INGEST01
Evidence intake
Source material is normalized, classified and attached to durable origin records before claims are admitted.
Ribosome
Information extraction, retrieval and resynthesis without neural models. Lower compute cost, deterministic classification with full provenance, and no generative hallucinations.
operational No embeddings · closed verb frames · explicit evidence authority
PB-02 operational
SYS/PREDICATE 04 CHANNELS PALOMA BB
- Neural models
- None
- Embeddings
- None
- Representation
- Predicate graph
- Authority
- Explicit evidence
01
02
03
04
What is real today
Retrieval that can explain what it retrieved.
Ribosome is a completely non-neural system for information extraction, retrieval and resynthesis. It converts source material into explicit entities, predicates and evidence relationships rather than compressing meaning into vectors.
Facts can remain tentative, become answerable when evidence supports them, and later be corrected or retracted. The graph preserves where a claim came from and which transformation produced it.
The point is not to imitate a language model with rules. It is to build a different kind of information machine whose representations remain inspectable all the way through.
System shape
Evidence enters as structure and leaves as structure.
The pipeline constrains extraction enough to keep meaning legible while still supporting practical recall and resynthesis.
FRAME02
Predicate extraction
Closed verb frames turn language into explicit relationships instead of unconstrained generated summaries.
IDENT03
Entity resolution
Names and mentions resolve into a graph with abstention when identity is not sufficiently supported.
RECALL04
Retrieval + resynthesis
Answers are assembled from answerable graph facts with their evidence and authority still attached.
Current evidence
Opacity is treated as a defect, not a feature.
Ribosome is designed around things ordinary retrieval stacks tend to hide: admission, identity, contradiction and provenance.
01
Semantics
Closed frames
Relationships use a deliberately bounded vocabulary rather than unconstrained latent similarity.
02
Correction
First-class
Retraction and supersession are represented inside the graph instead of patched over at response time.
03
Identity
Abstains on miss
The system prefers unresolved identity to an attractive but unsupported merge.
04
Answering
Evidence-gated
A stored proposition is not automatically authorized to answer a human question.
Operating principles
Constraints that keep the work honest.
No hidden semantic layer
Meaning remains represented in entities, relations and evidence.
Abstention over invention
Unsupported identity or claims remain unresolved.
Correction is native
Retraction is part of the data model, not cleanup.
Authority stays visible
Every answerable fact carries a reason it may be used.
Connected work