Hybrid Reasoning · your choice, workflow by workflow

An outside AI can help without ever seeing a name.

Hybrid Reasoning brings in outside help without letting anything identifiable leave the building. It is off until you switch it on, one workflow at a time.

This page has a twin — the record of every job the machine runs, all readable afterwards. the run record

The record of what leaves

Read what actually left the building.

One real question, step by step. Your IT team can read the real record the same way.

Egress ledger

STEP 01 · LOCAL · 03:41:19 HKT

The question, as the user typed it

            Between Priya Raman and Wai-kit Cheung, who fits the Head of
Compliance role at Meridian Partners? Budget is HK$1.85M and
the client wants someone who has taken an SFC Type 4 licence
through a first inspection.
          

NOTHING

The question never leaves the machine in this form. It is only used to search your own files.

Local · 0 bytes out Session opened, map empty

1 / 5

STEP 02 · LOCAL RETRIEVAL · 03:41:20 HKT

What the machine found in your own files

            4 documents, 11 passages, all resident on the appliance:

  · CV — Priya Raman (2024-11).pdf              p.1–2
  · CV — Wai-kit Cheung (2025-02).pdf           p.1, 3
  · Meridian — role brief, Head of Compliance.docx
  · Placement notes — Meridian, Q1.md           §2

Retrieval: BGE-M3 · Qdrant · in-memory · 380ms
          

NOTHING

All of this searching happens inside the machine. No filename or passage is sent anywhere.

Local · 0 bytes out 11 passages held in memory

2 / 5
Boundary

STEP 03 · OUTBOUND · 03:41:22 HKT · 1,204 BYTES · LOGGED

Exactly what left the machine, in full

            Between Priya Raman and Wai-kit Cheung, who fits the Head of
Compliance role at Meridian Partners? Budget is HK$1.85M and
the client wants someone who has taken an SFC Type 4 licence
through a first inspection.

(Toggle to "What we sent" to see what actually left the machine.)
          

CODE WORDS

Code words only — no name, figure, filename or firm. The key that links code words back to real names never leaves the machine.

1,204 bytes out · logged Tripwire: pass · ZDR, no-training terms Frontier cost: US$0.015

3 / 5

STEP 04 · INBOUND · 03:41:24 HKT

The outside AI’s answer, still in code words

            The outside AI answers about [PERSON_11] and [PERSON_12]. It
has no way of knowing who they are.
          

OUTLINE

Reasoning and judgement, but no facts about your firm. The machine swaps your real details back in during the next step.

Inbound · 812 bytes · logged No identities present

4 / 5

STEP 05 · LOCAL COMPOSITION · 03:41:25 HKT

The answer the user actually sees

            Priya Raman is the stronger fit for Head of Compliance at
Meridian Partners.

She carried an SFC Type 4 licence through a first inspection
at her 2021 firm [CV — Priya Raman (2024-11).pdf, p.2] and
has 11 years in asset management [ibid., p.1].

Wai-kit Cheung has 8 years and no first-inspection exposure
[CV — Wai-kit Cheung (2025-02).pdf, p.1], so would need
supervisory cover for the first cycle.

HK$1.85M sits at the market range for this seniority
[Placement notes — Meridian, Q1.md, §2].

— 3 claims cited · faithfulness gate passed · 0 orphans
          

NOTHING

The machine translates the answer back, puts your own names and figures in, and cites every fact to its source. It is checked before you see it.

Local · 0 bytes out Orphan scan: clean Sending this outside the firm requires approval

5 / 5

An illustrative example — no client data. The structure is exactly what really leaves the machine. The key that links code words to real names exists only on the machine.

Cloud AI asks you to trust a contract. Midas asks you to read the log.

The privacy sandwich

Four steps. The machine reaches outside exactly once.

01

Find the facts on the machine

It searches everything you have given it, entirely inside your own office — your documents never leave.

02

Swap every name for a code word

Three passes swap every name and number for a code word — [CLIENT_A] means the same firm every time. When in doubt the machine removes the detail, and the key to the code words never leaves.

03

Ask the outside AI — code words only

The question — now entirely code words — leaves the machine exactly once, recorded and checked against your never-send list. Anthropic Claude answers under written terms — nothing stored, nothing used for training — and the code words come back untouched.

04

Put the real names back — then check everything

Back inside, code words become real names again, and exact figures come from your own files. Every answer is checked first; if anything fails, a person reviews it instead.

INFORMATION LEAVES THE MACHINE ONCE — IN STEP 03, CARRYING CODE WORDS ONLY

Recorded on the way out · checked against your never-send list · every byte kept in the log

For each question, the machine decides whether to bring in outside help. On hard questions the outside AI supplies the outline, your machine the facts — an answer neither could produce alone.

Measured results

The test results, and the check that makes them believable.

0 / 38

Outbound messages that contained anything sensitive

Zero leaks across all testing to date, re-checked independently against a list written in advance.

11 / 15

Canary check — mistakes sent through on purpose that set off the alarm

An alarm that never goes off proves nothing — this one does, and we publish the misses.

13 / 15

Answers with outside help beat the machine working alone

The judges compared answers side by side, not knowing which system wrote which.

12 / 15

Matched or beat handing the outside AI everything, unmasked

Exact figures from your own files, with sources named, beat an outside AI's vague recall. Outside-AI cost per question: about US$0.015.

We tested on a set of files invented for testing — no client data — built so that we held a complete list, written in advance, of every name and number that must never leave, so we could search for each one exactly. We recorded what left at the machine's one door to the outside, not in the software's own notes. Answers were compared side by side, and the judges did not know which system wrote which. For details with a fixed pattern — ID numbers, phone numbers, sums of money — the protection is certain by design. Beyond those, it is layered and measured, not absolute, with a physical stop in the path behind both. We do not claim more than that. Certain where certainty is possible; layered and measured where it is not; and a stop built into the hardware behind both.

We didn't trust our own logs until we tapped our own wire.

Bring a hard question. We'll show you exactly what left the machine.

Your IT team can read the record in the showroom, then again on your own machine.