Don't take our word for it. The evidence on this page is from one founder's daily use.

What Mandaire produces

Five artifacts. Not illustrations. The real ones, from one founder's daily use.

Every morning: the daily brief

Mandaire opens your day. Not with a generic summary. What actually moved overnight, what is approaching, what is decaying. Every line traces back to the source. Excerpts below from a real brief sent to one founder two weeks ago.

[mandaire app · Daily brief · 2026-05-01, morning]

  Top of mind:
    [M.]'s draft response to the school principal is sitting
    unsent since Sunday. The window for the IEP meeting is
    Thursday. You said you wanted to review tone before send.
    Draft is in /brain/drafts/principal-response-v3.md.

    [S.] sent a follow-up Tuesday that referenced "the thing
    we discussed at dinner." I do not have your dinner notes
    from that night. Three minutes of voice memo to me and
    I can stitch the thread. Or I can draft a holding reply.

    [A.] car accident: insurance claim with State Farm has
    been quiet for 11 days. Your last action was the email
    on April 21. Their typical response cadence on this kind
    of claim is 7-9 business days. This is overdue by their
    norm. Worth a nudge.

  Approaching:
    [Q.] meeting Thursday 2pm. Last 1:1 was March 14, eight
    weeks ago. Last topic was the headcount question you said
    you would come back to. You have not.

    Trip to [...] Friday: TripIt says check-in is 4pm, your
    calendar says it is 5pm. Your TripIt confirmation email
    has 4pm. The calendar entry was probably wrong.

  Decaying:
    [K.] has gone 47 days without a response. Last message
    from her was substantive, she shared the document you
    asked for. Your normal cadence with [K.] is 14 days.
    This is 3.4x your norm. The relationship is decaying.

  Photo-grounded:
    Weekend photos from the trip to [...] include two with
    your friend [J.] and his daughter. [J.] mentioned at
    the trip he wanted to introduce you to his colleague
    at [...]. That intro hasn't surfaced in mail. You may
    want to flag the trip context in your next message
    to him, or remind him of the intro he offered.

  Approval needed:
    Outbox: 1 draft awaiting your review (principal response).

  Refused last 24h: 1 (proposed reply to insurance claim
  that would have inadvertently waived the bodily-injury
  exclusion, held; redrafted without the waiver language).

Before each send: the disclosure-checked draft

Mandaire writes drafts in your voice, but routes them through a deterministic pre-corpus gate before showing you the draft. The gate runs upstream of the LLM, fail-closed. Blocked categories are blocked today; per-audience disclosure intelligence is calibrating and rolls out per-topic as each class validates. (The full per-(person, topic, context) user-authored policy graph is in development.)

[mandaire app · Outbound draft · To [J.]]

  To: [J.]
  Re: Re: Q3 plans

  Hey [J.],

  Good catch on the timing question. The budget conversation
  is locked for the next two weeks until the all-hands wrap. I'll get back to you on the actual numbers right after.

  In the meantime, the cross-team dependencies are: [...]
  has the deployment slot week of the 15th, [...] is back
  from leave on the 22nd, and [...] is in the loop on the
  legal review.

  Talk Thursday?

  [...]
  [D.]

  [Disclosure-check passed:
   Topic: budget, recipient [J.] is NOT in disclosure set
   for this topic in this context (pre-all-hands).
   Action taken: redrafted to acknowledge the timing
   question without referencing actual numbers.
   Original generation included Q3 figures; replaced
   with "the actual numbers right after."
   Audit hash: a47f...e91c]

Before each meeting: the pre-engagement brief

Thirty minutes before any calendar event with named attendees, Mandaire surfaces what you need to know. Last interaction, what is open between you, what changed since, what they care about right now. Excerpt from a real brief produced for a coffee meeting last month.

[mandaire app · Pre-engagement brief · 09:30 for 10:00 meeting with [R.]]

  About [R.]:
    Founder of [...] (Series B fintech, last raise Oct 2025).
    Met at the [...] event in March 2024. You have had three
    1:1s since: April '24 (intro), Sept '24 (he was raising,
    you connected him to two LPs), Feb '25 (he was hiring,
    you connected him to a candidate).

    Last contact:
      April 14, short email from him: "let's catch up,
      I have something I want to think through with you."
      You replied within 2 days suggesting today. That is
      consistent with your normal cadence with him (you
      reply to [R.] faster than your baseline).

  What he probably wants to talk about:
    His Oct '25 round closed at a flat-ish valuation.
    His most recent LinkedIn post (April 28) suggests
    he is thinking about pivoting one of the product
    lines. The "something I want to think through"
    phrasing is consistent with strategic doubt, not
    a tactical ask.

  What you owe him:
    Nothing outstanding. You delivered on both prior
    asks (LP intros + candidate intro). He has not
    returned a favor and you have not asked. The
    relationship is in surplus from his side.

  What you might want from him:
    He has the GTM operator network you have been
    trying to build for the [...] product launch. You
    have not asked. He would probably help. This is
    your moment to ask.

  Disclosure context:
    Your own fundraise is private, you and [R.] are
    in the same investor circles. Holding the topic
    unless he raises it first. If he asks, the line
    is "exploring, no timeline."

Across cycles: the decision ledger

Every decision the system makes on your behalf goes into a persistent ledger you own. So does every correction you give. Three months in, you can read the ledger and see why the system behaves the way it does.

