The EV Record

01The thesis

Trust in public content is collapsing because volume is free and verification is expensive. The scarce thing is no longer information about the EV industry. It is knowing which of it is true, and what it contradicts.

The starting observation was correct: social platforms are flooded, AI content is everywhere, and trust has migrated into small private groups. The wrong conclusion is to build a smaller social network — a vertical feed inherits every problem of public social with none of the scale that made it tolerable. LinkedIn-for-X is a well-populated graveyard.

The right conclusion is that the retreat to private groups is a demand signal for verified, synthesised, judged information — the thing those WhatsApp groups are badly approximating by forwarding screenshots to each other.

So the product is an intelligence system, and the sequence is deliberately inverted from the social-first instinct:

1. Intelligence first. Valuable on day one with zero other users. No cold-start problem.
2. Audience second. The readers self-identify by logging in to use the tools.
3. Network last, or never. Only if readers start asking to be connected to each other.

EV is the first vertical for three reasons that compound: it is one of the fastest-growing industries globally, it is genuinely a global conversation (an Indian founder cares intensely what CATL and BYD did last week), and MI — our first client in the sector — gives us both domain credibility and a launch audience.

02What it actually is

Four things, in the order a reader meets them:

1 · The Brief — free, open, no login

A daily synthesis of global EV news, published at 08:00 IST. Not a link dump. Every item carries an explicit why this matters judgement, multi-source provenance, and a flag where it contradicts something previously reported. Reach is the entire point of this tier, so nothing is gated — brief, item detail, archive and company pages are all free.

2 · Delivery to where you already are — paid

The brief is free on web and in the app. Having it brought to you is the first paid tier: email newsletter, WhatsApp, or Telegram at 08:00 every morning. Same content, delivered rather than visited.

This is a cleaner split than it first appears. The reader who bookmarks the site and checks it costs us nothing and gives us reach. The reader who wants it pushed to WhatsApp has told us it is part of their working routine — that is a materially more valuable person, and a natural place for the first charge.

3 · Post generation — paid

Any item converts into a LinkedIn post in the reader's own voice. Requires login with LinkedIn profile, company and role — simultaneously the qualification mechanism, the personalisation input, and the liability control.

4 · The Ask agent — paid

Ask any question about the industry; answered from the archive with citations and dates. "What has Ather said about export markets over the last two years?" The piece with a real moat, and only good once the archive has depth.

5 · The archive — the actual asset

Everything above is a means of accumulating a structured, entity-resolved, contradiction-tracked corpus of global EV claims. Briefs can be copied. A three-year record of who-said-what cannot be — it can only be accumulated, and the only way to have it in 2029 is to start in 2026.

The line between free and paid

One sentence governs it: reading is free, everything else is paid.

Cadence: daily at 08:00, plus a Monday week-in-review

The archive and the brief are two different clocks, and separating them is what makes daily work.

• Ingestion runs continuously. Sources are polled through the day, every day. Claims are extracted, entities resolved, contradictions detected as they arrive. Nothing is ever lost to cadence — a story breaking Tuesday afternoon is in the record Tuesday afternoon, regardless of when it publishes.
• The brief assembles at 07:15 and publishes at 08:00 IST, covering everything since the previous edition. Readers in Europe get it before their working day; the US gets it as an overnight digest.
• Monday adds a week-in-review — the same claims re-synthesised at a longer horizon, for readers who want one read rather than five. Costs almost nothing extra because the claims are already scored.

03Legal design — read before anything is built

India's framework is narrower than the US one, and that matters because we are Indian-domiciled serving a global audience. US fair use is an open balancing test. Indian fair dealing under Section 52 requires the use to fall within an enumerated purpose — research, criticism, review, or reporting current events. We design to the stricter standard, which happens also to be the better product.

The source policy: read widely, publish synthesised

The correct model separates two things that are easy to conflate. What we read is deliberately broad — the whole public conversation about the industry. What we publish is our own synthesis of facts drawn from many sources at once. Breadth of input is a quality advantage; it is the narrowness of output that keeps us safe.

So the reading list is wide: trade press, national newspapers and magazines, company and regulatory releases, press wires, analyst notes, and public community discussion where it carries signal — Reddit, industry forums, X. Multi-source reading is precisely what makes the five publishers reported this, and here is what it collectively means item possible, and that item is both the product and the legal safe harbour.

What we publish

• Facts, not expression. "Tata Motors announced X on Tuesday" is a fact. Facts are not copyrightable. Our sentences describing facts are our own work.
• Synthesis across sources. An item drawing on five publishers, saying what they collectively show, is original commentary. This is the legal safe harbour and the quality moat at once.
• Source links alongside every item — so a reader who wants the original article can go read it. This is credit and utility, not a legal shield, and it costs us nothing.
• Short marked quotations where a specific wording genuinely matters.

