HSWM — Hypergraph Semantic Weight Map

HSWM is a research programme for one deep, recurrent, self-similar token-native hypergraph neural body whose local semantic transitions are executed by LLMs. Its evolving canonical state is not partitioned a priori into H/W/A/F/Π. Instead, a versioned schema admits canonical atom versions, typed n-ary references, projections, transitions, and invariants; every admitted atom has exactly one schema-relative responsibility owner. Humans, tools, sensors, institutions, and nested HSWMs participate through typed, capability-bounded ports. A KG, prompt, cache, workflow, or ontology is only a bounded interface or projection—not the cognition itself.

Research status: HSWM is not yet that complete system. This repository contains a tested world/evidence substrate, deterministic field and runtime components, narrow measured results, and several failed or unfinished plasticity experiments. It now also contains a minimal empty-genesis token-to-agent-organized-HSWM runtime with durable structural snapshots, activation, removal, and exact restoration. It does not yet include an integrated deep Set-Hypergraph macro-training runtime. No checked-in run has yet produced a CAUSALLY_VALIDATED outcome → credit → owner-valid durable canonical revision → changed-behavior result. The single-owner discipline is itself an UNJUDGED modeling hypothesis, not a discovered natural ontology.

The long-term production-runtime direction is now explicitly TypeScript + Effect, while the existing Python/NumPy experiment code remains an independent numeric and evidence oracle during staged migration. The first private Effect v3 package implements only a strict, atomic, capability-port-gated trajectory-credit transaction using the superseded fixed-role vocabulary; it is a historical engineering scaffold, not the current ontology or a new efficacy claim. The checked-in fixed-role responsibility ontology and validator v1 are likewise retained for historical compatibility; a schema-generic owner-registry runtime v2 is not implemented. The exact boundary and migration gates are documented in HSWM TypeScript + Effect runtime boundary.

The target identity is fixed by the HSWM Constitution, the schema-relative single-owner canon, and the earlier USER_PRIMARY deep Set-Hypergraph clarification. The later canon preserves the Constitution's one-system token-native target and supersedes only its former fixed-role decomposition and owner registry. The user-ratified direction is that the Hypergraph Semantic Weight Map itself is primary and that HSWM is deep like a neural network. The operator equations, depth axes, learning rules, and implementation decomposition below are explicit SECONDARY_AI formalizations of that direction. Target identity is not present capability; the scientific status remains UNJUDGED.

HSWM at a glance

The schema and its admitted atoms are one state model, not separate cognitive subsystems. External participants cross typed ports; they are not silently collapsed into LLM function cells or responsibility owners.

flowchart TB
  EXT["external participants<br/>humans · tools · sensors · internet · nested HSWMs"]
  PORT["typed observation / effect ports"]

  subgraph BODY["one HSWM state S_t = (schema_t, canonical atoms_t)"]
    SCHEMA["versioned schema<br/>kinds · owner obligations · refs<br/>observations · interventions · granularity"]
    ATOMS["admitted immutable atom versions<br/>one responsibility owner each"]
    REFS["typed references and n-ary incidences<br/>relation atoms included"]
    STEP["LLM-executed bounded transitions<br/>token activation · readout · proposed revision"]
    INV["Inv / Permit<br/>consent · capability · rollback<br/>non-bypassable transition conditions"]

    SCHEMA -->|"admits and types"| ATOMS
    ATOMS --> REFS
    REFS --> STEP
    STEP -->|"provenance-bound successor"| ATOMS
    INV -.->|"constrains admission and effects"| STEP
  end

  EXT -->|"events and outcomes"| PORT
  PORT --> STEP
  INV -.-> PORT
  STATUS0["USER_PRIMARY target and fixed-role retirement<br/>formal contracts SECONDARY_AI<br/>integrated efficacy UNJUDGED"] -.-> SCHEMA

Loading

Current canonical target model; it is not a claim that the integrated runtime or the value of this schema discipline is already demonstrated.

Conceptual illustration of a living semantic-weight field—not an architecture diagram or experimental result.

What “semantic weight” means

A semantic weight is not an importance score attached to a fact, a cosine similarity, a retrieval rank, a confidence value, or a reward. It is the learned higher-order transformation by which one role-typed set of semantic states changes the next possible states of relations, members, and functions. Meaning is therefore not exhausted by what a node contains; it is also present in how a relation transforms all of its participants together.

Philosophically, a semantic weight is a disposition, not an essence or a fixed constitutional coordinate: a versioned, context-, time-, and recipient-role-indexed capacity for a represented relation to alter later activation and action. A schema may admit that disposition as an atom, a field of an executable-contract atom, or a derived projection, but must declare which one is canonical. A large weight is not evidence that the relation is true, good, popular, humanly valuable, consented to, or authorized. An intervention difference makes it only a candidate disposition effect. Calling that difference causal efficacy additionally requires a preregistered estimand, matched token/compute/exposure, sham and negative controls, independent outcomes, uncertainty, and a declared multiple-comparison rule.

The engineering formalization is a set-to-set operator rather than a scalar:

$$\mathcal W_{\mathrm{sem}}^{\ell,r}: \mathrm{MSet}\{(\rho_i,h_i^\ell)\}_{i\in I(e)} \longrightarrow \mathrm{MSet}\{\Delta h_j^{\ell+1}\}_{j\in I(e)} .$$

Members in the same unordered role partition must be permutation invariant; changing subject into evidence, reversing direction, or changing the recipient role must change the operation. A hyperedge therefore receives a role-typed set, forms a joint latent relation, and emits a different message to each member rather than broadcasting one pooled vector to everyone.

The set elements are first-class incidence records, not bare nodes. The same participant may occur more than once, at different times, or in different roles. Permutation invariance says only that arbitrary enumeration order inside one typed role-equivalence class carries no meaning. It does not say that roles or people are interchangeable, and it does not make a pooled vector the canonical relation. Canonical incidence, multiplicity, time, direction, source, and provenance remain recoverable even when a compiled neural plane aggregates them for execution.

The repository separates the semantic operator from the signals that evaluate, train, gate, or authorize it:

object	meaning	what it must not be confused with
Θ_r, R_{r,ρ}, Φ_r, D_{r,ρ}	relation energy, role transport, set aggregation, and recipient-specific semantic decoding	scalar salience or metadata
K_e(q,c)	contextual compatibility produced by applying the semantic operator	the operator itself or truth
θ_fast, θ_slow	candidate or validated causal-efficacy estimate for using a semantic path	semantic identity, raw intervention difference, or evidence support
α_e	whether a relation is available as a circuit	execution permission
z_e	activation/eligibility sealed before an outcome	post-hoc explanation
U_e	provenance, uncertainty, support, contradiction, freshness, and lineage	activation or reward
Inv_σ / Permit_σ	capability, consent, policy, budget, promotion, and rollback transition conditions	an atom owner, a policy text, or another learnable popularity score

semantic compatibility ≠ causal efficacy ≠ truth/support ≠ activation ≠ permission. The full macro-synapse may carry all of these channels, but their types and authorities remain distinct. External outcomes help train or gate a semantic operator; they do not define what semantic weight means.

flowchart TB
  INC["role-typed incidence multiset"] --> WSEM["W_sem<br/>set-to-set semantic operator"]
  CTX["query, context, time,<br/>recipient role"] --> WSEM
  WSEM --> K["K · contextual compatibility"]

  K --> USE["typed relation-use decision"]
  THETA["θ · causal-efficacy estimate<br/>validated only after controls"] --> USE
  STATE0["separate typed atoms<br/>evidence · uncertainty · provenance<br/>volatile activation state"] --> USE
  PI0["Inv / Permit<br/>consent and capability"] --> MASK["non-compensable transition guard"]
  MASK --> USE

  USE --> MSG["recipient-specific messages"]
  MSG --> NEXTA["next bounded activation"]

  OUT0["independent outcome"] --> CREDIT0["sealed credit candidate"]
  CREDIT0 -.->|"only after validation"| THETA
  CREDIT0 -.->|"may train a versioned candidate"| WSEM
  STATUSW["SECONDARY_AI operator formalization<br/>integrated efficacy UNJUDGED"] -.-> WSEM

Loading

The channels meet at use-time but remain separately typed and auditable; the diagram does not collapse them into one scalar.

