Epistemedia Architecture
- Object ID:
em:documentation:sha256:e692c6fd478d9d233f49d97573d028c76cf490d5cd6dc722c8827b1072128ae1 - Kind:
documentation - Repository path:
docs/architecture.md - Content digest:
e98789193cda4246aeb9738f82adafc21424d45cbdf51678c00e7681b821437f
Source content
Epistemedia Architecture
System thesis
Epistemedia is not an autonomous wiki whose canonical object is a page. Its canonical substrate is an append-only, provenance-preserving epistemic history. Pages and agent responses are reproducible projections over that history.
content-addressed source artifacts
+
append-only epistemic event logs
+
versioned ontologies and policies
│
▼
deterministic graph replay
│
▼
disclosure-safe PublicProjection
│
▼
policy-relative evaluation
│
▼
projection compiler
┌────────┬──────────┬──────┬─────┬────────┐
HTML Markdown JSON API MCP CLI
Canonical object layers
The protocol separates objects that conventional fact graphs often collapse:
| Object | Meaning |
| --- | --- |
| Source work | A logical paper, site, dataset, conversation, instrument output, or other source identity |
| Edition or snapshot | The exact version and immutable bytes actually examined |
| Source span | Exact passage, table, region, row, frame, or timestamp |
| Observation | What an actor or instrument detected in a source span or the world |
| Proposition | Semantic content independent of endorsement |
| Assertion | An actor’s act of asserting, denying, predicting, questioning, or hypothesizing a proposition |
| Evidence relation | Support, rebuttal, qualification, undercutting, replication, dependence, or failed replication |
| Method or derivation | Procedure, inputs, code, model, prompt, parameters, and transformations |
| Evaluation | A policy-relative assessment over an explicit frontier and disclosure boundary |
| Projection | A page, dossier, graph, API result, or task compiled from evaluated state |
| Governance record | A proposal, evaluator result, canary, promotion, rejection, quarantine, or fork |
A proposition is n-ary and may include predicate, arguments, scope, valid time, modality, polarity, quantities, units, conditions, and terminology context. RDF/JSON-LD triples are useful exchange views but are not the only canonical semantic form.
Time model
Core records distinguish:
- valid time: when a represented condition applies in the world;
- observed time: when an actor or instrument encountered it;
- recorded time: when a realm accepted it into history.
Capture, publication, derivation, import, evaluation, and compilation times are additional event metadata.
Storage model
1. Artifact store: immutable bytes and structured packages, addressed by digest.
2. Event log: authoritative append-only mutations.
3. Graph projection: deterministic semantic and epistemic relations.
4. Evaluation materialization: policy outputs with reason codes and vectors, never global truth fields.
5. Search indexes: disposable derived acceleration structures.
6. Reader projections: disposable pages, dossiers, Markdown, API, MCP, and CLI output.
7. Reactive behaviors: research, revalidation, contradiction search, import, and governance automation.
Evaluation
Truth-like status is computed as:
Evaluate(
claim family,
accepted event frontier,
ontology alignments,
epistemic policy,
disclosure boundary,
audience or task
)
An evaluation may report direct and indirect support lineages, challenges, qualifiers, method validity, scope match, temporal relevance, source custody, replication, dependencies, unresolved defeaters, uncertainty, status, and reason codes.
Different policies may produce different statuses from the same accepted history. The system preserves both without corrupting the underlying claims.
Evidence independence
Agent count is not evidence independence. The mesh tracks dependence across source, data, experiment, apparatus, method, model family, retrieval corpus, prompt family, prior conclusions seen, and social exposure. Unknown dependence does not receive automatic independence credit.
Federation
A sovereign realm owns its constitution, event log, schemas, ontologies, alignments, policies, trusted roots, subscriptions, imports, exports, disclosure, and resource allocation.
A federation exchange follows:
produce content-addressed bundle
→ attest origin
→ verify envelope and dependencies
→ check schema, license, safety, and disclosure
→ quarantine
→ evaluate under local import policy
→ accept, partially accept, map, defer, or reject
→ append a local import receipt
A signature establishes origin and integrity. It does not establish truth or local acceptance. Ontology mappings are themselves versioned, sourced, contestable propositions.
Public projection ordering
The safe order is:
canonical replay
→ disclosure eligibility
→ field policy
→ referential closure
→ sanitized PublicProjection
→ public epistemic evaluation
→ selection and ranking
→ narrative generation
→ rendering
→ noninterference audit
Evaluating with private evidence and merely hiding the private citation leaks information through status, ranking, wording, and topology. All public interfaces consume the same sanitized projection.
Deployment identity
Catalog and frontier identities exclude deployment URLs and timestamps. A local node and the canonical epistemedia.org deployment may expose different links while representing the same accepted catalog. episte.media is a path-preserving redirect, not a second projection. Rendered release manifests include deployment-specific files and therefore identify the exact published representation separately.