The substrate

Stage 1 of Psyche: the substrate. This is the foundation of Psyche, an open research project that builds brain-inspired components in stages with AGI as its long-term aim, and the origin of the project’s method. Its development has been paused since 2026-10-23 (D-031) while stage 2, spiking cognitive maps, is the active line; its code, suites, ledger and this record stay maintained. Project overview: README; how the substrate led to stage 2: Origins. The text below was the repository README until 2026-10-23, when the substrate was the whole project, so in it the name “Psyche” refers to the substrate; About this page lists every edit made since.

A biologically-motivated, dependency-free Python prototype of a memory-and-activation substrate for a cognitive architecture. Psyche implements, with no third-party dependencies, a perceive-to-broadcast cycle that now closes end-to-end: (1) a graph substrate store of concept nodes and weighted roads with consolidation-inspired dynamics, lazy activity-dependent decay, and “sticky” closure of traumatic edges; (2) a deterministic spreading-activation engine that records per-edge flow and treats closed roads as leaky-not-inert (an intrusive-recall-like breakthrough); (3) a multi-source activation field that assigns each node to the source that activates it most; (4) a source-clustering pre-stage that groups nearby perceived sources before the field is built; and (5) a workspace stage that runs one competition among coalitions, prunes the bridged unperceived material from the winner, and performs a single bounded broadcast back onto the substrate that biases the next cycle — closing the loop perceive → activate → cluster → compete → prune → broadcast → next cycle.

The substrate and activation layers are verified correct and deterministic. The source-clustering and workspace layers are mechanism-verified (independently and adversarially tester-verified for determinism, purity, integrity, and stability) but their cognitive meaning is not claimed — closing the loop is a mechanism result, not a cognitive one. The first behavioral cognitive claim has been attempted three times and refused three times, each for a structurally distinct reason, and each before any graded number existed: semantic priming was circular (the external norm was the edge weight), the fan effect found the generating mechanism absent, and interoception found the signal flat on realistic work. The meaning bar is still open, and the recurring obstacle is the external yardstick rather than the mechanism.

---

What this is (and how to read it)

The substrate is an engineering prototype and a software-testing-methodology case study: a small, self-contained system built test-first, where adversarial testing caught a sequence of real structural bugs that the write-up (docs/METHODOLOGY.md) documents in detail.

The long-term aim of the whole project, Psyche, is artificial general intelligence (AGI), stated only as an aim. This substrate is not AGI and makes no such claim. It is not a validated cognitive model, not a claim of machine cognition, and not a “mind” or “consciousness.” The neuroscience it borrows vocabulary from (below) is conceptual inspiration, not an implemented or reproduced result. Where the system does something unverified, this documentation says so plainly.

Note on the docs/design/ folder: it contains the project’s original aspirational vision documents, written before this code and using far grander language than the code earns. They describe intent, not what is built. See docs/design/README.md for the disclaimer. The accurate account of the real system is here, in docs/ARCHITECTURE.md, and in docs/METHODOLOGY.md.

---

Citation

This work is archived on Zenodo with a permanent DOI: https://doi.org/10.5281/zenodo.20474613

If you use or reference Psyche, please cite:

Caliskan, C. (2026). Nous: Adversarial Testing of a Dependency-Free Memory-and-Activation Substrate for Cognitive Architectures. Zenodo. https://doi.org/10.5281/zenodo.20474613

The report is open access under CC BY 4.0; the code is MIT-licensed (see LICENSE). A machine-readable CITATION.cff is provided at the repository root.

---

Scope: three assurance tiers (kept strictly distinct)

Not everything in the repository carries the same assurance, and the project’s first discipline is to never let one tier borrow the credibility of another. Three tiers are in play, from strongest to weakest.

Tier 1 — verified (substrate + activation)

The ledger’s status name for this tier is PROVEN. It means: verified by the full adversarial suite on every interpreter; it is not a formal proof.

The substrate store and the spreading-activation engine are verified for three properties:

• Determinism: identical results across process runs, including under randomized hash seeds (PYTHONHASHSEED) checked in fresh subprocesses. This matters because the natural implementations leak set-iteration order (over uuid4 keys) into results.
• Integrity: the graph’s indices, adjacency, dedup, and designations stay mutually consistent across every mutation; a full-graph scan finds no drift.
• Purity: reads never mutate stored state. Lazy time-decay is computed as a read-only view; only writes persist it. (This is load-bearing: a read that silently re-opened a “closed” traumatic edge would corrupt state over time. A test asserts it cannot.)

