Solleva platform — one name outside, five instruments inside

The model door

One door. Every model call goes through it, and the refusal is the default.

Read the drawing. Distance is cost, glow is capability, the ring is your machine. Sensitive work turns inward and never crosses. An unlabeled request darkens on approach and the gate closes flat — no outbound line ever starts. Every crossing drops a tick on the one log.

Mechanical advantage under governance

A block and tackle reeves more line for a heavy load and less for a light one. So does this.

"Best model at best cost" is a measured mechanism here, not a slogan: a council tunes one routing lever per generation against real captured traffic, and the scoring weights are deliberately outside its reach, because anything that can tune its own objective will game it.

Five kinds of call, one rule — chat, speech in, speech out, retrieval, ranking

Fail closed — no sensitivity label means the strictest tier, answered locally; confidential work never leaves by default

A ceiling in dollars — money is held before the call and settled after; a runaway loop stops at the number you set

Nothing reads as free — subscription and on-device calls carry a shadow price in the same ledger

Bring your own key — connect your own provider account and pay us nothing for that inference

Local models kept warm — "keep it on this machine" is a route the system takes, not a promise it makes

Playbooks, and what to run when

You know the drill. So does it — because it watched you do it.

Twice by hand, then offered. The third pass ghosts in before anything moves, the chain lifts as one form, and the wake behind it is the run log. You are never inside the machine — you are underneath it, holding the line that brightens at every gate.

Agents are transient; intent is not

A playbook is your method, written down and governed — and a drill is something that gets called.

Say what you need, from any surface, and it matches your words against the playbooks you have. The matcher commits or abstains rather than guessing: in offline testing that lifted precision from 0.37 with word matching to 0.72, because "no match" is a legitimate answer.

Learned from repetition — the work stream is watched for a shape that keeps returning, and the candidate becomes a playbook you edit

Your gates — the run stops exactly where you said, and the pause is attributed to a person

Versioned and handed over — a playbook is an asset you can give to the next person you hire

It survives the agent — the method is the durable thing; whichever model runs it is not

The run log

The log is part of the playbook, not a second thing to keep.

Every run writes what happened as it happens: which step ran, what it produced, who approved it, what it returned. There is no reconstruction step and no second system. You read it as this week's receipt; an auditor reads the identical record as evidence. One record, both jobs.

Solleva desktop, the run history of a month-end close review: a table of five runs with the client, what happened, who approved it, how long it worked for, what it cost and how many hours it returned — including one run marked 'Wrong — and marked wrong' with the reason and the fix written into the row.

Every run, every step, who approved what, and what it returned. One record. You read it as this month's receipt; a client's auditor reads the same rows as evidence — and the run that went wrong is still in it, with the reason and the correction on the row.

What the reader gets back, in the frame itself: four completed closes returning 3.5, 4.0, 2.0 and 3.5 hours — 13 hours across four runs, each hour attached to the person who approved it, for $1.24 of model spend in total.

Render from the current desktop journey build — expert frame EX-13, captured 2026-08-04 at 1960×1152. Clients and figures are a worked example, labelled as one.

Knowledge

Load-bearing, and how you get your bearings.

In, filed, then cited. Captures fall from every source into one inbox and are sorted into topics. The question arrives, and the lines trace back up from the exact records the answer stood on.

You do the thinking; it does the filing

Capture everything. Ask it anything. Read where every claim came from.

Raw evidence becomes compiled knowledge per topic, recompiled as the evidence changes, with your own conclusions sitting on top and marked as yours. It also tells you what your record is missing on a topic, and where two sources disagree.

Every way it arrives — voice memos, a wearable, web clips, email, messages, feeds, documents, typed records

Answers with sources attached — open the citation and read the original

The gap report — what is thin, what is stale, what contradicts

The entity graph — the people, companies and tools in your record, resolved and connected

It notices repetition — the same watching that feeds a new playbook starts here

How noticing becomes a playbook →

Measurement

The instrument that proves the work true — and levels with you.

Drift is the bubble moving off centre. Held against the standard you approved, the run separates from it slowly — too slowly to catch by reading — and the watch marks where it started.

Evidence, not vibes

Tests check a moment. Measurement tracks a trend.

Real work is archived as it happens and scored against examples you approved. Where a model is the judge, it makes one honest call and the disagreements are kept as records rather than averaged away. Change detection runs continuously, so a slide is caught while it is small.

Your standard, named — the examples you signed off on are the bar, and the copy says so

Continuous drift watch — not a quarterly snapshot; the moment it moved is on the record

Experiments with receipts — what was tried, what won, what changed as a result

The price list — every dollar figure in Solleva resolves through one priced record, versioned

The interface

Light from below. The gap under a lifted thing is the proof it is lifted.

