Aggregator
Substrate joins
Connectors to the structured stores that hold your real entities — your DB, your Drive, your Linear, your custom app. Output is a typed packet, not a stringified blob.
§ 16Stack · Isilon
Isilon. The evidence-packet builder underneath every AI call we make.
Pulls structured data from your substrate. Validates it. Hashes it. Hands a clean, single-source-of-truth payload to whatever AI tool runs next. Cache-aware, audit-ready, and the reason our downstream products can cite their sources.
§ 01 / The problem we are solving
Most AI workloads fail at the layer below the model.
An LLM is only as good as the evidence packet it sees. Most teams we look at are assembling that packet on the fly inside the request handler — a join here, a stringify there, a quiet truncation when the context blows past the window — and then wondering why the output is inconsistent across calls or impossible to audit after the fact.
Isilon is the layer we built to stop doing that. Every AI call we make in production goes through it. It pulls the structured data the model needs (sessions, journals, psychometrics, external documents, whatever the substrate is), assembles a typed payload, hashes it for cache invalidation, attaches source-attribution markers, and writes a record of exactly what got handed to the model. The downstream tool (Terra, Soqratic, Atelier) consumes it. The audit trail is automatic.
§ 02 / What is in the layer
Five things, one job.
Hash & cache
Content addressing
Every packet hashed by content. The same evidence twice produces the same hash twice. Downstream cache hits become deterministic; re-runs become free.
Attribution
Source markers
Every fact in the packet labelled with its origin (session id, message id, journal day, document chunk). The downstream model cites against these markers; auditors trace claims back to source.
Validators
Schema gates
A packet that fails validation never reaches the model. Missing fields, broken FKs, out-of-window timestamps, rights-cleared flags — all caught at the boundary.
Ledger
What the model saw
The exact packet, with hash and timestamp, written for every AI call. Six months later when someone asks “what did the model see?”, the answer is one query, not a forensics project.
External docs
Bring-your-own corpus
Upload zone for files outside your structured substrate (PDFs, briefs, contracts). Parsed, chunked, attribution-tagged, and slotted into the packet alongside the structured data.
§ 03 / What this is not
An honest non-promise.
Isilon is not a vector store. If you need a production vector database, use one of the three good ones already on the market. Isilon does the layer underneath that — assemble, validate, hash, attribute, ledger — the layer most teams skip and most AI failures trace back to.
We built it because we needed it. We are productising it because every operator team we know is doing this layer badly or not at all.
§ 04 / The shape
Installed, not subscribed.
Isilon ships inside an Avira install today. The standalone SDK extraction is in progress; library users get it next.
Embedded
Embedded
part of an Avira install
- Tuned to your substrate
- Connectors written in week 1
- Schema gates & attribution
- Ledger queryable from your admin surface
Library
Library
extraction in progress
- Library, your runtime
- Connector authoring kit
- Validator + hash + ledger primitives
- Pairs natively with Observatory
Sovereign
Sovereign
regulated industries
- Self-hosted, your domain
- Attribution signing options on request
- Custom validator rules
- Compliance review tailored to your industry
Stop assembling the evidence packet inside the request handler.
If your AI calls cannot answer “what did the model see?” six months later, this is the layer you are missing.