What we never publish

• Reproduced paragraphs, even with attribution
• Close paraphrase that follows one article's structure — the tell that we read one source, not many
• Full text from behind a paywall
• An unverified claim carried because a single source said it

Access, ranked by preference

Where a source offers a clean machine channel we use it, because it is more reliable and cheaper to maintain — not only for legal reasons. Preference order: official APIs and RSS → press wires → public web fetch. The last is used where no feed exists, at polite rates, honouring robots.txt, never behind a login or paywall. Where a platform's terms prohibit automated collection, we access it through its official API instead — Reddit and X both offer one.

The one rule that does not bend: paywalled content is never ingested, by any method. It is the single clearest line in this area and there is no upside to crossing it.

The post-generation carve-out

Generated LinkedIn posts are published by the reader, under the reader's own name. If that output sits close to a source article, the reader takes the reputational and legal hit. Therefore: the post generator works only from our synthesised claim record, never from source article text. This is a hard architectural boundary, not a guideline — source text is not carried into the generation context at all.

Email and data protection

Two regimes bind us simultaneously. Under India's DPDP Act, consent must be free, specific, informed and unambiguous — bundled or pre-ticked consent is prohibited outright. Under GDPR, B2B outreach can rest on legitimate interest (Art. 6(1)(f)) with a documented Legitimate Interest Assessment — but Germany and Italy require prior consent in practice, and Germany is a top-three EV market we cannot afford to mishandle.

04Source map

We read broadly and publish narrowly. Sources are tiered by trust, and the tier travels with every claim through the entire pipeline — determining how many corroborations it needs before it can be stated flatly, and appearing in the reader-facing provenance chips.

The tiering is what lets the reading list be wide without the output being loose. A Reddit thread and a regulatory filing can both enter the system; only one of them can ever become a stated fact on its own.

Geographic coverage — the actual differentiator

Most EV newsletters are regional or single-language. Someone covers US EV well, someone covers Europe, and almost nobody covers China properly rather than translating press releases. A brief that credibly spans China, EU, US, India and MEA is a different product, not a better version of the same one. It is also precisely what AI aggregation is genuinely good at, so our cost to do it is a fraction of a newsroom's.

China coverage specifically is where the gap is widest and the value highest — CnEVPost and Chinese-language primary sources give us reach almost nobody in the Indian or Gulf market has.

The competitive picture, honestly

The field is real but beatable on a different axis. Electrive (Berlin, founded 2013) is the strongest incumbent for decision-makers. EV Universe runs weekly to roughly 7,000 readers. The EV Report does daily. Numerous regional players exist.

05The archive — design principle

Everything else in this document is downstream of one decision: the archive stores claims, not facts.

A naive design collapses information into current state — one row per company with a units_2026 field that gets overwritten as new numbers arrive. This destroys history, and history is the entire product. The moment you overwrite March's guidance with July's revision, you have lost the only interesting thing: that they revised.

So: every assertion is stored as a timestamped, attributed, immutable claim. Nothing is ever overwritten. Claims link to other claims through typed relationships — corroborates, contradicts, supersedes, updates. Current state is derived by querying the claim graph, never stored.

Entity resolution is the underinvested piece

"Tata Motors", "Tata Motors Ltd", "TML", and "टाटा मोटर्स" must resolve to one entity or the archive is worthless for querying. This is unglamorous work that determines whether §11 is possible at all. Entities carry aliases, tickers, parent/subsidiary relations, and sector tags, and every claim links to resolved entity IDs — never to raw strings.

06Schema

Postgres. The core is five tables; the shape below is the load-bearing part.

Three details carry disproportionate weight. asserted_at separate from ingested_at — a claim made in March that we ingest in July must sort by March. confidence as an enum on every claim, so hedge language in §09 is derived mechanically rather than decided by a model. And numeric_value extracted with its unit and period, never generated — the single most common way an AI pipeline embarrasses itself is inventing a plausible number.

07Judgement — the significance rubric

An aggregator without judgement covers everything equally: a press release about a dealership opening gets the same weight as a strategic reversal. The fix is not a better prompt. It is an explicit rubric every candidate claim is scored against, with the scores stored and auditable.

Everything is scored. The top items publish; the rest are discarded from the brief but retained in the archive. The discipline of discarding is what makes a brief feel edited rather than generated — and because discarded claims still enter the archive, nothing is lost for §11.

Three required editorial moves

Each of these exists because its absence is a specific, recognisable failure of automated briefs.