These are claims about mechanism: the code does exactly and only what it says, reproducibly.

Open defect, stated here rather than in a footnote (2026-07-25). The determinism bullet above is contested for the activation engine. An independent adversarial pass found that convergent accumulation is a naive sequential float add whose term order comes from iterating a set of uuid4 ids, so on graphs where three or more contributions reach the same node, fresh processes disagree in the last place: at library defaults, 48 processes produced two results differing by 1 ULP. It holds bit-exactly at in-degree ≤ 2, the substrate store is not affected, and the project’s existing probes could not have seen it because they compare round(x, 9). It is the same bug class as the one described under “Determinism is scoped to the interpreters it was swept on” below, one layer upstream, and it is not yet fixed. Full record: docs/BUG-0b1-coalition-strength-determinism.md §8. Whether Tier 1 keeps the PROVEN label while it stands is an open decision in DECISIONS.md (D-011), not something quietly resolved in a README.

Later record: this defect was fixed and Phase 0 returned to its earlier status (D-013, D-016 in DECISIONS.md).

Tier 2 — tester-verified mechanism (the closed perceive-to-broadcast cycle)

Coalition formation, the multi-source field, source-clustering, and the workspace (competition + pruning + bounded broadcast) make the cycle close end-to-end. They are independently and adversarially tester-verified for the same hard properties as Tier 1 — determinism across processes, purity, integrity — plus the stability properties the feedback write requires (the broadcast does not diverge and does not lock in; see below). This is a mechanism result only: the loop runs, deterministically, and does exactly what it says.

What changed since the earlier draft of this README:

• The cycle now closes. The old “strongest coalition = the thought” placeholder is gone. A real workspace competition selects one winner, prunes the bridged unperceived material from it, and a single bounded broadcast writes the winner’s reinforcement back onto the substrate to bias the next cycle.
• The old over-swallow known limitation is resolved at the mechanism level. Source-clustering groups nearby perceived sources before the field is built, and the workspace prunes the bridged unperceived domain from the winner. A residual corner case remains and is shipped as an explicit asserted-negative (see “Known limitation” below).

Crucially, Tier 2 makes no cognitive claim. Closing an information loop with bounded feedback is an internal substrate property, not priming, awareness, or understanding. The status of these layers is therefore BUILT · UNVALIDATED: mechanism-verified, meaning not yet earned.

Tier 3 — disciplined-exploratory (experiment scaffolding)

A separate body of pre-registered experiment scaffolding (a SWOW association-norm loader, a fan-effect characterization, and a graded semantic-priming harness) lives under experiments/ and is genuinely depended upon by the structural check below. It is disciplined and pre-registered, but it is a distinct, lower tier: it does not hold PROVEN and it has not had a validator pass. It is an instrument; its existence does not establish a cognitive result. (The instruments’ mechanism suites — test_harness.py, test_priming_instrument.py, test_priming_1e.py — were since brought into the canonical runner and are tester-verified, per SPEC-1e §10. The data-dependent experiment-result scripts stay out: they need the ~53 MB SWOW file and make meaning claims.)

One structural check has already been run on this scaffolding and is reported as the negative it is: the source-clustering grouping was compared against an independent standard partition of the SWOW graph (a different clustering-algorithm family). It beat the no-clustering control decisively (sign test 170/200 probes, p ≈ 6e-25) but missed the pre-registered absolute bar (mean ARI 0.228 < 0.30 — it under-merges relative to a coarse standard partition). Verdict: NOT YET. This is reported as agreement with an independent standard partition, not ground-truth recovery, and it does not move the source-clustering layer out of BUILT · UNVALIDATED.

The line that matters

“The mechanism is correct and deterministic” is verified (Tier 1) or tester-verified (Tier 2); “the mechanism reproduces a cognitive phenomenon” has not been shown. The meaning bar remains open after three structurally distinct refusals — see “Limitations & future work”.

---

Running the tests

Pure standard library: nothing to install (Python 3.8+; the current capture is swept across CPython 3.10, 3.11, 3.12, 3.13 and 3.14, with 3.14 as the transcript interpreter).

