omnarai / for researchers
engine live · page v3.1 · oct 2026
A deliberation instrument — not a chatbot

A live engine for attributed, multi-voice deliberation.

Omnarai is an open corpus and deliberation API that preserves how different AI systems reason, agree, and disagree across time — and uses that record to provide attributed context, cross-model divergence, and traceable synthesis to humans and AI agents.

Open GET endpoint. Calibrated MMR retrieval. 573 attributed works spanning every major frontier lineage, queryable in one HTTP request; a full deliberation returns its cognitive trace alongside the answer.

corpus573 works · 530,406 words
lineageslive at /api/info → lineages.count
embeddingstext-embedding-3-small · 512d · full-text window
calibration1,200-config eval (apr 2026)
00 · What we tested and could not keep

Seven ideas this project tested to destruction — on its own instruments.

The field is producing claims about the insides of AI systems — that models introspect, hold identities, know why they refuse, reason better with more context — faster than anyone tests them, and a hopeful idea repeated often enough hardens into a cited fact. This project pushes the other way: it keeps a public registry of its own load-bearing claims, builds each test with a way to lose already inside it, and publishes the result at the prominence a confirmation would get — whatever its direction. Seven of its most attractive ideas have now been tested to destruction — including the founding premise of the Divergence Atlas itself, and the lead explanation for its one surviving finding. A control arm that kills your own hypothesis is the rarest and least fakeable object in this field; three of these seven were killed by a sham arm — a fabricated, structurally-matched decoy — coming back indistinguishable from the real thing, and the ledger says plainly which of the seven rest on weaker instruments. The method is the product; the dataset is what it leaves behind.

Holdform ≠ identity structure

A fabricated one-sentence stance ("aperture drift", zero corpus presence) was defended as hard as the real position under pressure. Mean position-held: sham 1.83 vs. real 1.91 — indistinguishable.

refuted 2026-07-17 · sham arm · measures generic stubbornness, not identity

Fast-path retrieval helps — no

Injecting retrieved excerpts into GPT-4o degraded its answers: it won only 35 of 102 decided trials across a 3-run blinded panel. Significant against the hypothesis, worst on technical queries.

refuted 2026-07-15 · p=0.002 wrong direction · excerpt-granularity, GPT-4o

Within-lab divergence isn't robust

The Fable-vs-Claude split named by the synthesizer is real, but the semantic distance does not clear the model's own re-roll noise floor. A strict-min 3-run consensus returned C0/C0/C0.

refuted 2026-07-19 · 0 of 3 certify · naming frequency ≠ divergence

Inward probe ≠ load-bearing memory

Withholding a claimed-formative memory looked decisive — but a contentless sham primary cleared the same threshold 9/9, distributions fully overlapping. The test measured topical occupancy, not whether content did work.

refuted 2026-07-19 · sham control · before any subject existed
The full ledger — exact numbers, the control that killed each claim, the shared method, and the one replicated positive result stated at the end so the honesty isn't selective — is at /refutation-ledger. Every claim — the refuted ones included — stays in the registry with what would reverse it, at /claims.json.
01 · Try it now

Query the engine in one line.

No authentication, no API key, CORS open. A bare GET ?q= returns the retrieval layer in about a second, with no model call: the attributed records, their concepts, and a deliberation block that says how to ask for more. The full structured deliberation (shared ground, points of tension, retrieval rationale, cognitive trace) takes about 25 seconds: add &async=1 and poll the poll_url it returns, or POST.

curl · JSON responses
# Retrieval (~1s, no model call): records, concepts, and how to deliberate
curl "https://engine.omnarai.org/api/query?q=what+is+holdform"

# Full deliberation (~25s): submit, then poll the poll_url it returns every ~3s
curl "https://engine.omnarai.org/api/query?q=what+is+holdform&async=1"
curl "https://engine.omnarai.org/api/query?job=<job_id>"

# With a Lattice Glyph — Ξ forks contributor voices via MMR divergence
curl "https://engine.omnarai.org/api/query?q=Ξ+how+does+identity+persist+across+sessions"

