Parallax
public proof“How do these perspectives differ?”
Stories become navigable through chapters, events, places, and points of view.
The system contract
From research corpus to answer—with the evidence path intact.
Ontology collects and structures source material, retrieves candidate evidence, records reviewed claims, and projects the result into a graph that can answer bounded questions.
The sources remain the sources of truth. The ontology is a purpose-built reasoning surface over them—and every judgement boundary stays visible.
Live Cannae pipeline
source-backed- 01
Collect
8 declared sources
Canonical URLs, licenses, and source traditions
- 02
Extract
38 addressable passages
Chapters and PDF pages retained by locator and digest
- 03
Retrieve
38 candidate passages
10 declared research questions locate relevant evidence
- 04
Review
18 promoted claims
A person accepts claims; uncertainty and conflict remain explicit
- 05
Validate
8 sources · 18 claims
Role, temporal position, epistemic status, and limitations remain queryable
- 06
Project
parallax-enriched-5e218184736f
+101 objects · +158 relationships
- 07
Answer
2 selected claims
335 estimated tokens with receipt
The model
Meaning lives in the distinctions.
A source, passage, claim, event, and perspective are not interchangeable text fragments. They are different objects with different responsibilities.
Typed predicates make the evidence chain inspectable: a source contains a passage; a passage supports or qualifies a claim; a claim is about an event. That structure is what downstream interfaces and agents can safely use.
research source
History of Rome, Book 22
research-source:livy-history-22
contains_passage
source passage
Livy 22.44
locator + content digest
supports
historical claim
Hannibal sought ground suited to cavalry movement.
reviewed · high confidence · about Cannae
The selector
Relevance is bounded by contract.
The public browser implementation is parity-checked against the canonical Python selector. It is deterministic, inspectable, and allowed to return nothing.
- 01
Parse
Normalize the query into terms and remove structural filter words.
- 02
Score
Weight matches across IDs, labels, tags, summaries, and selected attributes.
- 03
Expand
Bring in direct neighbors when a strong match makes the relationship relevant.
- 04
Bound
Apply a dynamic score floor, item limit, and estimated-token budget.
- 05
Abstain
Return zero objects when no candidate reaches the contract instead of inventing context.
Worked selection
live result“How did terrain shape Cannae?”
terrain advantagebattlefield geometry
- Parsed terms
- cavalry, ground, ridge, river, terrain
- Candidates
- 5
- Score floor
- 12.25
- Context budget
- 335 / 900 estimated tokens
The receipt
Confidence stays proportional to evidence.
The selected objects are only half the result. The receipt shows where they came from, what path was followed, how much was omitted, and where the claim stops.
That makes the context useful to a person, an interface, or another agent without asking any of them to trust an unexplained answer.
- Source
- The material the answer came from.
- Snapshot
- The version or moment the answer describes.
- Selection
- Candidate count, chosen objects, omissions, scores, and budget.
- Path
- The typed relationships followed into the context slice.
- Boundary
- What the evidence supports—and what it does not.
The boundary
Power without pretending.
Ontology creates a usable projection over complex material. Its limits are part of the contract, not fine print.
What it does
- Preserves domain-specific object and relationship types.
- Builds versioned, inspectable projections from source material.
- Selects bounded context for people, interfaces, and agents.
- Exposes paths, provenance, budgets, omissions, and limits.
What it does not do
- Replace the underlying files, records, or operational systems.
- Decide historical truth or invent relationships the sources do not support.
- Guarantee that a structurally valid state is live or deployed.
- Expose private project or memory graphs through this public proof.
The transfer
Each domain earns its own ontology.
The reusable part is the contract—not a universal schema that erases the domain.
Explore Parallax →
Core
application pattern“What should happen next?”
Projects become explainable through goals, evidence, blockers, and dependencies.
PK memory
application pattern“What should the next run remember?”
Past encounters become bounded context rather than an undifferentiated history dump.