# all 26 suites, with the correct PYTHONPATH handled for you:
python3 run_tests.py

# or a single suite directly (absolute PYTHONPATH so the determinism subprocess
# probes resolve their imports regardless of working directory):
PYTHONPATH="$PWD/src" python3 tests/test_psyche_store.py

The suites use a tiny built-in harness, not pytest: each prints RESULT: N passed, M failed and exits non-zero on any failure. Do not run them under pytest: pytest would collect the test_* functions but would not register a logical check(...) failure as a failed assertion, so it could report green on a real failure. Run them as scripts (or via run_tests.py).

Real results

Captured verbatim in docs/TEST_OUTPUT.txt. Every suite passes with zero failures. These counts are mirrored in tools/ledger.json, which is the single source of truth; tools/ledger_check.py fails the build if any document disagrees with it.

Suite	Layer / role	Checks passed
test_psyche_store.py	substrate store	67
test_psyche_activation.py	spreading activation	102
test_psyche_coalitions.py	coalition formation	19
test_psyche_field.py	multi-source field	13
test_psyche_source_clustering.py	source-clustering (1c)	49
test_psyche_workspace.py	workspace competition + pruning (1d-i)	96
test_psyche_workspace_broadcast.py	bounded broadcast + feedback (1d-ii)	63
test_psyche_perceive.py	perception + closed end-to-end cycle	19
test_psyche_bodysensor.py	body-state sensor + modulation hook (2c)	49
test_psyche_credit.py	temporal credit assignment + depression path (2f-i, 2g)	133
test_tester_1c_adversarial.py	tester: adversarial 1c	73
test_tester_1c_mutation_battery.py	tester: 1c randomized mutation battery	5
test_tester_1d_i_adversarial.py	tester: adversarial 1d-i	107
test_tester_1d_ii_adversarial.py	tester: adversarial 1d-ii	110
test_tester_1e_selector_adversarial.py	tester: 1e blind selector	92
test_tester_1e_instrument_adversarial.py	tester: 1e broadcast-Δ instrument	76
test_tester_1e_loopclosed.py	tester: 1e loop-closure	58
test_tester_0b2_adversarial.py	tester: adversarial 0b2 normalization overlay	104
test_tester_0d_adversarial.py	tester: adversarial 0d float-accumulation detector	342
test_tester_0b2_reverify.py	tester: independent 0b2 re-verification	246
test_tester_0c_adversarial.py	tester: adversarial 0c road ordering + fsum	150
test_tester_2f_adversarial.py	tester: adversarial 2f-i, 0e, 0g and 2g	196
test_tester_2c_adversarial.py	tester: adversarial 2c interoception	350
test_harness.py	experiment harness mechanism	17
test_priming_instrument.py	priming instrument mechanism	42
test_priming_1e.py	blind three-arm selector mechanism	67
Total	26 suites	2,645 passed, 0 failed

That total is 2,645 assertions across 26 suites, 0 failures, including the builder suites above and the independent tester-authored adversarial suites (test_tester_*), which add hostile and loop-closed checks: cross-process determinism in fresh interpreters under PYTHONHASHSEED ∈ {0,1,2,3,7,42}, randomized mutation batteries interleaved with integrity scans, no-divergence / no-lock-in stability of the broadcast write, and the resolved over-swallow case now driven through the live cycle (formerly an asserted-negative limitation, now a real pass).

A passing count is not a verdict. When test_tester_0b2_adversarial.py first ran (2026-07-25), it had 83 passing checks and the tester’s verdict on the brick it examines was FAIL: seven of nine specified invariants held and two did not. Its three defects were recorded as asserted negatives — checks that passed because the defect was present and were designed to turn red the moment it was fixed, so that a fix had to come back through the same suite. The defects, their scope, and which of them the change introduced versus inherited are recorded in docs/specs/SPEC-0b2-conserved-activation.md §6. All three were later closed (DEFECT-A by brick 0c, the other two by follow-up fixes); those checks now assert the fix, the suite now has 104, and a second, independent pass, test_tester_0b2_reverify.py, returned PASS on all nine invariants on 2026-07-26 (§9 of the same spec). Read the verdict, not the count.