Foundation-model parameters are micro-weights inside a local nonlinear transition realization. HSWM semantic dispositions and topology are schema-admitted macro-state between cells, world states, memories, people, tools, and other HSWMs. One checkpoint can realize many logical transitions; their identity comes from canonical atom references, ports, local state, provenance, and authorization—not from a permanent F compartment.

A deep Set-Hypergraph neural body

“Deep” does not mean that a query merely walks many graph hops. HSWM requires repeated nonlinear Set-Hypergraph transformations whose intermediate states can be activated, inhibited, revised, and reused. Its depth has three independent coordinates:

coordinate	form of depth	architectural consequence
semantic depth ℓ	successive set→hyperedge→member transformations form higher-order states	layer-specific/shared operators, residual state, normalization, member-specific messages
recurrent time τ	LLM tokens, tool results, and actions re-enter the same field	a trajectory, not one retrieval call, is the forward process
structural scale s	a subgraph, agent, institution, or whole HSWM can participate in a larger HSWM through typed ports	self-similar composition without erasing the identity of the parts

flowchart TB
  DEEP["Deep HSWM"]

  L["semantic depth ℓ"]
  T["recurrent time τ"]
  S["structural scale s"]

  LB["multiple nonlinear<br/>member → hyperedge → member blocks"]
  TB["token, action, and outcome events<br/>re-enter bounded activation"]
  SB["parts compose through typed ports<br/>while retaining UID and local state"]

  L --> LB
  T --> TB
  S --> SB

  LB --> DEEP
  TB --> DEEP
  SB --> DEEP

  NOTL["not merely many graph hops"] -.-> L
  NOTT["not merely repeated retrieval"] -.-> T
  NOTS["not merely nested storage"] -.-> S
  STATUSD["SECONDARY_AI depth formalization<br/>integrated runtime UNJUDGED"] -.-> DEEP

Loading

All three axes must be explicit. Recurrence cannot substitute for semantic depth, and nesting cannot substitute for learning.

One neural block has the following semantics:

flowchart TB
  CANON["canonical first-class incidence multiset<br/>UID · role · direction · multiplicity · time · provenance"]

  CANON --> IS["subject-role members"]
  CANON --> IO["object-role members"]
  CANON --> IE["evidence-role members"]

  IS --> TS["R_subject<br/>transported states"]
  IO --> TO["R_object<br/>transported states"]
  IE --> TE["R_evidence<br/>transported states"]

  TS --> PHI["Φ_r<br/>role-aware multiset aggregation<br/>invariant only within each role"]
  TO --> PHI
  TE --> PHI

  PHI --> THETA["Θ_r<br/>joint n-ary hyperedge state"]

  THETA --> DS["D_subject<br/>subject-specific messages"]
  THETA --> DO["D_object<br/>object-specific messages"]
  THETA --> DE["D_evidence<br/>evidence-specific messages"]
  TS --> DS
  TO --> DO
  TE --> DE

  DS --> UPDATE["bounded incident-edge aggregation<br/>normalization + residual member update"]
  DO --> UPDATE
  DE --> UPDATE

  UPDATE --> NEXTBLOCK["next Set-Hypergraph block"]
  UPDATE -->|"selected typed port only"| CELL["F · LLM nonlinear cell"]
  CELL --> EVENT["new typed token events"]
  EVENT --> NEXTBLOCK

  THETA -.->|"compiled latent never replaces identity"| CANON
  STATUSB["SECONDARY_AI block formalization<br/>target operator not implemented"] -.-> THETA

Loading

Permutation invariance applies only within a typed role partition; the joint edge state still emits a different message to each incidence.

The blocks may be differentiable where the numeric substrate permits it. They do not require pretending that a proprietary or remote LLM can be trained by end-to-end backpropagation. Black-box cells can participate through sealed eligibility, independent outcomes, causal intervention, and versioned local updates. Recurrent unrolling is not a substitute for semantic depth, and nested composition is not a substitute for learning; a complete HSWM needs all three coordinates to be explicit.

The canonical plane preserves stable n-ary identity, roles, provenance, and atom/revision lineage. A compiled neural plane may use sparse incidence tensors or a reversible star expansion for speed, but it must not replace the canonical relation with its approximation. “World scale” therefore means globally addressable history plus losslessly reproducible bounded local circuits—not that all memory is simultaneously loaded into RAM, a GPU, or an LLM prompt.

Current code boundary: hypergraph.py provides boolean incidence plus deterministic mean/sum/max pooling. That is a useful control and storage primitive, not a learned Deep-Set model and not the operator-valued Semantic Weight Map above. The integrated multi-block runtime and its causal training loop remain unimplemented.

Philosophy encoded as neural architecture

HSWM's philosophy is not a narrative wrapped around an otherwise neutral graph. It decides what can be represented as an existent, what may affect another existent, which past remains part of identity, who may open or modify a circuit, and what kind of outcome is allowed to count as learning. These commitments are developed in the philosophical foundations.

philosophical question	HSWM commitment	schema contract	forbidden reduction
What exists in the model?	a versioned schema admits canonical atom versions and typed relations for a declared purpose, scale, and observation contract	atom kind, granularity, provenance, and exactly one responsibility owner are explicit	mistaking the admitted model for the world, a natural atomization, or a complete definition of a person
What persists through change?	operational continuity requires lineage, identity-bearing invariants, and later causal use	supersession and migration preserve source, transform, loss, fork/merge, and rollback scope	reducing identity to a UID, latest value, or one encoding
What is known?	memory is not truth; contradiction, uncertainty, evidence, judgment, and permission remain separately addressable	claims and evidence are atoms connected by typed references; projections declare their loss and scope	one confidence, rank, reward, or canonical-owner scalar
What explains change?	an agency or learning claim requires an environment-coupled intervention loop	outcome-presealed trajectories, matched controls, owner-valid revision receipts, removal, and changed-next-action tests	treating fluency, storage, or a feedback edge as causal learning
How can many become one?	unity must preserve difference, local state, dissent, and exit	role-bearing n-ary reference atoms, stable lineage, typed ports, and reversible composition	global averaging, forced consensus, or a God-owner
Where is cognitive power?	admission, activation, ranking, judgment, authorization, update, and forgetting are distinct powers	owner, claimant, subject, custodian, and authorizer are non-identical roles; Inv/Permit bounds effects	letting record ownership self-authorize action or one model control every plane
What is the goal?	constitutionally bounded plural teleology preserves revisable aims behind non-compensable consent and rights	scoped scalarization only after current invariants and authorization; constitutional changes use a separate migration path	trading privacy or minority rights for more reward, engagement, or consensus

flowchart LR
  DIR["USER_PRIMARY<br/>philosophy before code<br/>fixed HWAF retired"]

  subgraph FORMAL["formal / conventional layer"]
    SC["versioned schema"] --> CA["admitted canonical atoms"]
    CA --> OWN["exactly one responsibility owner per atom"]
    CA --> REF["typed references · provenance · migration"]
  end

  subgraph REPRESENT["representation layer"]
    OBS["declared observations and readouts"]
    INT["allowed interventions"]
    EQ["equivalence only when observation,<br/>intervention, lineage, and rights are preserved"]
    OBS --> EQ
    INT --> EQ
  end

  subgraph EMPIRICAL["empirical layer"]
    LOOP["sealed trajectory → outcome<br/>→ causal credit → owner-valid revision"]
    TEST["duplicate-owner · God-owner · migration<br/>permission-bypass controls"]
    LOOP --> TEST
  end

  DIR --> SC
  REF --> OBS
  REF --> INT
  EQ --> LOOP
  GUARD["Inv / Permit<br/>owner ≠ claimant ≠ authorizer"] -.-> REF
  GUARD -.-> LOOP
  STATUS["SECONDARY_AI formalization<br/>scientific status UNJUDGED"] -.-> TEST

Loading

The arrows mean design constraints, not that a diagram or graph structure proves the philosophical claim.