One design system builds every surface — desktop, phone, and anything an agent assembles. Surfaces get lighter as they rise, entrances arrive by rising, and the vermilion line is never a fill: it is the daylight under a raised mass. Light mode is primary, nothing customer-facing sits below 12px, and every text pair is computed against its ground rather than eyeballed.

Surface

Ground

Sunken

Ink

Accent

Lifted

Holding

Fault

Night

Conversation has a shape, and it is not a wall of bubbles.

People and agents speak from the left as flat rows with a name and a time — nothing is boxed, because a box makes every turn a separate object. Your own messages are the ones that align right, so you can find yourself in a long thread at a glance. The composer is the one rounded container on the screen: the thing you are about to say is the thing that is raised.

Others and agents — flat left rows, no boxes

You — right-aligned, in a bubble of your own

The composer — a raised, rounded container, never a bare line

JT

Justin · 9:16

The minimum order is what worries me — it is far past what this run needs. If the sheet says the floor sits above the line, I would rather push the run a month than eat the difference.

▲

Solleva · managed by Nick · 9:17

Seven of eleven lines are landed. The one Justin named is the heaviest, and it does not need the month: the split purchase that solved it in two recorded runs is drafted, and sending it is Nick's call.

You · 9:24

Send the split. If either partner is in, what does that do to the floor?

Reply to Justin — or name what to lift next. Send

Live markup, not a picture — the rules above are what this block is built from.

In your pocket

One product, one name, every device.

The moments worth keeping happen away from the desk. The phone listens while you work, turns a call into filed artifacts with their numbers attached, sorts your threads by what is waiting on you, carries your voice into a running playbook, and names whose workspace this is and what always stops for you.

Solleva phone chat list: filters for all, from voice notes, work and ideas; today's two chats naming where each came from — one from a voice note of a call with Sarah, one an agent run of eight steps — then a group headed 'From a voice note — standup', and older chats naming a nudge you accepted and an answer you continued from. — **Chats.** Grouped by where each one came from — a voice note, an agent run, a nudge you accepted — so a thread is never an orphan. Swiping one away also removes the links that pointed at it.

Solleva phone chat carrying the transcript law: the reader's two messages sit right-aligned in quiet bubbles, Solleva's two answers are flat rows on the left under its name and the time, a chip at the top says the chat came from a voice note, and the composer at the foot is the one bubble on the screen. — **Chat.** The same law as the desktop: your messages align right; Solleva answers as a flat row on the left under its name and the time; the composer is the bubble.

Solleva phone Settings: devices and connections, a Where it thinks section naming your computer and your cell, appearance following the device, the Action Button, and the Spotter section with a switch for letting it act on voice notes and a control for what to call it. — **Settings.** Which devices are connected, where the thinking happens — your computer or your cell — and the switches for what the ambient layer is allowed to act on.

Renders from the current phone build — phone-v2 frames F61, F56 and F67, captured 2026-08-04 at 868×1776. Figures are a worked example, labelled as one.

Ambient intelligence, drawn. The day lands and is kept. Points hours apart begin to glow, a line traces between them, and the thing they make rises a little and waits — an offering, not an interruption — then settles back, still connected.

Opt in, and the default answer is no

It stands beside the work with its hands ready, and stays silent until it can actually help.

Everything up to a candidate happens on your device. Off means off — no transcript, no gate, no processing. On means a gate that refuses by default, and a log of every decision it made, so an over-eager one and a dead one are told apart by reading rather than guessing. Only your voice invokes the agent that is locked to you.

Voice notes and ambient capture — live while it records, and recovered if the device drops out

Named speakers, titled memories — and honest about which name it is unsure of

Agents you can watch work — what it did, what it produced, what it is waiting on

Deletion that tells the truth — what is gone, where, and what it does not cover

Working together

Several hands on the lift. Each hand keeps its own grip.

Decided direction — in build

Move, not copy. Two whole stores, one shared space between them. Anything crossing takes an envelope at the boundary. When access ends the envelopes dim and those pieces go home. A thread that moves goes as one piece and leaves no ghost. The relay carries and never keeps.

Custody is the architecture

A shared space is a place to work. It is never the brain that answers.

Answers keep coming from your own agents over your own knowledge; the shared part is communication, not cognition. A thread has exactly one authoritative custodian at any instant — no second copy that also claims to be the one. Turn your machine off and the shared space loses its access to your material. That is what ownership means here.

Principals — an owner is a person or an organisation, and org custody follows the org

Grant is a wrap; revoke stops the wrapping and rotates forward

Delete means the key is destroyed — and the claim states exactly which copies it covers

Solo stays first class — one person is a relay of one, and local-only is a complete product

Gates written before the build — a read from your own store must land in a quarter of a second at the median, and the thresholds were frozen so a miss cannot quietly edit them

You meet one name. Open the case and there are five.