What it looks like

01Design principles

Five rules govern every screen. Each exists because breaking it would undermine the thing we are actually selling, which is verified judgement.

1 · Provenance is always visible

Every claim carries its tier, its corroboration count and its gate status, on the surface, in the reader's eye-line. Not hidden behind a tooltip. The reader must be able to audit us at a glance — that is the entire trust proposition, and burying it would be self-defeating.

2 · Judgement is typographically distinct from fact

Facts are set in the body serif. Why it matters sits in an amber-railed block. The reader always knows which sentences are the record and which are our opinion. Blurring the two is what makes AI content untrustworthy; keeping them visibly separate is the fix.

3 · Amber means judgement or action, nothing else

The single accent is spent only on our opinion and on the things a reader can do. It is never decorative. When the eye catches amber, it has found either a position we are taking or a button.

4 · Memory is shown, not claimed

Archive callbacks — this supersedes what was reported in March — are rendered as first-class interface elements with dates and claim IDs, not as prose asides. This is the feature no competitor has and it should look like a feature.

5 · Reading is free; everything else is paid

No truncated items, no "read more" teases, no blur-and-upgrade. The brief, item detail, archive and company pages are whole and free on web and app. The wall sits where the reader wants the brief brought to them — email, WhatsApp, Telegram — or wants to act: generate a post, query the archive. Reach is the business model; convenience and action are the products.

02The system

Six colours, three typefaces. Deliberately tight — a constrained palette is what makes provenance chips legible at 9px.

03The brief — web

The front door. Free, no login, fully readable, published 08:00 IST every morning. Roughly 60% of readers will only ever see this screen, and it has to be good enough on its own that they come back tomorrow.

Note the hierarchy. Item 01 gets the amber wash, the full judgement block and the archive callback because it scored 84. Items 02–04 are compressed to headline, one line and chips. The brief looks edited because it is edited — the layout is driven by the significance score, not by arrival order.

04Email

The first paid capability, alongside WhatsApp and Telegram. Same edition as the free web brief — the charge is for delivery into the reader's routine at 08:00, not for different content. Designed to survive plain text, since half of corporate inboxes block images.

Every email CTA is a login trigger. "Turn into a post" is the most-clicked element in the email and it lands on the sign-up flow in §06. This is how the reader list converts into an identified, qualified audience without a wall at the top.

05WhatsApp & Telegram

The same 08:00 edition, pushed into the channel the reader already lives in. In India and the Gulf this is where the audience actually reads — and it is the lowest-friction paid capability in the product.

Chat channels get the summary, not the whole edition. Nobody reads eleven items in WhatsApp. The push carries the day's synthesis paragraph and the headline list with contradiction flags intact — then links out. The judgement blocks live on web where there is room for them.

06The paywall moment

This is the most commercially important screen in the product, and the one most easily got wrong. It appears only when a reader tries to act — never while they are reading.

What this screen does not say: no "unlock", no "premium", no blurred content behind it. The reader has finished reading — nothing was withheld. The ask is framed as whose voice to use, which is a genuine product requirement, not a pretext. That is why it converts and why it does not feel like a wall.

07Sign-up — 40 seconds

The whole flow, in four steps. Nothing is asked that does not either improve the output or qualify the lead.

08Post composer

The first paid tool. Source pane on the left, draft on the right — so the writer can see exactly what the claim says while editing what they are about to publish under their own name.

The architectural boundary is visible here. The left pane shows our synthesised claim — never the source article's text. The generator has no access to publisher prose, so the reader's post cannot accidentally paraphrase a copyrighted article. The legal firewall from §03 of the build document is a literal UI division.

09Ask agent

The second paid tool, and the one with the real moat. It narrates the record — it does not answer from model knowledge.

Every factual sentence carries a citation. An uncited assertion is a bug, not a stylistic choice. The confidence word in each source line — stated vs reported — is inherited mechanically from the claim record, not chosen by the model.

What people will actually ask

• "Has anyone announced sodium-ion at volume, or is it all pilot lines?" — the question the dissent slot keeps answering
• "Which Indian OEMs have revised export guidance in the last year?" — needs the claim graph; impossible with a search index
• "What did we know about the Hungary plant before the January briefing?" — the archive as memory
• "Who supplies cells to Ather?" — entity relationships, resolved

10When it declines

The screen that protects everything else. Below a retrieval-coverage threshold the agent says what it lacks rather than generating something plausible.

This is a feature, and it should look like one. The refusal is specific — it says what exists, what does not, and why the near-miss is not good enough. A vague "I don't know" wastes the reader's time; this one tells them where the edge of the record is.

11Company pages — the archive made visible

Where the source-of-record principle becomes something a reader can look at. Every entity accumulates a claim timeline, and the timeline shows revisions rather than hiding them.

Read it bottom-up and the story tells itself: confirmed, quietly slipped, then vaguer still. No individual article says this. The record does — and only because nothing was ever overwritten. This is the schema decision from §06 of the build document, rendered.

12Mobile app

A reader plus the two tools. Deliberately not a feed — no infinite scroll, no notifications begging for attention, no social layer. The brief ends.

Four tabs, and one of them is the account. There is no discover tab, no trending, no follower count. The app is finishable — a reader opens it Monday, reads for four minutes, and is done. That restraint is a positioning choice: we are the opposite of the thing that broke trust.

13Full screen map

Everything that gets built, by surface and phase. Screens marked P2 ship with publishing; P4 waits for archive depth.

14What this interface deliberately is not

Each of these is a thing a product like this normally has, and each is left out on purpose.

• No feed. The brief ends. There is nothing to scroll into. A reader who finishes is finished.
• No follower counts, likes or comments. No social layer at all — that was the conclusion from the very first conversation and the interface holds the line.
• No engagement notifications. One push a week when the brief lands. Nothing else, ever.
• No truncated free content. Nothing is blurred, teased or cut short to force an upgrade.
• No personalised filter bubble. Follows reorder the brief; they never hide items. Knowing what the industry is doing is the point.
• No AI chat bubble in the corner. Ask is a destination with citations and a refusal threshold, not an assistant that will say anything.