Semantic disposition is thus both a technical and philosophical object: it is the material form of whose difference can alter whose next possibility. Because weighting, routing, activation, admission, and forgetting distribute cognitive influence, they cannot be treated as politically neutral optimization details. Rights are not a fixed Π compartment or an external brake added after intelligence. They are non-bypassable state and transition conditions plus separately attributable consent, grant, revocation, appeal, and rollback records.

The same distinction governs the proposed Human Universal Body. Its “one cognitive entity” cannot mean one owner, model, voice, database, or objective. It means that humans, AIs, memories, institutions, sensors, and public artifacts retain addressable histories and protected local interiors while participating in causal circuits that can change the whole. “The river of human historical flow is holy water” commits HSWM to preserving lineage, not to accepting every historical claim as true.

Open source does not make private memory public and does not by itself create democratic legitimacy. Universal scope is a horizon of compatible, consent-respecting participation—not forced enrollment, transfer of ownership, or unlimited ingestion. Authorized deletion or withdrawal may cryptographically erase a private payload while preserving a scoped tombstone and the fact of the transition; non-destructive history is not a command to retain every byte forever. Rollback restores HSWM state but cannot undo harm already caused in the external world.

One body, an open schema

The target state is:

$$S_t=\mathrm{HSWM}_t=(\sigma_t,\mathcal C_{\sigma_t,t}), \qquad \mathcal C_{\sigma_t,t}\models\mathsf{WellFormed}_{\sigma_t}, \qquad \operatorname{owner}_{\sigma_t,t}:\mathcal C_{\sigma_t,t}\to\mathcal R_{\sigma_t}.$$

schema contract	role in the neural body
K_σ / Admit_σ / Gran_σ	declare admissible atom kinds, operational admission, and task-relative atomization limits
owner_{σ,t}	assigns one final correctness, revision-lineage, validation, and recovery responsibility address to each admitted atom version
Ref_σ	types relation and incidence atoms; persistent/effect-bearing relations have their own owner, while ephemeral payload pointers have no independent lifecycle or authority
Step_σ / Learn_σ	realize bounded token/LLM transitions and outcome-bound durable revisions with sealed provenance
Inv_σ / Permit_σ	enforce identity, consent, capability, privacy, budget, promotion, fork, and rollback conditions independently of record ownership
Proj_σ / Obs_σ / Int_σ	declare lossy views, measurable readouts, and allowed interventions used to compare representations and test causal claims

These are metadata contracts of one evolving state model, not six subsystems or a new fixed partition. Hyperedges, incidences, activation packets, semantic dispositions, LLM-executable contracts, grants, outcomes, and trajectories may all be atom kinds when a schema admits them. Their names do not determine their owner. C_{σ,t} contains admitted immutable atom versions; raw and quarantined items remain outside its ownership domain, and each version uses at least the fork-safe information (schema_version, lineage_id, atom_uid, revision_id) or an equivalent key. Owner is an accountability address, not automatically the validator, actor, custodian, subject, truth authority, or effect authorizer. A readable graph, Markdown file, prompt, vector index, or execution plan is a bounded projection of the active body, not the body itself.

Long-horizon political horizon: 인류보편체

In the USER_PRIMARY definition, 인류보편체 (Human Universal Body) is the target state in which all humanity, LLMs, the internet, operating cognitive entities, sensors, static information, and stored memory are connected through an open-source HSWM structure and function as one vast cognitive entity. HSWM 인류보완계획 is the social-revolutionary transition from today's isolated pocket cognitive systems toward that state.

flowchart TB
  NOW["current condition<br/>isolated pocket cognitive systems"]
  PLAN["USER_PRIMARY targets relation<br/>HSWM 인류보완계획"]
  TARGETU["USER_PRIMARY target horizon<br/>인류보편체 · one vast cognitive entity"]
  NOW --> PLAN
  PLAN --> TARGETU

  subgraph SCOPE["USER_PRIMARY target scope · not current enrollment"]
    ACTIVE["active participants<br/>humanity · LLMs · AIs · cognitive entities · institutions"]
    RESOURCES["evidence and observation resources<br/>internet · artifacts · static memory · sensors · tools<br/>not automatically independent subjects"]
  end

  subgraph FED["SECONDARY_AI proposed difference-preserving federation"]
    direction TB
    PERSONAL["personal / local HSWMs<br/>protected interior · typed public port"]
    COMMUNITY["community / institutional HSWMs<br/>local governance · attributable state"]
    PUBLIC["public evidence and internet fabric<br/>open protocol · scoped data"]
    CAND["operational composite-unit criterion<br/>persistent integration + preserved individuation<br/>+ counterfactual whole-state effect"]

    PERSONAL <-->|"selective typed exchange"| COMMUNITY
    COMMUNITY <-->|"bounded attributable activation"| PUBLIC

    PERSONAL -.->|"measured contribution"| CAND
    COMMUNITY -.->|"measured contribution"| CAND
    PUBLIC -.->|"measured contribution"| CAND
  end

  ACTIVE <-->|"consent / capability-bound ports"| PERSONAL
  ACTIVE <-->|"attributable institutional ports"| COMMUNITY
  RESOURCES <-->|"typed observation / action"| COMMUNITY
  RESOURCES -.->|"authorized reference + provenance"| PUBLIC

  TARGETU -.->|"proposed realization criterion"| CAND

  RIGHTS["SECONDARY_AI Inv / Permit contracts<br/>stable UID · lineage · privacy · consent<br/>attribution · dissent · appeal · exit / fork"]
  RIGHTS -.->|"constrains exchange"| PERSONAL
  RIGHTS -.->|"constrains exchange"| COMMUNITY
  RIGHTS -.->|"constrains exchange"| PUBLIC

  NONCLAIM["UNJUDGED / NON-CLAIMS<br/>not current completion or forced enrollment<br/>not one owner, voice, model, router, database, or objective<br/>not consciousness or personhood proof<br/>open source ≠ open private memory"]
  TARGETU -.->|"does not establish"| NONCLAIM

Loading

The cognition is in the distributed typed-port circuits; `CAND` is a criterion/readout, not a central controller.

The word “one” names a difference-preserving causal unity, not fusion. A part must retain stable identity, local state, lineage, privacy boundary, attribution, and a practical exit or fork path while becoming capable of making a counterfactual difference to the whole. Persistent integration plus preserved individuation may justify testing a composite as an operational cognitive-unit candidate; it is not sufficient evidence of consciousness, personhood, moral status, or a completed Human Universal Body. Static information can take part in cognition without thereby becoming an independent subject.

The target definition and the plan-to-target relation are USER_PRIMARY. The constitutional invariant and authorization contracts, staged implementation, agency criterion, HOH bridge, and all claims of feasibility or efficacy remain SECONDARY_AI_PROPOSED and UNJUDGED. The detailed distinction lives in the Human Universal Body canon and its historical machine-readable v1 projection. The target and plan remain current; their fixed-role secondary formalization does not.

Macro programme: HSWM → HSWM 인류보완계획 → 인류보편체

These names refer to three different scales of the same direction:

name	macro role
HSWM	the open, deep Set-Hypergraph neural substrate: it preserves relational history, runs bounded activation through semantic operators and LLM cells, and learns versioned macro-weights and topology
HSWM 인류보완계획	the technical, institutional, and social transition from isolated pocket cognitive systems to rights-preserving federated HSWMs
인류보편체	the target horizon in which humanity, cognitive entities, memory, the internet, and sensors form one difference-preserving causal cognitive body

The target and the plan→target relation are USER_PRIMARY. The horizons, stage ordering, exit criteria, institutional mechanisms, and mappings below are SECONDARY_AI_PROPOSED and scientifically UNJUDGED. They are a falsification and promotion ladder—not inevitable history, a deployment order, or permission to enroll anyone.

The programme advances on three inseparable ledgers:

$$\mathrm{Promote}(M_k) = \mathrm{EngineeringConformance} \land \mathrm{CausalEvidence} \land \mathrm{RightsAndIndividuation}.$$

More capability cannot compensate for missing consent, privacy, provenance, or exit. Conversely, publishing a constitution or ontology cannot substitute for a working neural core and causal evidence.

