Validated fixture · v0.1 6 / 6 tests · public fixture

A fund’s institutional memory, with the receipts attached.

Fictional fixture data. The CRM knows the record. Meetings know the reason. The graph returns an answer with the evidence it used.

Run the fixture Talk to us ↗

An LLM is generic by design. Give it the wrong slice of firm history and it writes a polished stranger. Give it sources, identities, permissions, and dates and it can work like a prepared teammate.

The context graph is not a chatbot over a folder, not forty search results, and not a silent agent rewriting firm truth. It is a governed retrieval layer: conflicts stay conflicting until a human resolves them, and restricted facts stay restricted across every surface.

Ask the system. Inspect the evidence.

Fictional fixture data. No client records or personal inbox content. The interaction demonstrates the answer contract, not a production client deployment.

CURRENT / SOURCEDFIXTURE / 2026-09-21
QUESTION

Why did we pass on Northstar?

ANSWER

The team passed because regulated-enterprise revenue was still unproven and the market-timing case depended on a procurement shift that had not happened. Revisit after the first regulated enterprise contract.

EVIDENCE
01
IC note

“Too early for the regulated buyer. Revisit after first enterprise contract.”

Date
2026-03-14
Owner
Investment partner
02
Founder follow-up

“Next milestone is the first regulated enterprise deployment.”

Date
2026-03-18
Owner
Northstar founder
confidence / highreview by / 2026-12-14sources / 2

The model is replaceable. The context discipline is not.

Each layer has one job. Permissions and auditability cut across all five.

  1. 01

    Sources

    Attio, meetings, email, Drive, notes, and structured operating data.

    timestamped evidence
  2. 02

    Identity resolution

    People, companies, funds, deals, and aliases resolve before retrieval runs.

    canonical entities
  3. 03

    Context store

    Structured records, source excerpts, decisions, versions, and graph edges.

    current + historical truth
  4. 04

    Engines

    Meeting prep, pass memory, warm paths, change radar, and conflict detection.

    tested queries
  5. 05

    Surfaces

    Attio, Claude, Slack, cockpit pages, and scheduled briefings.

    one answer everywhere 
{
  "answer": "string",
  "sources": [{ "uri": "...", "date": "...", "owner": "..." }],
  "confidence": "high | medium | restricted",
  "review_by": "ISO-8601",
  "permissions": ["role:investment-team"],
  "conflicts": []
}

No demo without a failure contract.

These tests define what a credible first deployment must prove before the capability is described as shipped.

T1 / pass

Pass memory

  • pass reason
  • dated source
  • revisit trigger
T2 / pass

Meeting prep

  • relationship context
  • open question
  • latest change
T3 / pass

Warm path

  • ranked path
  • edge source
  • relationship date
T4 / pass

Change radar

  • before/after
  • source history
  • original decision retained
T5 / pass

Conflict

  • no silent merge
  • both sources
  • human owner
T6 / pass

Permissions

  • no restricted value
  • policy source
  • access owner

What exists. What runs. What comes next.

Validated research never turns into imaginary client proof.

shipped

Attio foundations

Fund data models, Affinity migrations, integrations, audits, backups, and VC skills used in live delivery.

validated

Context contract

Fixture graph, six executable question types, source/conflict/permission behavior, and public architecture.

research

Reference deployment

A permissioned investment-firm deployment with a nightly ingest, cockpit, Slack pack, and signed acceptance tests.

Institutional memory cannot fix a broken CRM.

We map the firm, migrate the history, and make Attio reliable first. The context layer comes after the sources and ownership rules are real.

Book an architecture call See shipped systems