, stage 2 of Psyche, may say about its model, what the line covers, and how a new idea is kept out of it. The rules govern every public sentence about the line: these pages, the paper, the project’s other public text and any talk. They extend the claims rule that the whole project follows (rule 1 in CONTRIBUTING.md).

The project’s aim

Psyche is an open research project that builds brain-inspired components in stages, with artificial general intelligence (AGI) as its long-term aim (D-033). The aim may be named anywhere, as an aim. No public text describes Psyche, the substrate, the spiking model or any other component as having reached that aim, and no result of this line is presented as a measured step toward it. The line asks a narrow question about one measurable class of behaviour, and its claims stay inside that question.

The core rule

In this line, “cognition” names a class of behaviour that a test can measure. It is never a claim about inner states.

A sentence that uses a cognitive word must be rewritable as a statement about what an agent did in a defined test, with a captured result behind it. If it cannot be rewritten that way, it is not written.

The ceiling on claims

The strongest statement the line may make about its model is fixed in shape, not in content: it names the behaviours of the test battery that the model showed, each backed by a captured result, and reports their cost as an analytic estimate of synaptic operations beside the same estimate for the comparators. A behaviour that was not shown is left out rather than softened. A cost comparison is stated in whichever direction the captured result gives, and only under the energy claims rules below.

Never

Each line below applies to every public text about the cognitive-map line.

• Never: say the model “understands” or “thinks”, or use “consciousness” about it.
• Never: call it a mind, sentient, aware, general intelligence or a validated cognitive model.
• Never: “the brain does it this way”. Say “brain-inspired” or “biologically plausible”.
• Never: a general claim that spiking networks are more efficient than conventional networks.
• Never: a priority claim, such as “the first” or “first model to”, before release. A decided Gate A does not lift this, whatever its outcome (D-034).
• Never: energy presented as measured. It is an analytic estimate or a lower bound.
• Never: a number without a captured source, meaning a results file produced by code.
• Never: a planned result written as if it had been obtained.

Efficiency is stated only in the narrow form “on this task, with this measure, under these assumptions”. The phrases on this list are also kept as data in tools/ledger.json under research_forbidden_claim_terms, and the project-wide terms under forbidden_claim_terms, so that tooling can check public text against them.

Scope

In scope

Area	What the line covers
Environments	The three families of the test battery: grid worlds with repeating observations, where the current observation alone does not identify the location; growing mazes; and abstract relational structures, such as an ordered chain.
Tests	The five classic paradigms of the test battery: latent learning, detour, shortcut, structural transfer and transitive inference.
Comparisons	The comparator families of the test battery: model-free reinforcement learning, recurrent model-free agents, a small transformer agent, the successor representation and published cognitive-map models.
Repetition	At least five seeds per model and test, with bootstrap confidence intervals (see the test battery).
Cost	An analytic estimate of synaptic operations, under the energy rules below.

Out of scope

• Measurement on neuromorphic hardware. Energy is estimated analytically instead.
• Three-dimensional or visual environments.
• Language tasks.
• Large-scale benchmarks, such as ImageNet or continual-learning image benchmarks.
• Theoretical analysis of new biophysical learning rules.

The scope-creep test

A new idea joins the line only if the answer to all three questions is yes:

1. Does it directly answer one of the research questions (RQ1 to RQ4 in the overview)?
2. Can it be finished before Gate C?
3. Would a reviewer accept it without asking why it is there?

If any answer is no, the idea is parked: it is recorded and not started.

The diagram asks the three questions in order: an idea joins the line only with three yeses, and a single no parks it.

Energy claims

A spiking network simulated on a GPU runs on hardware that does not exploit spikes, so its energy use says nothing about spiking efficiency. Any energy figure the line reports is therefore an estimate of what the computation would cost, never a reading of what it did cost. The principles:

• Estimates and lower bounds only, named as such. Every energy figure is called an analytic estimate or an analytic lower bound wherever it appears: in tables, in figures and in prose. The limitation is stated in the text before a reviewer has to ask for it.
• Hardware readings are never evidence. GPU wall-clock time and watt readings are never presented as evidence of energy efficiency.
• No borrowed hardware figures. The estimates are not set against energy figures reported for neuromorphic chips.
• No general efficiency claim. Whether spiking networks save energy depends on the hardware, on how sparse their activity is and on what is counted, memory access included. Dampfhoffer et al. (2023) show this with a hardware-aware analysis, and it is the reason the line makes no general claim. Efficiency is stated only for this task, this measure and these assumptions.
• Like for like. Every model is estimated under the same assumptions, so a comparison reflects the models and not the accounting.
• Sensitivity analysis is mandatory. The estimate rests on assumptions about the cost of each operation and about what is counted, so every energy result is reported across a sensitivity analysis of those assumptions.
• A ranking must survive every setting. A ranking, including any cost comparison in the statement above, is stated only if it holds in every setting of the sensitivity analysis. Otherwise the text says the comparison depends on assumptions.
• The counter is checked before it is trusted. The operation counter is tested and checked against an established reference count before its numbers are used.

The formula, the counted quantities, the cost constants and the sensitivity settings are part of the withheld method.

Writing rules

Rule	In practice
Paragraph shape	A claim, then its evidence (a figure, a table or a citation), then a limitation or a transition.
Sentences	Short and active. “X does Y (Fig. N)” rather than “we show that”.
Terms	One name for one thing, used the same way everywhere. A second name for the same thing is a defect.
Numbers	Mean with a confidence interval. A value rather than “approximately”. “Significant” only with a statistical test behind it.
Restraint	“Novel” and “original”, as claims about the work, only with evidence. No priority claim before release, whatever a gate decides, no marketing language, no exclamation marks.
Figures	Every axis labelled with its unit, a colour-blind-safe palette, and every figure cited in the text. Figures are generated from code.
Citations	Primary sources, with a DOI, and the year and venue checked. A reference is used only after it has been checked against its primary source; a second-hand citation is not enough.
Status words	Lowercase in the line’s prose, exactly the vocabulary of tools/research_plan.json: planned, active and done for phases; upcoming and decided for gates; upcoming, met and missed for deadlines.
Drafting	Skeleton, then a few points per section, then prose, then alignment with the figures, then a cut.

Every public text of the line follows the same rules as the paper.