AIGISAIGIS SOLUTIONS Find your solution

AIGIS Context Engine

Architecture answers. AI only if it must.

The Context Engine reads your systems, maps them into one trusted language, routes to the one artifact that is actually current, and answers from verified facts. Every answer carries a receipt. A model is the last step in the pipeline, not the first — and on most questions it never runs at all.

>_ See the app on a desk Run the demo Why this exists

01 · DESIGN

Model the objects first

What an employee, a document, a policy, or a plan actually is — named, versioned, and validated before anything reads it.

02 · ARCHITECTURE

Make the rules mechanical

Effective dates, supersedence, jurisdiction, audience, and review state become code paths that fail closed, not instructions someone hopes a model will follow.

03 · AI

Add a model at the edge

Only for language, only behind a policy gate, only over a minimized packet — and it is never allowed to supply a fact the architecture did not verify.

Deployed · what your team actually opens

The engine is the back end. This is the application installed on your desks.

Nobody on your staff logs into a context engine. They open an application carrying your name, ask in your own terminology, and get the current answer with the receipt attached — or reach it from the system tray without opening anything at all. The wrapper is yours. The engine underneath is the same one in every deployment. Switch companies below and watch what changes, and what cannot. The engine that answers is demonstrated directly below.

Viewing as

managed workstation · governed deployment

Knowledge Desk — Acme Fabrication
AF
Acme Fabrication
Knowledge Desk · engine by AIGIS Solutions
1,284 governed artifacts · SSO: Entra ID

Asked in the app

What is our current PTO carryover limit for salaried staff?

Salaried staff may carry over a maximum of 40 hours into the next calendar year. Hours above that are forfeited on 31 December unless HR approves a written exception.

CurrentEmployee Handbook v4.2 — §7.3 Paid Time Off (effective 2026-01-01 → open)
SupersededEmployee Handbook v3.1 — §7.3 (retired 2025-12-31, stated 80 hours)
answer_source.... verified canonical facts
model_calls...... 0
excluded......... 1 superseded artifact (on the record)
receipt.......... JCS-RFC8785 · SHA-256 4f2a9c…e17b
AFKnowledge Desk GOVERNED
▦◱✉
^
2:41 PM7/26/2026
WRAPPER CHANGES → name · accent · terminology · corpus · sign-in | ENGINE DOES NOT → routing · currency · exclusion · receipt

Why it lives in the tray

A wrong answer happens at a moment — someone needs a number and reaches for whatever is closest. Two clicks in the tray beats opening a portal, and it beats asking a general assistant that has no idea which version is current.

What every answer carries

Source, effective window, what was excluded, and a receipt — in the same box as the answer. A bare answer with no provenance is indistinguishable from the thing it was built to replace.

The deployments above are illustrative wrappers over synthetic fixtures, shown to demonstrate the client-facing shape of a rollout. The interactive engine demo is the section directly below this one.

Interactive · executed on AIGIS infrastructure

Watch it refuse to answer from a document that expired last year.

The same engine, at two depths. Start simplified: one ordinary question, three documents that all mention it, and the plain-language reason only one of them is allowed to answer. Switch to technical when you want to drive the engine yourself — generate an unfamiliar API shape, edit the payload, and watch every mapping, routing decision, and SHA-256 receipt as it executes.

Illustrative fixtures, shown to make the shape of a governed answer legible. The technical view runs the real engine on AIGIS infrastructure.

Why this product exists

An employee asked a simple question and got an answer that stopped being true a year earlier.

A client came to us after an incident with their general-purpose assistant, wired to their internal document store. Nothing crashed. Nothing was flagged. The system did exactly what it was built to do — and that was the problem.

01 · What happened

The wrong document, delivered confidently

An employee asked how to complete a routine HR process. The assistant searched the document store, found a close textual match, and returned it — a superseded revision that had expired the previous year.

The employee had no way to tell. The answer read exactly like a correct one.

02 · Why a prompt cannot fix it

“Only use current documents” is not a fact

A prompt is an instruction. Supersedence is a fact about your documents — and it lived nowhere the retrieval layer could see. The store held eight revisions with no notion of which one was in force.

You cannot instruct a system toward a fact it was never told. Better wording narrows the odds; it does not close the hole.

03 · What we built instead

Make “current” a property, then enforce it

We modelled the document — lifecycle status, effective window, review date, jurisdiction, audience, supersedence chain, content hash — and made the router prove exactly one artifact is current before an answer is composed.

If it cannot prove that, it says so and escalates. It does not guess.

Retrieval that has no concept of “current” will eventually hand someone a document that is no longer true. That is not a model failure. It is a missing design.

Same question · same document store

One returns a match. The other returns the current answer.

Both systems can see all eight revisions of the same synthetic HR document family. Only one of them knows what “current” means.

General-purpose assistant over a document store

Similarity

>_ What is the current process for changing my address?

Returns the closest textual match and summarises it fluently. There is no version reasoning in the path, so the answer inherits whatever the index surfaced first.

address-change-requestv2effective 2024-01-01 → 2025-12-31Expired
  • ×Superseded by v3 — the chain is not modelled, so it cannot be checked.
  • ×Review overdue since 2025-06-15 — no review state exists in the index.
  • ×Effective window closed 2025-12-31 — dates are text, not policy.
  • ×No record of what was considered or rejected. Nothing to audit after the fact.

AIGIS Context Engine

Governed routing

>_ What is the current process for changing my address?

Selects exactly one artifact that satisfies every governance predicate, then composes the answer deterministically from it — with the rejected candidates recorded as evidence.

address-change-requestv3effective 2026-01-01 → openCurrent
  • ✓Published, inside its effective window, review not due until 2027-06-15.
  • ✓Supersedes v2 and v1; both are excluded on the record with a stated reason.
  • ✓Jurisdiction and audience checked against the canonical employee object.
  • ✓Content hash, route ID, and output hash returned as a receipt you can re-verify.

Both columns describe the same synthetic fixture catalog used in the demo above. The left column characterises a general retrieval pattern, not any specific vendor or product.

Next step

Run it against your documents.

The demo above uses synthetic fixtures. The interesting version uses your real catalog — and the first thing it usually finds is how many of your live answers are quietly out of date.

Recommended

Request a scoped pilot

One document family, one workflow, your own catalog. We model the objects, stand up the governed route, and show you the receipts.

>_ start a pilot Half a day

Architecture review

Bring the assistant or retrieval system you already have. We map where facts enter the path, where they cannot be verified, and what fails closed.

>_ book a review Reading

Design and Architecture > AI

The written argument behind this product: why the order of operations decides whether an AI system is trustworthy.

>_ read the note