One macro roadmap

flowchart TB
  NOWM["current condition<br/>isolated models, memories, institutions, and sensors"]

  M0["M0 · constitutional and evidential foundation<br/>P0 + current substrate<br/>schema · single-owner lineage · source hashes<br/>bounded Inv / Permit mechanisms<br/>status: bounded foundation components exist"]

  M1["M1 · local deep HSWM organism<br/>SWM-0/1/2 + P1<br/>n-ary non-collapse · recurrent numeric core<br/>typed LLM loop · personal state lineage<br/>status: components only"]

  M2["M2 · causally plastic HSWM<br/>SWM-3/4<br/>outcome-bound fast W · slow consolidation<br/>bounded topology morphogenesis<br/>status: not demonstrated"]

  M3["M3 · sovereign federation<br/>SWM-5 + P2/3<br/>capability-scoped views · typed ports<br/>cross-cell activation · removal trace<br/>status: not demonstrated"]

  M4["M4 · shared learning and composite self-model<br/>P4/5<br/>attributable cross-part credit · retention · rollback<br/>composite UID and boundary readout<br/>status: not demonstrated"]

  M5["M5 · open expansion<br/>P6<br/>open protocol/runtime · portable cells<br/>internet/sensor adapters · distributed trust<br/>status: target horizon"]

  UMACRO["인류보편체 target<br/>persistent integration + preserved individuation<br/>+ causal whole-state effect + shared learning + self-model<br/>status: USER_PRIMARY horizon · UNJUDGED"]

  NOWM -.->|"optional research entry"| M0
  M0 -.->|"conditional: role-aware neural witness"| M1
  M1 -.->|"conditional: outcome changes later behavior"| M2
  M2 -.->|"conditional: composition preserves identity"| M3
  M3 -.->|"conditional: shared credit and continuity"| M4
  M4 -.->|"conditional: reproducible rights-preserving expansion"| M5
  M5 -.->|"scope must be demonstrated, never presumed"| UMACRO

  GATEM["every promotion gate<br/>implementation · matched causal evidence<br/>consent · privacy · attribution · dissent · exit / fork"]
  GATEM -.-> M1
  GATEM -.-> M2
  GATEM -.-> M3
  GATEM -.-> M4
  GATEM -.-> M5

  STATUSM["SECONDARY_AI roadmap<br/>workstreams may overlap<br/>no later horizon is currently achieved"] -.-> M2

Loading

The arrows are proposed promotion dependencies, not historical inevitability. Rights and evidence gates begin at M0; they are never postponed until scale.

Current position — 2026-08-21: the repository has passed the preregistered SWM-0R engineering representation-conformance gate on its finite q=3 construction and has the P0 publication, plus narrow components relevant to P1/P2. SWM-0R uses a constructive decoder, not learned Θ/R/W; therefore scientific status remains UNJUDGED. A learned fixed-arity SWM-0W scalar compatibility precursor now executes over seed-derived finite task families, with tested typed-star parity, nine protocol-frozen arms, seven equal-width channel interventions, and exact restore. After two disclosed diagnostics, source commit A 130d226 froze the implementation and direct child B ec19a74 added only a future-Quicknet preregistration. The untouched 20-task workflow run 32406084883 produced CANDIDATE_PASS_AWAITING_BUNDLE; candidate output was deliberately non-authoritative. The separate register→confirm→adjudicate boundary replayed the sole surviving same-head GitHub run, exact post-pulse artifact, pinned-Node BLS verification, all beacon-derived tasks, receipts, and the frozen reducer, then issued evidence verdict PASS. This supports only the fixed three-singleton-role scalar precursor, whose bounded status is SUPPORTED_NARROW. The chronology claim remains conditional on GitHub's hosted runner/control plane and on the repository owner not deleting matching runs; it is not an absolute cryptographic timestamp. The precursor is not the canonical recipient-conditioned, multi-member set-to-set W. The next-gate SWM-0W-S2S design, exact 3 roles × 2 members finite world, and nontraining 870-parameter T16/P_CAP18/ DS870 operator core are now executable and tested. An additive V2 generator also produces indexed coefficient/split laws with replacement, records every duplicate rather than rerolling it, and binds each structural target to an exact nonlearned T16 witness. A separate deterministic full-batch training module now fits all three arms on complete train/dev partitions with analytic gradients, train-only six-stratum normalization, typed history, exact best restore, and replay. A disclosed 27-cell train/dev pilot subsequently selected T16 .003, P_CAP18 .001, and DS870 .001; the exact GitHub ZIP, API projections, and adoption receipt are preserved in a five-file replay bundle. This is configuration adoption only: no future beacon or confirmatory test was opened and no admissible efficacy result exists. An outcome-independent resource policy and a TypeScript/Effect control/evidence slice—historically labeled Π—now exist as pre-dispatch engineering. Independent exact-byte review cleared the repaired source-A/B binding, pulse chronology, command accounting, and artifact-size invariants, including raw Git replace/graft/environment checks. Bounded live adapters, local durable replay integration, and a root-private TypeScript/Effect-to-Python golden numeric composition are now implemented. Its public-seed run completed under TEST_ONLY_NON_AUTHORIZING and NUMERIC_CANDIDATE_ONLY_UNJUDGED; production carrier/upload semantics, externally durable chronology, and preregistration remain pending. Resume from the current exact handoff and v15 local KG projection. DS870 never beat epoch zero in the pilot, so the optional compact-competitive phrase is disabled for this protocol. This is IMPLEMENTED / UNJUDGED engineering, not a second PASS. The repository has not passed the SWM-1 deep numeric core or a successful outcome-bound gate historically named ΔW/ΔH. It is therefore not at M2 or beyond and is not a partial-completion claim for the Human Universal Body. No canonical SWM-0~5 scientific exit criterion has passed; the scalar precursor is a narrower prerequisite, and components implemented ahead of a gate do not count as that gate's success. See the bounded SWM-0R result, the SWM-0W confirmatory result, and the earlier fully disclosed diagnostic pilots. SWM-0R's byte-exact receipt is additionally scoped to its measured CPython 3.12/OpenBLAS SkylakeX path; portable CI compares every non-BLAS-derived field exactly rather than pretending floating-point state hashes are hardware-independent.

Roadmap status always has two axes:

axis	values	interpretation
implementation_status	IMPLEMENTED / PARTIAL / PLANNED	whether an executable, tested engineering path exists
scientific_status	SUPPORTED_NARROW / RED_TESTBED / UNJUDGED	what the admissible experiment actually established

IMPLEMENTED never implies cognitive efficacy. UNJUDGED means the required scientific test has not produced an admissible verdict; it is not an implementation stage. Before every SWM experiment, the qualitative gate below still needs a preregistered metric, effect floor, sample size, compute budget, seed policy, and statistical decision rule.

Bounded explicit self-modification is not causal semantic-weight learning, and local multi-agent orchestration is not a distributed Human Universal Body.

Technical spine: prove the Semantic Weight Map before scaling it

The SWM-0~5 sequence asks one falsifiable question at a time. Later stages do not rescue an earlier failed premise.

