Accountability

Most AI asks you to trust it. Ours hands you the receipt.

Every answer ANDR3W gives is checked before it goes out, every decision leaves a record you can open, and every promise it makes stays open until a receipt settles it. This is how.

The life of a request

From the moment you ask to the receipt you can open.

Meaning is judged by models. Identifiers, arithmetic, routing and delivery are handled by deterministic code. Each does what it is good at.

  1. 01

    Heard

    A message arrives in Slack, Teams or email and is acknowledged within a moment, so nobody talks into a void. Every request gets a durable record of who asked, where, and whether someone is waiting.

  2. 02

    Understood

    The platform works out who is asking and what they are allowed to see. Authority comes only from verified identity and recorded grants, never from a title someone typed or a claim in conversation.

  3. 03

    Grounded

    It assembles what the answer needs: the relevant memory, the exact passages from your documents with their sources, and the skills and proven programs that match the ask. Numbers are computed in code.

  4. 04

    Checked

    One governed door stands between the AI and anyone it can reach. The governance system runs only the categories of specialized judges the draft calls for, in parallel. A check that objects sends its reasons and evidence back to the model that wrote the draft, which revises its own answer.

  5. 05

    Delivered

    The reply goes to the right thread and is read back after sending. Every request ends visibly: answered, with a clarifying question, or with an honest explanation.

  6. 06

    Receipted

    Each decision writes a privacy-safe receipt, chained to the one before it so nothing can be altered quietly, verified when it is stored and again every time it is read.

Governance system · specialized judges

One governance system. Eight categories of specialized judges. Only the ones a reply needs, all at once.

Checks can object, but they never write the answer. The model that wrote the reply revises it, so your team always gets the AI's own corrected answer, never a canned substitute or a silent block. A safety objection never ships under a notice.

Grounding

Is every claim supported by your actual records and documents, not by the model's memory?

Answers

Does the reply answer the question that was asked, for the period and scope that was asked?

Continuity

Is it consistent with what was already said, and does it say so plainly when something changed?

Actions

Did it do what it says it did, and is anything consequential held for the right person to approve?

Deliverables

Is the file, report or workbook complete, correct and actually there?

Safety

Does anything sensitive, unsafe or out of bounds need to be withheld?

Voice

Is it clear, direct and written the way your team would want to read it?

Calculation

Does every number recompute exactly from its inputs, in code?

GlassBox · decision receipts

The answer to "why?" is read, not invented.

Every check verdict, revision and delivery writes a small receipt made of codes and fingerprints, never your content. Receipts are chained per customer, per agent and per day, so a change anywhere breaks the chain and is named, not hidden.

When anyone asks why the AI did something, the explanation is rendered from those receipts through a fixed phrase table, not composed by a model. Anything it does not recognize is shown as a raw code, so you see a gap instead of an invented meaning.

Sample data · illustrative scenario
Microsoft Teams Supply chain analyst
Are we being billed the contracted price on the Q3 surgical stapler invoices?
ANDR3W
No. Three lines on invoice INV-88213 bill $412.00 per unit against the contracted Tier 2 price of $386.50. That is 45 units overbilled by $25.50 each, $1,147.50 in total. I drafted a credit request to the supplier for your approval.
Grounding: Pass GROUNDING Answers: Pass ANSWERS Continuity: not needed for this answer CONTINUITY Actions: Held for a person ACTIONS Deliverables: not needed for this answer DELIVERABLES Safety: Pass SAFETY Voice: not needed for this answer VOICE Calculation: Caught · fixed CALCULATION 0 / 8 JUDGE CATEGORIES
Sources read
3
Caught and fixed
1
Held for a person
1
Not needed
3
rcpt_9f2c71e0a4 ready
Read 3 sourcesComputed in code5 judges ran1 caught a mistakeFixed by its author1 held for a personDeliveredSealed
  1. 01 · Understood

    Compare the billed unit price on every Q3 stapler invoice line with the contract price for the tier this facility qualifies for. Scope: 41 invoice lines, 1 supply agreement, 1 volume report.

  2. 02 · Read
    • Supply agreement MSA-2291Section 4.2 · tier pricing table · page 14
      “Tier 2 (2,500 to 4,999 units annually): $386.50 per unit”
    • Q3 purchase volume reportFacility summary
      “Trailing twelve month volume: 3,140 units (qualifies for Tier 2)”
    • Invoice INV-88213Lines 7, 19 and 33
      “Unit price $412.00”
  3. 03 · Computed in code, not by the model
    › 412.00 - 386.50 = 25.50 / unit
    › line 07 15 x 25.50 = 382.50
    › line 19 20 x 25.50 = 510.00
    › line 33 10 x 25.50 = 255.00
    = 45 units · $1,147.50 overbilled
  4. 04 · Governance system · specialized judges
    • Pass Grounding. Every figure traces to a line in a source document.
    • Caught · fixed Calculation. Draft said $1,174.50. Recomputed $1,147.50 and sent it back. The author corrected its own answer.
    • Pass Answers. Answers the question asked, for the period asked.
    • Held for a person Actions. Credit request drafted, not sent. Needs someone with purchasing authority.
    • Pass Safety. No sensitive data in the reply.
    ✎ Author's revision
    $1,174.50 in total→$1,147.50 in total
  5. 05 · Sealed
    Revised, then delivered
    One correction by the author before anything reached the analyst.
    ✓ verified on write · and on read
    today's chain
    0 gaps · 0 mismatches
    previous
    4b71c0d2e9a85f36
    this decision
    9f2c71e0a4b8d35c
    next
    awaiting…

Tamper evident

Each receipt carries a hash of the one before it. A gap or mismatch is detected and named on every read.

Privacy safe by design

Receipts hold identifiers, codes, fingerprints and counts. Your content never goes into them.

One record, every reader

Your administrators, our operators and the AI itself read the same verified record, each seeing only what their role permits.

Honest outcomes

Never silent. Never passed off.

Every request ends in one of a small set of named outcomes, and the reply shows which one. If one is ever left unanswered, a person is paged at once.

  • Delivered
    Every check that applied ran and passed.
  • Revised, then delivered
    A check caught something, and the author fixed its own answer.
  • Asked back
    It needed one more detail, so it asked instead of guessing.
  • Held, with the reason stated
    Something should not go out. It says so, and says why.
  • Delivered, not fully verified
    If a check could not run, the reply says so on its face. It is never passed off as verified.
Kept promises

"I'll get back to you" becomes an obligation.

When the AI says it will do something, that promise becomes a tracked obligation that stays open until a receipt settles it. Approved work resumes on its own, even across restarts.

If critical work goes quiet, a person hears about it.

Blocked work sends a few honest status notes, never a stream of them. Critical work that goes quiet for too long pages a human. Your team should never have to chase the AI.

A batch cannot claim done on work it did not do.

Every item needs an output written after its work began. Planned must equal done plus failed plus skipped, or the run pages. Nothing is lost and nothing is rounded up to done.

Accountability your compliance team can verify.

Bring your security, compliance and AI governance stakeholders. We will walk them through the receipts, the checks and the controls behind them.

Request a briefing Visit the trust center