TL;DR
- 01A Context Capsule keeps nine named graphs in one portable object: evidence, claims, situation, temporal, ontology, reasoning, memory, governance, runtime.
- 02The layering is epistemic, not organizational. What a source said, what an analyst concluded, when each became known, and which method produced the judgment are different kinds of statement — collapsing them is how confident nonsense gets made.
- 03Each layer has its own mutability contract: evidence is append-only, claims promote through review, runtime is a contract on consumers.
- 04The payoff is symmetric precision: agents cite at the right level; humans contest at the right level.
What a source said, what an analyst concluded, when each became known, and which method produced the judgment are different kinds of statement. Why capsules keep nine named graphs with distinct mutability contracts instead of one big graph.
Why one graph is not enough
Put everything in one undifferentiated graph and the first complex query destroys you: an edge asserted by a hostile source, an edge inferred by a model, and an edge confirmed by your own reviewer all look identical to traversal. Weighting helps until it doesn’t — a confidence score cannot express “true as testimony, disputed as fact, superseded as of Thursday.”
The fix is old and good: separate the epistemics. Database people learned it as bi-temporality; intelligence analysts as source-versus-assessment discipline; philosophers as the difference between what is said and what is warranted. Capsules encode it as nine named graphs in one JSON-LD document, each with a declared role and mutability contract.
The layers, briefly
- ▸Evidence — sources, spans, hashes, rights. Append-only bedrock; everything above must point into it.
- ▸Claims — atomic typed assertions with confidence, corroboration, trust tier, disputed flag.
- ▸Situation — entities and relations; edges carry temporal scope and review state, because relations are claims too.
- ▸Temporal — episodes, two clocks, Allen-13 interval relations, causal links, supersession chains.
- ▸Ontology — the kernel grammar plus capsule-specific extensions, proposed by models, promoted by humans.
- ▸Reasoning — devices, heuristics, traps, critique prompts: how the expert thinks, typed and attributable.
- ▸Memory — the build history: what was proposed, asked, decided. The capsule remembers its own making.
- ▸Governance — trust tiers, review ledger, sign-off, recorded dissent.
- ▸Runtime — the contract for any consumer: citation rules per tier, must-surface-disputed, composition policy.
Layers × primitives, not layers vs primitives
The nine layers are orthogonal to the eight kernel primitives (actor, claim, interest, constraint, leverage, commitment, event, narrative). Primitives are the vocabulary; layers are the epistemics. A Commitment is a primitive — whether a given statement about it lives in claims (asserted), temporal (scheduled), or governance (reviewed) depends on what kind of statement is being made. This is also why every capsule ships an embedded, read-only graph projection: traversal questions (“what connects the army command to the certification body?”) are answered locally, with Cypher, against structure whose epistemics are intact.
SOURCES
