Start here
An open research project on brain-inspired components
Psyche
Brain-inspired components for machine intelligence, built and tested one stage at a time.
Psyche is an open research project. Its current stage tests whether spiking neural networks that learn only through local plasticity build a map of their environment, on the classic cognitive-map experiments, and will report the result at release, whichever way it falls.
- Stage 1The substratePaused, kept
- Stage 2Spiking cognitive mapsActive
- Later stagesToward AGIOpen
Now in stage 2: P0 Novelty review and setup, 23 October 2026 to 15 November 2026
Next gate: Gate A Novelty, 15 November 2026
One project, in stages
Each stage stands on the one before it and is kept as a citable artifact, so the work behind it stays usable wherever the road stops.
| Aspect | Stage 1 The substrate | Stage 2 Spiking cognitive maps | Later stages Toward AGI |
|---|---|---|---|
| State | Paused, kept Since 23 October 2026, decision D-031 | Active Since 23 October 2026, decision D-031 | Open |
| What it asks | Can a memory-and-activation substrate move information without loss and set how much activation it spends without producing output its inputs do not support? An aim, paused and not achieved. | Does a spiking network that learns only through local synaptic plasticity show, on classic cognitive-map tests, the behaviours that an internal map would explain and memorised responses would not? The plan does not depend on a positive answer. | Later stages build toward the long-term aim on the same discipline and on the project's own designs. |
| What is kept | A dependency-free Python memory-and-activation substrate whose lower layers are verified correct and deterministic. Its three attempts to grade a cognitive effect each stopped before any graded number, at the external yardstick; that lesson shaped stage 2. | If the spiking model falls short, a documented account of where and why is kept as a result in its own right. The summary is public here; the detail stays private until release. | Not specified publicly yet. |
| Read | None |
The test battery
The classic behavioural paradigms the battery is built from, drawn as illustrations, not results.
- 01ExploreReward-free exploration: the trail covers the space, no goal exists yet.
- 02Goal introducedAfter exploration, a goal is introduced.
- 03Direct routeThe direct route to the goal.
Stage 2 roadmap
The current research line, Cognitive Map Formation and Energy Efficiency in Spiking Neural Networks with Local Plasticity, runs in 6 phases with 4 decision gates. Choose a phase to see what it delivers; the full schedule is on the timeline page.
P0: Novelty review and setup
Active23 October 2026 to 15 November 2026
Outputs
- novelty decision recorded
- development environment
Ends at Gate A (Novelty), on 15 November 2026: Is this line's combination of aims new, or is its difference from the closest published work written down and defensible?
P0 Novelty review and setup
23 October 2026 to 15 November 2026 Active
Gate A (Novelty): 15 November 2026
P1 Environments, battery and baselines
16 November 2026 to 31 December 2026 Planned
Gate B (Comparisons): 31 December 2026
P2 Spiking model
1 January 2027 to 28 February 2027 Planned
Gate C (Spiking model): 28 February 2027
P3 Scaling, ablations and the energy estimate
1 March 2027 to 15 April 2027 Planned
P4 Paper
15 April 2027 to the submission deadline (not yet announced) Planned
Gate D (Paper): one week before the submission deadline
P5 Public release
Starts after submission; end set by the disclosure policy Planned
How the work is built
The method came with the substrate and carries into every stage.
One ledger
Every count and status lives in one ledger. A check runs in CI and fails when a document disagrees with it.
Separate lanes
A builder implements test-first, a tester attacks the work independently, and a validator tests any cognitive claim against external data. The lanes do not cross.
Adversarial and loop-closed tests
Tester suites are hostile and written apart from the code they attack. Each part is driven with input generated by the part before it.
Determinism across 5 interpreters and 6 hash seeds
Every registered suite runs on each CPython version under each hash seed. A property that holds on one interpreter only was never a property of the program.
Read the engineering standardsRead the substrate's methodology
What is public, and what is not yet
Open
Published here
On this site and in the open repository.
- The research line's overview and questions
- The roadmap and the decision gates
- The test battery's paradigms and fairness rules
- The claims discipline and engineering standards
- The substrate's code, tests and published record
The claims discipline sets what these pages may say.
Read the claims and scopePrivate
Not public yet
Kept private. Only a release under the disclosure policy moves material here.
- The spiking model and its code
- Hypotheses beyond the research questions
- Pre-registered criteria, committed here by hash only
- The experiment plan and its order
- The analysis of related work
- The verdict, evidence and reasoning behind a decided gate
Releases follow a fixed rule: complete, working, older versions only.
Read the disclosure policyThe substrate, maintained
The substrate's development is paused, and its record stays here and stays verified.
- Suites
- 26
- Assertions
- 2,645
- Failures
- 0
- CPython versions
- 5
- 3.10 to 3.14
- Hash seeds
- 6
- 0, 1, 2, 3, 7, 42
From the blog
Dated posts on what is public: the substrate's record, how the work is done and checked, and the reading behind it.
Testing the test before the model: a random player at 22%
Psyche's second stage tested its test equipment before any model: a random player solved 22% of cases, so the check did not pass and the test is redesigned.
Three questions a test answers before its number counts
Three checks Psyche runs on an experiment before any number is graded, shown in code on the substrate's priming, fan-effect and interoception tests.
Why Psyche keeps a dated public blog: what its posts will cover, what they will not, and how each post is checked before it reaches the site.
Cite and contact
New to Psyche? The Reader's guide is the place to begin.
- Cite
To cite Psyche today, cite the archived report of its first stage.
Caliskan, C. (2026). Nous: Adversarial Testing of a Dependency-Free Memory-and-Activation Substrate for Cognitive Architectures. Zenodo.
The archived title uses the project's earlier name; the project is Psyche.
https://doi.org/10.5281/zenodo.20474613- Contact
To ask a question or raise a point, write in GitHub Discussions on the repository.
Open GitHub Discussions