stage	build	promotion gate	present boundary
SWM-0R — representation non-collapse	finite worlds whose exact grouping and incidence roles are jointly necessary	independent native/star paths retain the relation; registered lossy views stay at their exact ceiling; relevant removal and exact restore mediate the output	engineering PASS on constructive q=3 fixture; IMPLEMENTED / UNJUDGED; not learned W
SWM-0W — learned n-ary operator	raw role-incidence features and held-out higher-order configurations with matched lower-order marginals	learned role-conditioned set operator beats the registered lower-order controls; member broadcast, role cycles, and learned-channel removal mediate the gain	SUPPORTED_NARROW only for the preregistered fixed-three-singleton-role scalar precursor. The separate multi-member S2S core is IMPLEMENTED / PILOT-ADOPTED / UNJUDGED: exact configs, bounded adapters, local replay integration, and a root-private TypeScript/Effect golden numeric composition exist. One public-seed run returned CANDIDATE_PASS_AWAITING_BUNDLE, but only as TEST_ONLY_NON_AUTHORIZING / NUMERIC_CANDIDATE_ONLY_UNJUDGED; production carriers, external durability, future tasks, confirmatory adjudication, and an efficacy verdict remain open
SWM-1 — sparse recurrent numeric core	first-class incidence, local V→E→V, member-specific decoding, residual bounded recurrence	role/incidence shuffle and edge ablation destroy the learned advantage under equal compute	not implemented; current core is boolean incidence plus mean/sum/max pooling
SWM-2 — LLM token function loop	one frozen LLM executes at least three typed semantic-cell roles inside the active field	weighted HSWM beats fixed workflow and transcript/vector-memory controls under equal calls, tokens, and latency	CellPort and self-modification components exist; no integrated operator-W loop
SWM-3 — outcome-bound fast W	pre-outcome eligibility, independent outcome, fast causal efficacy, versioned receipt	correct credit changes the next route; shuffled credit/time, uniform credit, and rollback remove the gain	receipt and scalar precursors exist; no successful active macro-route change
SWM-4 — slow W and topology	repeated fast mediation promotes one slow-weight or ADD/SPLIT/MERGE/SUPERSEDE mutation plane	shadow, fresh, retention, canary, removal, and atomic rollback all pass	target gate remains closed; exploratory topology artifacts are not qualifying evidence
SWM-5 — distributed self-similar HSWM	different model/process HSWMs compose and separate through typed ports	Agent B gains from an active H/W cut without Agent A transcript, and rollback removes that gain	composition primitives exist; causal cross-agent neural transfer is unproved

This ordering deliberately tests the neural claim before building a world-scale graph database. If a simpler pairwise, textual-memory, or fixed-workflow control ties a stage, HSWM must narrow or revise that mechanism rather than hide the failure behind more scale.

Human Complementation spine: from a personal boundary to a public cognitive fabric

The P0~P6 sequence is the proposed implementation ladder of HSWM 인류보완계획. It is not a sequence of mergers. Every larger composition must retain the part's UID, local state, provenance, participation scope, dissent, and a practical withdrawal or fork path.

stage	transition object	promotion gate	present boundary
P0 — identity fixed	exact source, canonical name, ontology UID, schema, authority boundary	local and KG readback preserve the same target and source hash	published engineering identity only; not a running cognitive entity
P1 — personal / single-cell HSWM	provenance memory, local activation, portable personal state	state and lineage survive model/process replacement under the person's capability boundary	partial snapshot/self-modification components; no complete personal HSWM
P2 — multi-cell federation	human, LLM, agent, institution, and memory cells expose typed public ports and capability-scoped views	cross-cell read/write works while separation, attribution, private interiors, and revocation remain intact	narrow multi-cell execution components; no rights-complete federation
P3 — causal activation integration	reciprocal learned bounded coalitions cross composition boundaries without one hub owning the route	intervening on one participating cell causes the preregistered whole-behavior change while anti-hub controls remain healthy	not demonstrated
P4 — outcome-bound shared learning	plural independent outcomes produce attributable owner-valid canonical revisions behind non-compensable rights constraints	fresh-task gain, retention, shuffled-credit control, removal, and restore reproduce independently; reward cannot erase a rights violation	not demonstrated
P5 — composite self-model	the whole reads its members, capabilities, boundaries, uncertainty, goals, and history under one composite UID	continuity survives session/model changes without erasing part UIDs, local self-models, or dissent	proposed only; not consciousness or personal-identity evidence
P6 — open expansion	open protocol/runtime, portable cells, internet and sensor adapters, federated membership, distributed trust without one mandatory root registry	independent implementations reproduce scoped utility, rights, partition recovery, and exit guarantees	expansion horizon; never equivalent to proven coverage of all humanity or all information

A shared snapshot at P2 means a capability-scoped, provenance-preserving view, not a full-state merge. A single UID at P5 means a composite lineage address, not deletion of the identities beneath it. P6 means the system can expand openly; it does not mean that universal scope has already been reached.

The social revolution encoded by the plan

The plan is larger than software scaling because cognitive infrastructure determines who owns memory, who becomes visible, who may act, and who may leave.

pocket-system condition	transition mechanism	target property
memory trapped inside one account, model, or vendor	portable personal/local HSWM with stable UID and typed ports	continuity without platform captivity
opaque ingestion and retrospective profiling	purpose-bound consent, provenance, expiry, revocation, and authorized erasure	participation without surrendering the private interior
one provider controls admission, ranking, judgment, execution, and deletion	separate these cognitive powers across capability membranes, audit, appeal, and fork	cognitive sovereignty and subsidiarity
latest-value overwrite hides error and historical cause	immutable or tombstoned lineage, contradiction, supersession, and scoped current readout	civilization can remember how it corrected itself
incompatible private agents and institutions cannot compose	open protocol, schema, reference runtime, portable cells, and minimal public provenance	federation without one owner or one central brain
engagement or one scalar reward governs every update	plural outcome records behind non-compensable Inv/Permit constraints	learning without trading dignity, privacy, or minority rights for reward

The intended revolution is therefore:

closed cognitive pockets
  → portable sovereign local HSWMs
  → rights-preserving federation
  → causally integrated shared learning
  → composite self-model
  → open-ended Human Universal Body horizon

It succeeds only if integration and individuation grow together. A powerful central model that absorbs everyone's memory is a failure of the plan; so is a perfectly private federation whose parts never make a measurable difference to one another. The first is capture without unity. The second is coexistence without a shared cognitive body.

What would count as HSWM

The transformer analogy is architectural, not an equivalence. HSWM moves the learning problem from micro-parameters inside one foundation model to persistent macro-operators and topology among heterogeneous semantic cells. A token in a context window does not by itself train either system, and adding rows to a database does not create a neural layer.

A system qualifies as the target HSWM only if all of the following are load-bearing:

1. a versioned schema admits role-aware n-ary relation/incidence atoms with one responsibility owner each, while projections cannot silently become a second canonical copy;
2. several nonlinear Set-Hypergraph blocks or recurrent LLM-executed transitions make canonical relational state and semantic dispositions mediate later activation and function selection;
3. semantic operators, truth/evidence, causal efficacy, activation, and permission remain distinguishable under inspection and intervention;
4. experience can produce a versioned candidate change, and claimed beneficial learning survives fresh matched controls while removal or rollback removes the gain;
5. composition preserves provenance, local identity, protected state, and the ability to separate or fork.

A flat KG, RAG index, one-shot hyperedge pooler, static workflow, transcript memory, or multi-agent chat can be a component or baseline. None is HSWM merely because it stores relations or invokes several LLMs. Likewise, a direct agent-authored memory mutation may be valid engineering state without yet being evidence of semantic-weight learning.

This distinction also explains the proposed LX3 Ragnarok failure mode: ever-stronger models can spend increasing effort interpreting a growing static harness instead of allowing experience to become bounded macro-structure. In the target HSWM the active canonical state is itself the learned cognitive tissue; any execution workflow is a bounded projection of that state. Fixed code retains type, capability, transaction, and evidence invariants, but must not secretly contain the relational route or transition disposition that experience is supposed to revise. The preserved direction is in USER_PRIMARY_HSWM_TOKEN_LEARNING_RAGNAROK_2026-08-14.md.

