Summaries and Synthesis

Summaries and synthesis are OKT's slow-accumulation layer — the "system 2" thinking that crystallizes knowledge per concept over time. They are distinct artifacts with distinct roles.

Summaries (per-concept, incremental)

A summary is an incremental slice that folds new facts into a per-concept record. The summaries stage runs as facts become linked to a concept.

Slicing design

As new facts arrive for a concept, the summarizer produces slices:

• Complete slices (is_complete=TRUE) — frozen, immutable, each covering a batch of facts (default 20). Once written they are never touched again.
• One open slice (is_complete=FALSE) — the accumulating slice that grows as new facts arrive. When it reaches the batch size it freezes into a complete slice and a new open slice starts.

The unique partial index uq_concept_summaries_concept_open on (concept_id) WHERE is_complete = FALSE ensures at most one open slice per concept at a time.

The summarization prompt

The summarizer is instructed to:

• Reason ONLY from the provided facts — no training-data knowledge.
• Be credulous — reflect what the facts say, not judge them.
• Prioritize variety, not exhaustion — merge same-point facts, preserve distinct perspectives.
• Use an inverted pyramid — core first, nuances last, so truncation loses the least.
• Cite facts inline with [text](<fact:fact_id>).

Why "slow accumulation"

A future reader can read the slices in order and see a concept's knowledge grow over time. Each slice is a crystallized batch; the open slice is the leading edge. This is why summaries are a slow artifact — they accumulate as sources arrive, not in one shot.

Synthesis (per concept group, authoritative)

A synthesis is the final fold: it takes all the summary slices across a concept group (all contexts sharing the canonical name) and weaves them into one authoritative definition. See the synthesis stage.

What goes in

• All summary slices for the canonical-name group.
• Related concepts loaded from the concept_relations materialized view — the strongest few carry their own synthesis text. This is the Graph-Aware Reasoning context: the model can name bridging concepts, identify interpretive battlegrounds, and flag suppressed topics.
• Image candidates from the group's image facts, optionally filtered by an image-picker LLM call when there are too many.

The synthesis prompt

The synthesis prompt is the most elaborate in the system. It frames the model as a "Synthesis Agent" producing the single authoritative definition for the concept. Eight core principles, including:

1. Attribution-grounded tone — every claim traces to a fact.
2. Radical source neutrality — no source gets prestige-based credibility.
3. Reason through the evidence — synthesize, don't summarize.
4. Preserve all perspectives, including minority ones.
5. Stakeholder motivation analysis — who benefits from each claim being true.
6. Detect institutional deception patterns.
7. Ground everything in the summaries — no outside knowledge.
8. Honest assessment — say what the evidence supports and what it doesn't.

Two distinctive mechanisms:

• Parallel Scenarios — for contested claims, build both "the claim is genuine" and "the claim is an artifact" at full strength, then compare. Only collapse to a single view if the evidence is conclusive.
• Anti-Asymmetry Rules — don't frame mainstream claims as "evidence" and dissenting claims as "claim"; both sides get the same scrutiny.

When synthesis runs

Synthesis is enqueued only when at least one complete summary slice was written in a summarization pass. Open-slice-only updates wait for more facts. This means synthesis reflects crystallized batches of knowledge, not the still-growing open slice.

One per group

The unique constraint (repository_id, lower(canonical_name)) on concept_syntheses ensures one synthesis per concept group. A concept like "Einstein" appearing in contexts Scientist and Person gets one synthesis covering both.

Summaries vs synthesis at a glance

	Summaries	Synthesis
Scope	Per concept row	Per concept group (across contexts)
Input	New facts for one concept	All summary slices for the group + related concepts + images
Mutability	Complete slices frozen; one open slice grows	Re-written when new complete slices arrive
Role	Incremental accumulation	Authoritative definition
When it runs	As facts link to a concept	When at least one complete slice was written

Key tables

Table	Purpose
okt_repository.concept_summaries	Per-concept summary slices: sequence_num, is_complete, content, covered_fact_ids, model
okt_repository.concept_syntheses	One synthesis per canonical-name group: content, covered_summary_ids, covered_concept_ids, embedded_image_ids