Allivista is the working proof of concept for Structural Intelligence — a structural layer that works alongside AI, not in place of it. It surfaces structurally-coherent connections in public research corpora that a field's dominant prose narrative hasn't yet stated; a model can then narrate those connections in plain language, never generate them. The discovery is pre-cognitive and deterministic — no LLM in the analysis path, no hallucination surface in the structure — with the raw result shown alongside as receipts the narration can be checked against.
Architecture
-
deterministic
same input produces bit-identical output; reproducible by any auditor; every result has a provable provenance path
-
non-generative
pure graph computation over a precomputed coordinate space; no probabilistic component; no surface on which hallucination can occur
-
public-corpora-only
ingestion architecturally refuses private email, social media, customer comms, healthcare records (HIPAA), educational records (FERPA), internal corporate documents, and surveillance corpora — the constraint is enforced at request validation, not at the service layer
-
audit-clean
every operation appends to a SHA-256 parent-event chain; refusals are logged alongside acceptances; any auditor can replay the full lineage of any output
Demonstrated
From 9,858 PubMed Alzheimer's abstracts, Allivista surfaces a structurally-coherent cluster of cardiovascular, vessel, cerebrovascular, heart, and kidney concepts as zero-direct-co-occurrence neighbors of the Alzheimer anchor — the classic Don-Swanson hidden-bridging pattern, produced from pure graph structure.
A clinically-informed reading of this cluster inverts the brain-anchored framing of Alzheimer's disease into a systemic-clearance-failure phenotype, of which brain pathology is the visible end-state. The interpretation is consistent with the emerging Klotho, neurovascular-unit, glymphatic, ARIA, AMBAR, and CKD-dementia literatures. The system did not generate the hypothesis. It surfaced the structural fingerprint a clinician then recognized.
9,858 abstracts ·
200,685 symbols ·
1.92M edges ·
45.82s training ·
18ms synthesis
We posed one research question to the same model two ways: once with Allivista's structural tools, once with raw PubMed papers loaded into context. The structural path used ~2.9× fewer input tokens and less than half the cost, and returned organ-system-level findings robust to which papers were sampled. The raw-context path could fit fewer than 1% of the corpus and failed outright past the model's token limit — beyond ~100 papers, structural distillation is not an optimization but the only path that runs.
57,006 tokens structural vs 164,952 raw ·
2.89× ratio ·
<½ the cost
Try it
The full Swagger interface lets you call any endpoint from the browser. The canonical Alzheimer's space is loaded and ready — try a Swanson-style ABC discovery to reproduce the cluster above, or run the 5-method convergence analysis to see how it partitions the structural space.
curl -X POST https://api.allivista.com/v1/synthesize/swanson \
-H "Content-Type: application/json" \
-d '{
"space": "_alzheimer_space",
"anchor": ["alzheimer","alzheimer'\''s","alzheimers","eoad","fad","sad"],
"weighting": "random_walk",
"rank_by": "discovery_ratio",
"top_n": 10
}'
Safety properties
-
no bioweapon synthesis
no prose generator, no instruction-output surface, allowlist refuses private weapons-research corpora at request validation
-
no self-modification
no code-generation component; the patent-protected substrate cannot be automatically replaced
-
no alignment faking
outputs are bit-identically deterministic; no agent to misalign; every output has a verifiable provenance chain
-
no mass surveillance
ingestion allowlist hardcoded at the request-validation layer; refusals logged; the constraint is architectural, not configurable
Patent-relevant anchors
-
Pillar 1 capability snapshot
4c669979ba35a58ce3918a1e2cd5cc2612de2647bf61ef63b612f630d6cf0d36
-
Paradigm framing · claim language constraints
98a25b84db48d178ce74649e89e6131739f3c9a6a86377808629b9e10c418e2c
-
Naming finalization (Allivista)
340fb27d854fe7eab1502929a4c9266f7930be6ef3b8181c81cdd5395efe8e82
Each event is retrievable at https://api.allivista.com/v1/evidence/{event_id}. Every event is signed by content-derived SHA-256 in a parent-event chain; no event can be retroactively altered without invalidating every later event's hash. The full append-only chain is available on request.
Intellectual property
Patent and trademark applications in process. The GeoLang substrate is the patent target; Allivista is the trademark target. Mechanism-level architecture details (placement, scoring, bond extraction, neighborhood patterns) are held privately pending utility patent disclosure. The evidence chain anchors the timeline of every architectural decision under content-derived SHA-256 hashing.
Contact
contact@allivista.com
Substantive academic and institutional inquiry welcome. Press and licensing discussion under separate cover.