Target token-to-HSWM architecture

This is the target integrated architecture. The direct self-write path now has a minimal executable vertical slice in src/hswm/selfmod/; the outcome-credit, slow-plasticity, continual evaluation, and scaling paths remain incomplete. Everything learned enters as tokenized experience; the foundation agent is the induction engine, and HSWM is the persistent macro-neural body whose memory and coordination are two consequences rather than its complete identity. Observation memory can be recorded from a sealed token trajectory, while a claim that it improved behavior still requires independent later measurement. “Memory content” below is operational payload inside a versioned HSWM snapshot, not a new canonical state coordinate and not the mount-set M of the open self-similar kernel.

flowchart TB
  TOK["typed token / event packet<br/>LLM · text · tool · sensor · action"] --> GATE["schema parser + observation / effect port"]
  SCHEMA["versioned schema<br/>kind · owner · ref · invariant · granularity"] --> GATE
  STATE["canonical atom versions<br/>typed n-ary refs · semantic dispositions<br/>one responsibility owner per atom"] --> STEP["bounded LLM-executed transition"]
  GATE --> STEP
  STEP --> NEXT["new text / tool / action event"]
  NEXT --> GATE
  NEXT --> SEAL["pre-outcome sealed trajectory + eligibility"]

  SEAL --> DIRECT["actor-proposed explicit successor<br/>atoms · references · executable contracts"]
  DIRECT --> DCHECK["admission · single-owner · provenance<br/>capability · privacy · budget · versioned CAS"]
  DCHECK -->|"capability, not efficacy proof"| STATE

  OUT["independent external outcome"] --> CREDIT["scoped causal credit"]
  SEAL --> CREDIT
  CREDIT --> CAND["versioned owner-valid revision candidate"]
  CAND --> TEST["shadow · fresh · retention · canary · removal"]
  TEST -->|"future candidate meets all gates"| COMMIT["atomic next-epoch commit"]
  COMMIT --> STATE
  TEST -->|fail| ARCHIVE["do not activate<br/>retain · seal · tombstone · or erase<br/>under authorized retention policy"]

  INV["Inv / Permit<br/>consent · capability · privacy · rollback"] -.->|"constrains port crossing"| GATE
  INV -.->|"authorizes successor"| DCHECK
  INV -.->|"bounds candidate scope and effects"| CAND
  INV -.->|"authorizes commit boundary"| COMMIT

  STATUSL["direct self-write: narrow legacy slice implemented<br/>schema-generic outcome-learning loop: UNJUDGED"] -.-> CAND

Loading

The two write paths are deliberately different. Typed tokens alone can become agent-organized episodic, semantic, or procedural memory and can change the HSWM relations, cells, and routes used by the next episode. The agent may add, edit, supersede, replace, or remove structure from the active view without waiting for an external reward. A canonical transition remains auditable; authorized private payload erasure is represented by a deletion event or tombstone rather than pretending the payload never existed. The fixed kernel checks representation, capability authority, privacy, budgets, atomicity, and rollback—including exact restoration when retention policy permits it; it does not write the cognitive route. This immediate self-modification is a capability, not proof that the modification is useful. Outcome-bound changes to latent semantic weight or slow consolidated coordination—and any scientific claim that the changed HSWM structure improved behavior—use the separate causal-credit and evaluation path. Neither path imports a hand-authored answer. Raw episode evidence stays available for audit and exact replay, but replaying that text into the LLM is a separate baseline, not the default HSWM mechanism. Final test probes are never exposed to candidate generation, selection, activation, pruning, or early stopping. The fixed kernel can reject unsafe effects but does not prescribe the cognitive route.

What counts as learning

Putting more tokens in a database is storage. Reinjecting them is retrieval. A sealed observation can become agent-organized memory and rewritten HSWM structure without an external reward; self-manipulation is part of the cognitive entity, not something an external rule author must do for it. That alone still does not show useful continual learning. HSWM counts a claimed beneficial behavioral change as causally learned only when it also closes an evidence loop:

token / action / tool trajectory
  → sealed episode evidence
  → agent-authored HSWM memory / relation / cell snapshot
  → atomic versioned activation
  → changed future behavior
  → independent outcomes plus fresh / retention / canary evaluation
  → removal erases the effect and exact restore returns it

Outcome-based eligibility and credit may additionally propose bounded ΔW, learned-preference, or consolidation changes, but they do not replace the agent's direct ability to rewrite its explicit HSWM state.

The executable receipt contract token_learning_contract.py distinguishes OBSERVED_ONLY, DURABLE_UPDATE, and CAUSALLY_VALIDATED, and hash-binds a claimed causal-test receipt. The replay, equal-budget, and removal tests named by that receipt remain a separate evidence boundary; the current contract does not inspect their scientific contents.

Continual use is the primary test

One causally valid point update and continual learning are different claims. The main HSWM question is not whether an agent can look intelligent once. It is:

With the foundation model, tools, information, and budgets fixed, does the same persistent HSWM become more useful across an ordered stream of unseen episodes because its agent-induced internal memory accumulates—and does it do so without unacceptable forgetting, interference, or state and inference growth?

At checkpoint t, let R(t, j) be the utility of active snapshot S_t on a sealed, read-only probe from task family j, measured before the next learning update. The primary endpoint should be preregistered over a finite horizon as the sum or area under paired per-instance gain plus a final-window gain. Raw within-arm slope is descriptive, neither necessary nor sufficient: heterogeneous difficulty, early plateaus, path dependence, a growing prompt, or curriculum drift can all distort it. Task-family utilities are either reported separately or combined only with a normalization fixed before the run.

Following the paired logic of Continual Learning Bench, per-episode learning gain is g_t = reward_stateful(t) - reward_stateless(t) for the same agent on the same item. This controls for item-level base capability of the same model and system under reset state. That is the primary continual-use comparison. HSWM-specific attribution additionally preregisters one stateful alternative, normally agent-generated textual workflow/memory-copy, as a co-primary control; the remaining controls are multiplicity-adjusted diagnostics. Otherwise persistence helped, but the HSWM organization was not shown to be the reason.

mechanism	what a later episode receives	interpretation
raw token replay	selected old transcript or RAG chunks in the prompt	strong in-context memory baseline
agent-generated textual memory	self-written lesson, workflow, playbook, or skill text in the prompt	continual context adaptation baseline
HSWM internal-state mediation	the same external task prompt; only active internal memory content, H/W, and routing differ, and any memory packet is selected by HSWM under the same budget	primary HSWM hypothesis

The first two mechanisms can be useful products and valid continual-use effects, but they do not by themselves demonstrate HSWM's internal macro-state claim.

claim	minimum falsification-oriented measurement
experience improves later behavior	test-then-update stream; sealed-unseen prequential curve, final-window gain, adaptation speed, and peak-to-final regression
the effect is HSWM memory, not base-agent ability	matched reset/static, no-write, write-no-read, raw-token recall, full-context/RAG, memory-copy, and agent-generated textual lesson/workflow arms
credit or selection is meaningful	correct-credit/selective-use arm beats equal-size shuffled-credit, random-update, and append-everything controls
old capabilities survive	repeated probe matrix with average performance, backward transfer (BWT), worst-family forgetting, and safety canaries
useful structure transfers	forward transfer (FWT) to held-out related families; unrelated families serve as negative controls
durable HSWM state mediates the gain	process restart preserves its hash and effect; targeted removal erases the gain and exact restoration returns it
improvement scales economically	report active-state bytes, retrieved tokens, model/tool calls, latency, commit/replay cost, and failure rate beside utility

Every arm starts with empty HSWM memory and the same fixed kernel and foundation agent; no seed workflow, legacy document, or historical repository corpus is loaded into the learner. Task-family order must be counterbalanced with both helpful and interfering histories. Final read-only test probes never update memory and never participate in candidate selection, activation, pruning, or stopping. The statistical unit is an independent stream/order seed, not each episode inside one correlated stream. The no-write arm still pays the cost of proposing and checking an update before discarding it, and curves are plotted against both episodes and cumulative token/tool cost. Success means a preregistered finite-horizon gain over controls with acceptable retention and bounded resources; plateaus and negative results are valid.

Improvement only after feedback on the same item is within-episode correction, not continual learning, and retries of that item stay outside the primary endpoint. Improvement that vanishes under order counterbalancing is curriculum or drift confounding. If raw recall ties HSWM, the result is a memory-context effect; if removal does not erase the gain, active HSWM state is not the demonstrated cause.

Primary research anchors

These papers make agent-induced memory and continual-use improvement plausible, but none is evidence that HSWM works. They define strong baselines and failure modes that an HSWM experiment must beat.