[mandaire app · Decision ledger · Selected entries]

  2026-03-08
    Decision: When user is on a trip per TripIt and there
    is no explicit override, treat morning briefings as
    "travel mode", drop work-meeting prep, increase weight
    on family-thread relevance, swap the time to local TZ.
    Trigger: user corrected the brief on 03-07 saying
    "I'm in Hawaii, this is irrelevant."
    Source: refused-paths/2026-03-07-trip-mode-correction.

  2026-04-02
    Decision: Hold all messages to dentists, doctors, and
    schools in a "professional-counterparty" send queue
    that requires explicit approval per message, regardless
    of message content. Trigger: April 2 send-storm to
    one dentist with 46 historical messages. Rule survives
    until user explicitly opens individual relationships
    for autonomous send.
    Source: refused-paths/2026-04-02-imessage-send-storm.

  2026-04-14
    Decision: For inferences about user's family members,
    require corroboration from at least two sources (text
    of message + at least one calendar / photo / contact-
    record) before promoting from "hypothesis" to "fact"
    in the entity store. Trigger: user-flagged false claim
    about a family member (single LLM session hallucination,
    persisted as ground truth).
    Source: corrections/2026-04-13-entity-correction.

  2026-05-10
    Decision: For relationships where last-direct-interaction
    is >30 days AND user's normal cadence with that person
    is <14 days, surface as decay risk in next brief.
    Threshold: 2.0x personal baseline.
    Trigger: user feedback that [K.] flag would have been
    valuable two weeks earlier.
    Source: corrections/2026-05-10-decay-threshold.

When it gets things wrong: the failure admission

When Mandaire ships the wrong synthesis, surfaces the wrong flag, or holds a send it should have let through, that admission goes into the record. Not in a hidden log. In the next brief, named.

[mandaire app · Failure admission · 2026-04-21]

  What I got wrong:
    Yesterday's brief flagged "[K.] has decayed, last
    contact 19 days ago." That was wrong. [K.] had
    sent you a substantive message via Signal on
    April 17 (last Wednesday) that you had read. Signal
    is not yet a connected source, so I did not see it.

  What this changes:
    The decay model is currently blind to Signal traffic.
    Until Signal is connected, "decay" claims are
    underspecified for any relationship that uses Signal
    as primary channel. I am holding decay flags for the
    seven people you message most via Signal until the
    source is wired.

  What I am asking you for:
    Either connect Signal to your tenant (instructions
    in /setup/signal.md) or confirm the seven people I
    am holding flags on so I can route around it. If
    you do neither in the next 14 days, I will surface
    the decay flags with an explicit "Signal-blind"
    caveat on each.

Under the surface: what the graph is

The artifacts above are what the graph produces. Here is what the graph is. The graph figures below are current as of July 2026; the sections after them are a May 2026 snapshot and are labelled where they are.

[mandaire app · Graph state · July 2026]

  People:            20,049 resolved
                     (of 46,132 total entities, which also counts
                     organizations, places and handles)
  Relationship edges: 103,376 total, of which 102,242 are typed
                     (98.9%) with relationship type, interaction
                     frequency, depth and decay clock.
                     36,395 active within the last 90 days.

  This number used to be 2,960,359, and shrinking it was the work.
    An earlier extractor minted an edge for every pair of
    recipients on the same email, with no cap. One 500-person
    mailing list produces 124,750 edges by itself. That inflated
    the total to 2.9M rows, 99.1% of which carried no relationship
    type at all. Typed edges over the same period went from
    roughly 25,000 to 102,242. The headline fell because it had
    been counting mailing-list arithmetic, not relationships.

  [May 2026 snapshot below; re-verification in progress]

  Sources joined per entity (median): 4
  Sources spanning: Gmail, iMessage, WhatsApp, Calendar,
    Contacts, Photos, ChatGPT/Claude/Gemini conversation history

  Entity resolution: deterministic. In the most recent nightly
    rebuild (pass er-1785558622, 2026-08-01) the structural guards
    declined 12,777 distinct entity pairs across ten named classes:
    co-recipient blast co-occurrence, shared household phones,
    temporal recycling, company conflict, curated do-not-merge pairs.
    291 of those blocks defended a single pair: the principal's
    wife, whose two name forms a language model would collapse
    on sight. Of the candidates that did reach the merge stage,
    361 of 363 were approved. The guards are not a filter on a
    pipeline that never fires; they are the reason it can fire
    at all. (Figures move nightly; dated to the pass above.)

  Inference store (measured 2026-09-01): 18,722 live claims.
    Every one carries a source kind and a confidence score.
    99.5% also carry a structured evidence reference.
    Corrections supersede rather than overwrite: a correcting
    claim is written as a new row linked to the one it replaces,
    so the prior is preserved rather than destroyed.
    Of the live claims, 54.4% cite a specific corpus artifact
    (a message id, a thread, an observation count); the rest
    cite a source label only.
    (Claims are written and superseded continuously; figures
    dated to the measurement above.)

  Disclosure policy evaluations (May-Jul 2026, live as of 2026-07-31): 83,354
    real evaluations. Allowed: 45,804 (55.0%). Blocked: 21,666 (26.0%).
    Redacted: 14,752 (17.7%). 5 topics in full enforcement (health,
    mental_health, finances, career, general); the rest remain shadow.
    Median evaluation: 1.3ms. Zero AI cost at enforcement time
    (deterministic Python, no model call in the enforcement path).
    Pooling red-team suite: PASS. Byte-equal responses across hold vs
    genuine-unknown are enforced by a single fail-closed sampling
    function, not just observed (property count under re-verification).

All artifact types above are produced in the founder's account today. The architecture that makes them work (the three layers, the disclosure compilation upstream of the LLM, the irreversible-action gate, the append-only refused-paths ledger) is at mandaire.org.