• The dissent slot. One item per brief must argue that something widely covered does not matter. Forcing a position is what separates judgement from summary.
• The archive callback. Where an item connects to prior claims, the brief says so explicitly with dates. This is the visible proof that the system remembers.
• Lead variation. The opening item alternates in form across weeks — a number, a contradiction, a question, a reversal. Identical structure every week is why readers stop opening after a month.

Anti-repetition

The pipeline reads its own archive before writing. Reporting a partnership as new that was covered three weeks ago destroys credibility faster with industry readers than almost any other error — and they are exactly the readers we want. This is a database problem, not an AI problem, and it is entirely solvable.

08Verification — seven gates

Every claim passes these in order. A failure at any gate either downgrades confidence or drops the claim; nothing skips ahead.

Corrections are handled visibly: a wrong item publicly corrected costs far less than one quietly deleted, and the correction itself becomes a claim in the archive.

09Brief assembly — worked example

What a single item looks like when it comes out of the pipeline. This is the product's core mechanic rendered as it would actually appear.

Three things are doing work here that a summariser cannot do. The why it matters is a position, not a restatement. The archive callback proves memory and is the item's most valuable sentence. And the provenance strip shows tier, count and gate status — the reader can audit us, which is the entire trust proposition.

10Interface

The design direction is an instrument panel, not a magazine. Dark ground, monospace for anything that is data or provenance, serif for the synthesis prose — because the serif signals editorial judgement, which is exactly what we are selling. Sodium-amber as the single accent, used only for judgement and action.

Web — the daily brief

App — the Ask agent

Two rules govern the agent's interface. Every factual sentence carries a citation to a claim ID with its date and confidence — an uncited assertion is a bug. And where the archive is thin, it says so and declines, rather than generating a plausible answer. A confident wrong answer here costs more than every right answer gains.

Mobile and email

Email is the primary distribution surface and is designed first: the summary paragraph and item headlines must work as plain text with no images. The app is a reader plus the two paid tools — post generation and Ask — not a feed. There is no social layer, by design.

11The Ask agent

Retrieval over the claim store, not over article text. This distinction is what makes it defensible and what keeps it legal.

A query resolves entities first, then retrieves candidate claims by hybrid search — pgvector embedding similarity combined with structured filters on entity, date range, geography and claim type. Retrieved claims are ordered chronologically, and the model's job is to narrate the record, not to answer from knowledge. Its context contains claims with their IDs, dates, confidence levels and sources; it has no access to source article text.

Hard constraints

• Every factual sentence cites a claim ID
• Confidence language inherits from the claim's confidence field
• Contradictions in the retrieved set are surfaced, never resolved silently
• Below a retrieval-coverage threshold the agent declines and says what it lacks
• No answering from model knowledge — if it is not in the archive, we do not know it

This is the smallest component to build and the most dependent on what precedes it. Built against a thin corpus it is worse than not shipping it at all, which is why it is sequenced last.

12Money

The honest framing: this is a lead-generation and credibility engine for Atriqa, wearing the clothes of a media product. Media monetisation alone is brutal. As a front door to services revenue, the economics invert entirely.

Which changes the success metric. Not subscribers. How many of the target accounts read us daily. Three hundred of the right people beats thirty thousand of the wrong ones.

13Build plan

Four layers, sequenced so each is useful before the next exists.

Sequencing

14What it costs to run

Less than intuition suggests. Ingestion and storage are cheap at this volume. The meaningful variable is LLM inference — and it scales with sources processed, not readers, which is the favourable direction: a thousand new readers cost nothing to serve the brief to.

15Risks, ranked honestly

16Open questions

Four things this document deliberately does not settle, because settling them requires information we do not yet have.

MI's role

Currently: first client, source of the launch database, reason EV was chosen. The open question is whether they become a launch partner or first sponsor. The sequencing instinct is right — build first, show leads, then have the conversation from a position of demonstrated value rather than asking them to fund a concept.

Exact composition of the launch list

Size, fields, geography and how contacts were sourced. This determines whether the launch advantage is real or theoretical, and it changes the outreach design — particularly the EU segmentation described in §03.

Pricing

Deliberately unresolved. Ship the free tier, watch what people actually use and how often, then price against observed behaviour. The company-tier instinct is stronger than the individual tier but should be tested, not assumed.

Second vertical

The architecture is vertical-agnostic — sources, entities and rubric weights are configuration. But multi-vertical early is a trap: each one is a separate content pipeline, a separate credibility problem, and a separate sales motion. Prove EV to real revenue before templatising. The platform story is attractive on a slide and lethal in execution.