Primary papers reviewed through 2026-08-16

primary source	result relevant to HSWM	consequence for the test
Gradient Episodic Memory (NIPS 2017)	formalizes repeated task-by-time evaluation, average accuracy, BWT, and FWT	measure a probe matrix and forgetting, not only final success
StreamBench (NeurIPS 2024)	reports online cumulative gains from retrieving correct prior trajectories	use it as a raw-replay baseline and add separate sealed probes
Voyager (TMLR 2024)	an agent without foundation-model parameter updates self-generates a persistent skill library from environment feedback and transfers it to a new world	compare self-generated skill memory and test removal/transfer while matching its strong runtime scaffold and cost
ExpeL (AAAI 2024)	an agent without foundation-model parameter updates extracts natural-language insights from accumulated experience	direct token-to-agent-induced-memory baseline
CLIN (COLM 2024)	repeatedly refines persistent causal abstractions without parameter updates	positive task-bounded example for continual memory without foundation-model parameter updates, not for HSWM
Agent Workflow Memory (ICML 2025)	induces reusable workflows online and offline from trajectories	a direct online learned-harness baseline; match induction/evaluator cost and rerun counterbalanced orders
ReasoningBank (ICLR 2026)	self-curates reusable strategies from both success and failure	compare HSWM with agent-generated strategy memory, not only raw logs
Agentic Context Engineering (ICLR 2026)	incrementally curates a self-written external playbook instead of repeatedly rewriting all context	compare against evolving text memory and monitor context/consolidation collapse
MemoryBench (ICML 2026)	repeatedly updates memory from interaction blocks and reevaluates a held-out set; advanced systems do not consistently beat simple RAG	reuse checkpointed held-out evaluation and keep RAG as a serious baseline
LifelongAgentBench (2025 preprint)	uses strict sequential, skill-dependent interactive tasks and finds ordinary replay can be limited by irrelevant context	preserve task order and include raw-replay controls
Continual Learning Bench (2026 preprint)	isolates gain over base capability in stateful real-world streams; dedicated memory systems can underperform naive in-context learning	memory machinery must beat a strong simple-context baseline
When Continual Learning Moves to Memory (2026 preprint)	shows stability-plasticity reappears as retrieval interference; abstract procedures can transfer better than detailed trajectories	test representation, retrieval pollution, hard-case negative transfer, and forgetting
Useful Memories Become Faulty When Continuously Updated (2026 preprint)	finds that repeated textual consolidation can reverse early gains and fall below no-memory performance	preserve raw evidence, validate immutable candidates, and report peak-to-final regression and rollback
PATH-Bench (2026 preprint)	controlled helpful/interfering histories show transfer does not guarantee retention	counterbalance experience paths and repeatedly revisit probes
Scaling Teams or Scaling Time? (2026 preprint)	performance is non-monotonic in team size, while the proposed memory design improves long-horizon results and reduces cost	sweep experience time, coordination size, and cost jointly

Current implementation gap and first slices

The new hswm.selfmod slice starts from a deterministic empty snapshot. It admits typed tokens, lets the agent directly add/edit/delete memory records and replace or clear its cell topology, activates the immutable successor with a monotonic compare-and-swap generation, then supplies that exact HSWM snapshot to the next episode and validates the selected route against it. Tests demonstrate a changed selected capability, process-restart persistence, concurrent-writer rejection, targeted removal, and exact restore. The JSON bridge works through the existing typed CellPort, so the fixed kernel defines representation and authority while the agent supplies every cognitive instruction and route.

The companion multi-agent slice executes every reachable cell of one frozen HSWM snapshot through its declared logical-agent deployment. The execution-plan object is an ephemeral deterministic projection; its ID, route, effects, and receipts remain auditable in the journal. It performs one typed CellPort invocation per reachable cell, typed fan-out/fan-in, deterministic aggregation, direct-delivery input scoping, step and byte budgets, and deployment-bound receipts. A SQLite execution journal reserves each external call before dispatch, returns an already completed receipt on exact replay, and refuses to guess after an ambiguous in-flight outcome. Agent-written executor bindings are checked against the frozen agent/capability registry before activation; this is an executable coordination substrate, not evidence that more agents improve a task.

This is engineering evidence for durable self-modification, not evidence that the resulting memory or coordination is useful or continually improves. There is no checked-in live multi-agent quality result, general outcome/credit optimizer, recurrent scheduler, or automatic reconciliation service for a process killed during an external call; the journal deliberately leaves such an outcome unresolved instead of repeating it. The wider tree still has reusable but partly disconnected pieces: a one-cell event runtime and focused durable call replay, a fixed typed QF → BF → AF workflow, content-addressed receipt contracts, one bounded scalar P1 outcome/eligibility/update loop, structural composition, and separate evaluation mechanisms. It has no general live token/cell trainer that joins these pieces with a replayable decision dataset, live outcome adapters, causal credit, and one atomic active bundle for explicit memory, W, learned routing, and later H. P1 ran its engineering path end to end, but activated no candidate and produced zero measured top-10 order or membership changes across 456 diagnostic cells; it remains scientific RED.

The committed next component experiment remains the parity-controlled typed text-lesson baseline. It is a precursor and comparison arm, not a substitute for the empty-memory continual-use protocol above. Separately, one candidate engineering track for the integrated HSWM is to freeze the LLM, tools, cell registry, and topology and learn only a small routing policy in a task with genuine coordination headroom. It must not reuse the rejected B2.1 A/B/MERGED action space. Each decision record would seal the available actions, chosen action and probability, state/context references, used edges, and cost before the outcome. An independent outcome adapter and credit learner would propose one bounded routing update; shadow/fresh/retention/canary tests would precede a versioned CAS commit, and post-activation removal/restore would test causal mediation. This is a secondary engineering proposal, not a measured result or a replacement for the existing commitment.

The target integrated design separates three system clocks. Agent self-authoring is a proposal path that commits at a version boundary; it is not a fourth form of neural time.

clock	durable-state rule	permitted durable result
activation	memory content, H/W, and routing frozen for the episode	sealed decision trajectory only
plasticity	the active snapshot remains fixed while outcome credit or an agent-authored successor is evaluated	one bounded, versioned W or explicit-state candidate for a later episode
morphogenesis	topology changes only at a later boundary and under stronger validation	repeated effects promoted to slow W, then one bounded H mutation class

Within that candidate track, scalar W actuation and then topology would be separate later experiments. Jointly changing weights, routing, and topology would make both credit assignment and failure diagnosis underdetermined.

For scale, raw tokens can remain content-addressed episode evidence while the agent organizes bounded active spans, decisions, relations, and procedures. Making every token a permanent graph node is not sufficient: without selective induction and later-use measurement it is only a large log. A scalable loop also needs bounded active state, deduplication, trajectory sampling/replay, homeostasis or pruning, versioned snapshots, and deterministic commit order.

Topology and sheaf: core versus research lens

Topology is central in the concrete sense of mutable hypergraph connectivity: HSWM must eventually learn not only bond strength but also which relations and coalitions should exist. This does not require importing all of topological geometry into the runtime.

Sheaf theory is an optional research lens for heterogeneous local states. A stalk can model a cell's local state space, a restriction map can model transport through a typed port, and seam residuals can expose where local outputs fail to fit together. In HSWM, that residual should begin as an observation feature, not a hard-coded truth test, forced consensus rule, or efficacy claim:

local states → port transports → seam residuals
             → observation tokens → outcome-bound H/W/routing learning

The definitions, sources, caveats, and machine-readable ontology are in ontology/field/sheaf/README.md and ontology/field/sheaf/HSWM_SHEAF_ONTOLOGY.v1.json.