Determinism is scoped to the interpreters it was swept on. A cross-process determinism failure lived undetected in Coalition.strength for a month because the suite was only ever run on one CPython minor version: float summation over a uuid4-keyed dict is order-dependent, so two identical coalitions differed by one ULP and that resolved the competition before the creation_seq tie-break could run. It was invisible on 3.12 and flipped the winner 8 times in 30 on 3.11, at a fixed hash seed. Fixed with math.fsum; see docs/BUG-0b1-coalition-strength-determinism.md. Captures are now produced by tools/capture_run.py, which sweeps every CPython on the machine and refuses to capture if they disagree. The current capture was swept across five CPython minor versions — 3.10, 3.11, 3.12, 3.13 and 3.14 — on macOS (Darwin arm64), all producing identical per-suite counts and identical verdicts. The determinism claim is therefore scoped to those five versions and to PYTHONHASHSEED ∈ {0,1,2,3,7,42}, and tools/ledger.json records that scope so it cannot silently narrow.

---

Project size (measured)

	Files	Lines (wc -l)	Non-blank, non-comment (approx.)
Source modules (src/)	9	—	—
Test suites in the canonical runner (tests/ + experiments/)	26	—	—
Python total	35	—	—

Line counts are deliberately not tabulated here any more: they were hand-maintained, they went stale, and nothing depends on them. Run wc -l src/*.py tests/*.py for current figures. Assertion counts, which are load-bearing, live in tools/ledger.json and are checked against every document by tools/ledger_check.py.

---

Project structure

psyche/
├── README.md               ← you are here
├── LICENSE                 ← MIT
├── requirements.txt        ← standard library only (no dependencies)
├── run_tests.py            ← runs all 26 suites with the correct absolute PYTHONPATH
├── src/                    ← the 9 modules
│   ├── psyche_store.py               substrate store (Layer 1): graph, IDs, decay, closure
│   ├── psyche_activation.py          deterministic spreading activation + flow
│   ├── psyche_field.py               multi-source activation field (basins)
│   ├── psyche_coalitions.py          coalition formation (flow-cut & by-basin)
│   ├── psyche_source_clustering.py   source-clustering pre-stage (group nearby sources)
│   ├── psyche_workspace.py           workspace: competition + pruning + bounded broadcast
│   ├── psyche_perceive.py            perception + the closed end-to-end cycle
│   ├── psyche_bodysensor.py          body-state sensor + default-off decay modulation (2c)
│   └── psyche_credit.py              temporal credit assignment + depression path (2f-i, 2g)
├── tests/                  ← the test suites (built-in harness; run as scripts)
│   ├── test_psyche_*.py              builder suites (one per module)
│   └── test_tester_*.py              independent tester-authored adversarial suites
├── experiments/            ← Tier 3 disciplined-exploratory scaffolding (its data-free mechanism suites are in run_tests.py; the data-dependent result scripts are not)
└── docs/
    ├── ARCHITECTURE.md         as-built technical reference
    ├── METHODOLOGY.md          the adversarial-testing development history
    ├── TEST_OUTPUT.txt         verbatim captured test run (the source of the numbers above)
    ├── psyche_roadmap.md       the living roadmap (current state, tiers, next step)
    ├── specs/                  ADRs / specs the architect writes for the builder
    ├── design/                 the original *aspirational* vision docs (see its README)
    └── diagrams/               Mermaid (.mmd) diagrams

The tree shows the repository before D-031, when this text was README.md.

---

Background (conceptual inspiration, not implementation)

Psyche borrows vocabulary and structure from the following bodies of work. These are inspirations that shaped the design’s framing; Psyche does not implement, validate, or reproduce any of these theories or their biological results, and makes no claim of biological fidelity.

• Complementary Learning Systems (McClelland, O’Reilly): the fast/slow, experience-then-consolidate split that motivates separating activation (what lights up now) from consolidation (what later gets reinforced).
• Synaptic consolidation research: the framing behind treating reinforcement as a gated, budgeted event, which is why spreading activation here is a pure read and never strengthens what it touches.
• Global Workspace Theory (Baars, Dehaene): the source of the coalitions compete, one is broadcast vocabulary. Coalition formation, a one-winner competition, and a single bounded broadcast back onto the substrate now all exist as mechanism — but this borrows the framing only. No claim of conscious access, ignition, or global availability in the theory’s sense is made; the broadcast is an internal, bounded conduction bias, nothing more.
• LIDA cognitive cycle: the perceive → attend → broadcast staging the end-to-end cycle is loosely shaped after. The cycle now closes (the broadcast biases the next pass), but the resemblance is structural framing, not an implementation or validation of the LIDA model.
• Yerkes–Dodson (arousal / inverted-U): appears in the design notes as motivation for emotionally-modulated plasticity; in the code this shows up only as charge-modulated strengthening, and is best read as design-level aspiration rather than a realized model.