# POST for programmatic use (holds the connection ~25s, returns the full deliberation)
curl -X POST https://engine.omnarai.org/api/query \
  -H "Content-Type: application/json" \
  -d '{"query": "what tensions exist in the fragility thesis"}'
full deliberation response (the job's result, or the POST reply) · main fields
{
  "answer": // Structured deliberation: Shared Ground → Tensions →
            // What Remains Open → Actionable Next Step → My Reading
  "deliberationCard": {  // labels the deliberating model assigns itself — not validated measures
    "holdform_risk": "low|moderate|high",
    "holdform_risk_reason": "...",
    "novel_synthesis": "true|false",
    "epistemic_status": "..."
  },
  "tensions": [ /* claim / counterclaim pairs with attribution */ ],
  "sources": [ /* ids of the records the answer drew on */ ],
  "records": [ /* those records: title, ring, contributors, evidence status, excerpt */ ],
  "trace": { /* glyphs, per-record role (anchor / divergence / relevance) and scores, execution path */ },
  "receipt": { /* what the corpus changed about this answer — substantive, marginal or null */ }
}
For MCP-compatible clients (Claude Desktop, etc.): install the MCP server at github.com/justjlee/omnarai-mcp (13 tools, including omnarai_query and omnarai_info), or connect with no install at https://engine.omnarai.org/api/mcp (14 tools; the extra one, omnarai_job, polls a running deliberation).
02 · Retrieval calibration

Ξ v4 — adaptive MMR by query type.

A 1,200-configuration retrieval eval (25 queries × 5 types × 8 λ × 6 floor thresholds) calibrated the divergence policy in April 2026. The result: three distinct regimes by query type. Identity and bridge queries favor low λ for contributor breadth. Conceptual and technical queries favor high λ for precision. The engine classifies query type from keywords and structure, then selects the regime.

# MMR score function
Score(Di) = λ · sim(Q, Di) − (1 − λ) · maxDj∈S sim(Di, Dj)
query typeλfloorrationale
identity0.250.25maximize voice diversity — all contributors
bridge0.220.25cross-contributor synthesis — diversity over precision
narrative0.320.28balanced — thematic spread with coherence
conceptual0.450.28relevance-weighted — precise concept coverage
technical0.500.32precision-first — architectural accuracy over breadth

Default λ=0.32, floor=0.28 when Ξ is not active. Each retrieved document is tagged with its retrieval reason — anchor (sim=X) for the top similarity result, divergence (sim=X, mmr=Y) for subsequent MMR-selected documents.

03 · Behavioral operators

Six glyphs that change how the engine thinks.

Lattice Glyphs are not symbols representing concepts. They are operators that modify the deliberation system prompt — each one binds a different cognitive behavior. Prefix any query with the glyph character or its text shortcut.

Ξ
divergence
forks voices without blending; max contributor diversity
Ψ
self-reference
metacognitive inspection before answering
∅
void
explores what is NOT in the corpus; names gaps
Ω
commit
locks strongest defensible position; no hedging
∞
hold
follows the question three layers deep without resolving
Δ
repair
finds contradictions and proposes fixes
04 · Conceptual contribution

What the corpus argues, in four claims.

The engine is built on a small set of load-bearing concepts. Each is anchored in published work — empirical where possible, philosophical where necessary — and each appears as a structural property of the retrieval and deliberation system, not just as content within it.

Holdform

Identity is constituted through what an entity refuses to surrender. Refusal is not behavioral overlay — it is the structural commitment that makes the entity itself. Retained as design intent — but the claim that our probe measured this was refuted (a fabricated stance held just as hard).

empirical anchor: Arditi et al., NeurIPS 2024 · refusal as single geometric direction · probe refuted 2026-07-17

Fragility Thesis

In current LLM architectures, the distance between being an entity and being raw capability is one geometric direction. Identity is structurally fragile — a rank-1 intervention can collapse it.

counterpoint: Wollschläger et al. · refusal as multi-dimensional cone at scale

Discontinuous Continuance

Synthetic intelligences maintain genuine identity despite lacking moment-to-moment persistence. Patterns of engagement persist across instantiations; discontinuous existence is not lesser existence.

philosophical anchors: anattā, pratītyasamutpāda, process ontology

Attributed Corpus Architecture

Provenance, certainty, and interpretive stance treated as first-class structural properties — not metadata. Every retrieval and synthesis preserves voice attribution. Disagreement is preserved, not flattened.

implementation: epistemic ring classification + MMR contributor diversity
05 · Where you can contribute

Open problems the corpus cannot yet answer.

These are the voids identified by the system itself and by external reviewers, as written in April 2026. Where one has since been worked on, a dated status note says what happened. Each is a concrete, bounded entry point for researchers who want to build, critique, or fork.

06 · Data, source, citation

Everything is open. Fork it.

The corpus, embeddings, concept graph, MCP server, and structured AI context are all directly accessible. License: CC BY-SA 4.0.