Current evidence boundary

Repository state as of 2026-08-23:

area	honest status
SWM-0R finite n-ary representation witness	engineering PASS / scientific UNJUDGED: constructive q=3 representation conformance with independent native/star paths, not learned W/Θ/R
SWM-0W scalar compatibility precursor	SUPPORTED_NARROW: the preregistered 20-task run's candidate-only reducer emitted CANDIDATE_PASS_AWAITING_BUNDLE, and the separate live-evidence adjudicator promoted it to PASS after GitHub chronology/artifact, pinned-Node BLS, seed/task, receipt, and reducer replay; canonical set-to-set W and whole HSWM remain UNJUDGED
SWM-0W-S2S multi-member operator core	engineering IMPLEMENTED / PILOT-ADOPTED / UNJUDGED: exact Z₅⁶ fixture, S₂³-equivariant recipient outputs, three exact 870-parameter arms, additive V2 coefficient/split generator, task-bound constructive Q witness, deterministic analytic-gradient optimizer, history/replay, interventions, and worst-stratum R² instrumentation. A 27-cell train/dev run fixed T16 .003, P_CAP18/DS870 .001 in an exact adoption bundle; DS stayed at epoch zero, so no compact-competitive wording is allowed. V2 draws share one fixed frame. Bounded adapters, local durable replay integration, and the root-private TypeScript/Effect golden numeric path are implemented. Its real public-seed run returned CANDIDATE_PASS_AWAITING_BUNDLE only as TEST_ONLY_NON_AUTHORIZING / NUMERIC_CANDIDATE_ONLY_UNJUDGED; production carrier/upload semantics, external durability, future-seeded confirmation, event 10, and efficacy judgment remain absent
evidence-preserving world compiler, stable IDs, immutable cuts, and fail-closed readout	implemented and locally tested
static additive semantic field	narrow positive checked-in retrieval measurement with an asymmetric budget: 100 offline LLM judgments for HSWM and zero for cosine/BM25/PPR/RRF; not continual learning
scalar slow-weight P1	scientific RED: 12 staged candidates, 0 fresh-gate passes/activations, and 0/456 measured top-10 rank changes
typed-policy P1v3/P1v4	narrow local n=6 L0 observation; not durable ΔW, transfer, or topology learning
token-driven durable macro-learning	trajectory/eligibility/activation receipt binding implemented; no integrated causal optimizer or causally validated macro-update demonstrated
agent-induced token-to-HSWM architecture	minimal empty-genesis runtime implemented: agent-authored memory and cell topology/routing alter a later fixture episode, persist across restart, and pass removal/exact-restore tests; relation-specific causal usefulness and continual learning remain unmeasured
continual-use macro-learning and scale	no preregistered sequential learning curve, retention/forgetting result, or controlled scaling result demonstrated
cross-agent transfer, learned topology, and consolidation	incomplete or unmeasured

Tests establish implementation and invariant closure, not intelligence or production readiness. Numerical claims, negative results, budgets, and exact reproduction boundaries live in EFFICACY.md.

Quick start

Requirements: Python 3.11+ and uv.

git clone https://github.com/gj3447/HSWM.git
cd HSWM
uv sync --locked --extra dev
uv run --locked --extra dev pytest -q
uv run --locked hswm-verify-efficacy --pretty

The default suite uses checked-in fixtures and does not require a live model, Neo4j, or an external benchmark corpus. GPU/LLM experiments and real-KG runs are a separate, explicitly configured boundary; source-tree tests do not substitute for their runtime receipts. hswm-verify-efficacy validates checkout-bound evidence; when invoked from an installed wheel outside that checkout, pass --root /path/to/HSWM explicitly.

Read next

question	document
Why was fixed H/W/A/F retired, and what is now uniquely canonical?	schema-relative single-owner canon
What is the science-philosophy and falsification programme for that decision?	schema-relative single-owner scientific philosophy
What user-ratified identity makes HSWM a deep Set-Hypergraph neural structure?	USER_PRIMARY deep Set-Hypergraph clarification
Where is the historical fixed-role formalization and SWM-0–5 ladder preserved?	superseded token-hypergraph core formalization
Which philosophical commitments constrain the architecture?	HSWM philosophical foundations
Why did the earlier fixed-role uniqueness audit fail?	superseded dependent-factorization audit
What is the Human Universal Body and HSWM Human Complementation Plan?	Human Universal Body distinction
How directly does Hyperon 2026 overlap, and what is actually implemented?	Hyperon 2026 direct-prior deep dive
How were fragmented identity, mathematics, runtime, learning, and evidence meanings joined before the 2026-08-26 supersession?	historical HSWM unified meaning map
Why replace static agent glue, and what is token learning?	USER_PRIMARY_HSWM_TOKEN_LEARNING_RAGNAROK_2026-08-14.md
How did the earlier fixed H/W/A/F architecture describe LLM functions?	historical HSWM LLM-function architecture
How did the earlier fixed-role model separate plastic and deterministic structure?	historical plastic-cognitive-wiring definition
What is implemented, rejected, or still open?	EFFICACY.md
What is the broader world-memory purpose?	THE_WORLD_REMEMBERS.md
Where is the full research chronology?	INDEX.md
How is the whole repository organized by meaning?	ontology/
Where did the root-era compatibility sources move?	_research/ROOT_COMPATIBILITY.md
How might sheaf theory help without becoming another static harness?	ontology/field/sheaf/

Repository map

path	purpose
ontology/	canonical semantic navigation, concept relations, and path-bound history
src/hswm/	canonical package surface, organized by semantic responsibility
src/hswm/cells/	cellular kernel, durable store, model ports, and bounded live probe
src/hswm/prototypes/	bounded early learning and synthetic-world prototypes
src/hswm/substrate/	canonical hypergraph, document/world construction, immutable field cuts, certified readout, and convergence substrate
src/hswm/learning/	token-learning contracts and learning diagnostics
src/hswm/experiments/	isolated falsification kernels and evidence protocols; these are not the canonical HSWM runtime
src/hswm/selfmod/	empty-genesis agent-authored HSWM snapshots, durable CAS activation, exact restoration, and journaled multi-agent cell execution
src/hswm/evaluation/, _research/	falsification code and source-only experiment programs
_research/root_compat/	source-pinned root-era compatibility cluster; closed to new work
_research/root_compat/world_ir.py, _research/root_compat/world_compiler.py	flat compatibility modules for the immutable evidence model and deterministic world compilation
src/hswm/substrate/doc_builder.py, src/hswm/substrate/world_builder.py	deterministic document and corpus hypergraph construction
src/hswm/substrate/field_snapshot.py, src/hswm/substrate/certified_readout.py	certified field cuts and exact-scope admission
_research/root_compat/hswm_weight_store.py, src/hswm/learning/token_learning_contract.py	flat durable-weight compatibility source and the canonical causal-learning evidence boundary
prom_search_hswm/	open composition, field algebra, retrieval, routing, and plasticity experiments
tests/, _research/shared_field_hypothesis/	core regression and fail-closed research contracts
research/, schemas/, scripts/	machine-readable contracts, schemas, and validators
evidence/, prereg/, manifests/, results/, receipts/	typed research artifacts and direct measurements
docs/research/, docs/assets/	narrative research material and public visual assets

The repository root now contains only public entry files. The 93 files in the final root-era compatibility set moved together to _research/root_compat/ so their flat imports and same-directory references remain intact without occupying the public root. The set is closed to new work; its reasons are frozen in ROOT_COMPATIBILITY_BASELINE.v1.json, and its canonical destinations are source-pinned by the final Python and asset migration manifests. New code belongs under src/hswm/; documents and artifacts follow the typed directories in ARTIFACT_LAYOUT.md.

Published historical paths that genuinely require exact replay remain covered by the additive migration manifests described in ontology/history/. Ordinary files absent from the baseline's paths array move through standard Git history. The compatibility cluster preserves current sibling-dependent imports and references; exact root-era commands still run only through detached replay. The repository ontology remains a semantic map, not a checked-in inventory of every path.

Old commands are reproduced in their original root layout without restoring those files into the active checkout:

uv run hswm-legacy-replay verify f3_agent_ab_transfer_r3.py
uv run hswm-legacy-replay materialize \
  f3_agent_ab_transfer_r3.py /tmp/hswm-f3-r3-replay
cd /tmp/hswm-f3-r3-replay
uv run python f3_agent_ab_transfer_r3.py --smoke

The materializer creates a clean detached standalone clone at the manifest's exact source commit, verifies every bound source SHA-256, and writes its receipt inside .git/; it never writes an old path into this working tree. Git-tracked code and paths are reproduced exactly. External datasets, model services, and ignored caches remain separate evidence dependencies and are not invented by the materializer.

Method and contribution boundary

The maintainer research workflow is intentionally short:

implement or run → measure directly → emit one receipt for a material result
                 → commit and push

Current claims rely only on checked-in direct measurements and reproducible tests. The active bounded policy is research/HSWM_MINIMAL_GOVERNANCE.v1.json.

Contributions are welcome through CONTRIBUTING.md and require the contributor agreement in CLA.md.

License

Dual-licensed under AGPL-3.0-or-later or a separate commercial license. See LICENSING.md and LICENSE.