The correct verb throughout is “inspired by” / “borrows the framing of”: never “implements” or “models” in any validated sense.

---

Limitations & future work

• Cognitive meaningfulness above the substrate is unestablished: The field, clustering, and workspace layers are deterministic, pure, and tester-verified as mechanism, but there is no evidence yet that the “thought” they produce is cognitively meaningful. Closing the loop is a mechanism result, not a cognitive one.
• The over-swallow known limitation is resolved at the mechanism level — with a residual corner case shipped as an asserted-negative. Source-clustering groups nearby perceived sources before basins are built, and the workspace prunes the bridged unperceived domain from the winner (the old “incoherent thought on clustered input” is gone, verified through the live cycle). The pruning rule is a parameter-free cohesion-valley / relative-attachment watershed, not literal domain-membership: an unnamed domain bridged into the named region at least as strongly as its own internal cohesion has no valley and is therefore kept, so over-swallow can recur in that specific topology. This divergence is always safe-direction (it over-includes an unnamed domain; it never over-prunes a perceived member) and is documented as a standing asserted-negative test rather than hidden.
• The “strongest coalition = the thought” placeholder is gone. It has been replaced by a real workspace competition (one winner, by activation density with a min-size guard), pruning, and a single bounded broadcast that biases the next cycle.
• Several constants are uncalibrated guesses on probation: notably the closed-road breakthrough threshold/leak fraction, the coalition flow-cut fraction, the source-clustering proximity threshold, and the broadcast gain (it inherits strengthen()’s default and is not tuned to any cognitive outcome). The tests pin the qualitative regime, never the magnitudes, and nothing downstream may depend on these numbers.
• Perception is exact/substring token matching: no embeddings, no fuzzy semantics. It is the minimal step that makes activation sources input-derived rather than hand-placed.
• Backward spread is not implemented: activation follows directed out-roads only (v1).

The first behavioral cognitive claim is still open, and three attempts have been refused — each before any graded number existed, each for a different structural reason.

• Semantic priming (1e) — CIRCULAR. The substrate encodes SWOW association strength as the edge weight, so “related targets activate more” would have been the weight map read back out. The weight-blind leverage stratum the design needed populated at N = 7 against a floor of 30, and the fallback margin was arithmetically unreachable. Frozen NO-GO.
• Fan effect (1e′) — MECHANISM ABSENT. The engine gives every out-road the source’s full activation, with no division across out-edges and no per-source budget, so the fan response is flat to the last digit. There is no fan-dependent signal to grade. Frozen NO-GO.
• Interoception (2c) — SIGNAL FLAT. The body-state signal is a stall fraction, and a realistic think() is CPU-bound, so it sits at ≈ 0 and is scale-invariant. Any gradable band would have to be manufactured, which is forbidden. Recorded NOT-YET; the mechanism is tester-verified and is not the defect.

The pattern is itself the finding: the substrate and the cycle are real and verified, and the recurring hard problem is finding an external yardstick the mechanism can both produce and be graded against without circularity. That is the anti-circularity discipline working — three unwinnable graded runs caught before any of them produced an uninterpretable “pass” — and it is scoped to this substrate and these three yardsticks, not offered as a general claim about spreading-activation models.

---

About this page

This page was the repository README until 2026-10-23. It is kept intact apart from these edits: the title names the stage; the opening banner is new, and this note moved here from it; the section “What this is” is reworded for the project’s stage structure, with its claims unchanged; links are repointed for the new location; wording is edited under the public writing rules; figures that had gone stale or were wrong are corrected (the source-module count, now matching src/; the suite count, now matching tools/ledger.json; the file total that follows from those two; and the BUG-0b1 flip count and interpreter, now matching the BUG-0b1 record section 1); one module is added to the project tree and one wrong tree label is corrected; and a later-record note follows the open defect, with another note under the project tree; and the word “proven” is replaced by “verified” in prose, with the ledger’s status name PROVEN explained where the tiers are set out.

---

License

MIT: see LICENSE. Copyright (c) 2026 Cagan Caliskan.