CLEAN-ROOM RECREATION

Median — Media Library Assistant

Turn clutter into reviewed, traceable next steps.

A fictional local-first workflow showing how an agent can search, inspect ambiguity, prepare a safe plan and wait for a human decision before it records a simulated audit event.

  • FICTIONAL DATA
  • NO LIVE BACKEND
  • LOCAL-FIRST DESIGN

01 / The problem

Automation is most useful when uncertainty remains visible.

An untidy personal library can contain clear matches, near matches and records that simply do not say enough. This clean-room recreation demonstrates the safety pattern: make each case legible, route ambiguity to review, and keep the final decision with a person.

Searchable

Start with a fictional catalogue record rather than an opaque collection.

Reviewable

Show why a case can proceed, needs a hold, or cannot be identified.

Traceable

Record a simulated decision without changing files or contacting anything.

02 / My bounded contribution

I designed the staged decision pattern and its public explanation.

The private-source pattern informed the capability story. This page is freshly authored from a safe description of that pattern: a fictional fixture set, conceptual locations, visible holds and a non-operational decision trail.

  • Break a complex workflow into inspectable stages.
  • Pair deterministic checks with an explicit ambiguity review.
  • Design a human gate before any simulated outcome is recorded.
  • Translate a technical workflow into a clear, privacy-safe interface.

03 / How it works

A decision path that stops when confidence is not enough.

These conceptual stages describe the safety pattern, not a live system. The review branch is intentional: a plausible answer is not treated as permission to proceed.

  1. InspectRead a fictional record and declared context.
  2. IdentifySeparate confident, possible and unknown identity.
  3. PlanPrepare a reversible simulated next step.
  4. Local reviewExplain ambiguity and keep it on hold.
  5. Human gateApprove or reject the simulation.
  6. AuditAppend a fictional decision event.

04 / Fictional workflow simulation

Choose a case. The simulated trail explains the decision.

Everything below is bundled in this page and marked synthetic. The controls do not search a library, send data, or change anything outside this browser view.

Five fictional review cases

Aurora Drift (2024)

Fictional result: a clear identity can receive a simulated organisation preview, but still waits for the human gate.

05 / What this demonstrates

Tools, judgement and accountability each have a different job.

A useful agent pattern is not a single instruction. It combines constrained inputs, deterministic stages, an explanation of uncertainty, a human decision, and an audit trail that makes the outcome reviewable later.

Transferable capabilities

  • Workflow analysis
  • Human-centred automation
  • Ambiguity handling
  • Local-first design
  • Interface design
  • Audit-minded delivery

06 / Boundaries

A safe explanation, not an operational service.

What this page does

  • Uses five new fictional records, each marked synthetic.
  • Demonstrates confident, possible and held paths.
  • Shows simulated approval, rejection and an audit event.

What this page does not claim

  • No live library, connection, file access, upload, copy or external action.
  • No automatic decisions, commercial deployment, accuracy result or adoption claim.
  • No reproduction of private code, operational details, records or tool surface.

Selected work

Explore the rest of the portfolio.

Back to all projects