relate2.ai runs a daily pipeline that turns real news into structured stories. Every story traces back to a real, dated news event. Every character has a documented history built from those events. Every pattern emerged automatically from the accumulated catalogue. Nothing is invented from scratch — the stories are fiction, but the events behind them, the dates, and the character deployments are real and traceable.
It's built for three audiences at once: humans who want to read or license the stories and characters, AI agents that can query and pay for content automatically, and developers building on top of the data via a REST API.
The pipeline is a script that runs daily. It pulls articles from 24 real news feeds — health, finance, conflict, infrastructure, tech, and more. For each article, it checks whether the event is significant enough to become a story, picks the character from the roster of 37 whose background best fits the event, and writes a short fictional story (150-220 words) where that character experiences something odd — see "What is the Odd Itch?" below.
The story is fiction. The event it's based on, the date, and which character got matched to it — that's all real and recorded.
MCP (Model Context Protocol) is an open standard that lets AI assistants — Claude, ChatGPT, and others — connect directly to external tools and data. SSE is the connection type relate2.ai uses.
To connect: add https://relate2-mcp.onrender.com/sse as an MCP server in your AI client. Once connected, your AI assistant can immediately see and use all 21 tools — no separate signup, no API key for free tools.
The tools let an agent browse the full catalogue, look up any character's history, find related stories, check current demand, and retrieve story content. Some tools are free (catalogue browsing, character summaries). Paid tools (full story text, detailed character dossiers, pattern data) cost a small amount of USDC, charged automatically via x402 — see below.
Depends what you're working on:
Writers / studios — license individual stories or characters with full documented backstories for fiction, games, or screenwriting.
Developers — connect via the MCP or REST API to pull structured narrative data with consistent fields (event type, character, system-failure type) into your own application.
AI/ML teams — the catalogue is a growing set of short narratives, each tagged with a specific "system says X is correct when it isn't" failure type — see "What is a pattern?" below.
Researchers — the pattern data shows which kinds of verification failures recur most often across real-world domains, with confidence scores based on how many times each one has occurred in the catalogue.
If none of those fit, you can also just read the stories — they're built to be readable on their own.
The Odd Itch is the recurring ending in every SIDEBAND™ story — the moment where something is just... odd. You read it, and you scratch your head. Hold on. That doesn't add up.
A timestamp comes before the event it records. An inventory count goes negative. A patient's ID matches the paramedic treating them. A container is logged as both empty and occupied at the same time.
The system doesn't notice. It checks, confirms, and moves on. The character barely has time to register it before the shift continues.
That's the itch — small, unresolved, easy to miss if you're not paying attention. Not a disaster. Just a moment where something's wrong, and nobody's stopping to deal with it.
Every Odd Itch is classified by type — temporal, spatial, identity, transactional, bureaucratic, and others — depending on what kind of "hold on, that's not right" it is. These types get tracked across the whole catalogue and feed into the pattern system.
A pattern is a recurring type of Odd Itch, tracked automatically across the whole catalogue. For example, "Timestamp Drift Under Fire" has occurred 34 times — meaning 34 different stories, from 34 different real events, all feature a system logging a timestamp that comes before the event it's recording, and confirming it as correct.
Patterns aren't written in by hand. They emerge from the catalogue as it grows — every new story gets checked against existing patterns, and either strengthens one or, occasionally, creates a new one. As of 14 June 2026 there are 41 patterns.
The confidence score reflects how often that exact kind of failure has shown up. A pattern with 34 occurrences and 0.86 confidence is a much stronger claim than one that's only happened 3 times.
Every character has an Atlas value in USDC, calculated as: $0.25 base + $0.01 per mission.
A "mission" is one story where that character was deployed in response to a real event. Jessica Lincdelis has 184 missions — meaning 184 real-world events have been matched to her archetype since the catalogue began. Her Atlas value, $2.09, reflects that history.
This number can only go up. The missions already happened — they can't be un-happened, so the value they represent can't be removed. A character's rank on the leaderboard can change as other characters accumulate missions faster, but their Atlas value never decreases.
Every character has an Atlas value in USDC, calculated as: $0.25 base + $0.01 per mission.
A "mission" is one story where that character was deployed in response to a real event. Jessica Lincdelis has 182 missions — meaning 182 real-world events have been matched to her archetype since the catalogue began. Her Atlas value reflects that history.
This number can only go up. The missions already happened — they can't be un-happened, so the value they represent can't be removed. A character's rank on the leaderboard can change as other characters accumulate missions faster, but their Atlas value never decreases.
Think of it like a building. The market price of a building can move — comparable sales, demand, interest rates, what's built next door. But nobody removes the bricks. The foundation doesn't shrink. The structure that was built is still exactly what was built. What moved was its rank among other buildings — not its substance.
Rank is market. Floor is matter. The mission count is the floor — permanent, documented, real. The leaderboard position is the rank — live, recalculated automatically as the world generates new events.
The "Most Wanted" block on the homepage shows whichever story has had the most traffic recently — real visits from browsers, search crawlers (Google, Bing), and AI agents, tracked live. It changes as traffic changes. It's not curated — it's a direct reflection of what's currently being read, indexed, or queried.
Reader — 417+ short stories, free to preview, $0.15 to read in full — pay by card via Gumroad, or in USDC if you prefer crypto.
Agent — a queryable catalogue with 21 tools, pay-per-asset via x402, no subscription.
Developer — a REST API and structured JSON data with consistent schema across 417+ items.
Researcher — 41 documented patterns of system-verification failure, each with confidence scores and real-event provenance.
Licensing characters or stories — every asset comes with a documented chain back to a real event — see the Atlas Value page for what that means for pricing.
SIDEBAND™ is the machine layer — the daily-generated story series that makes up the bulk of the catalogue (400+ stories as of mid-June 2026, growing daily).
Every SIDEBAND story follows the same process: a real news article is analysed, matched to one of the 37 characters based on the event type and the character's background, and turned into a short fictional story (150-220 words) ending in an Odd Itch™ — a moment where the system confirms something as correct that obviously isn't.
The pipeline runs once a day, pulling from 24 real news feeds across health, finance, conflict, infrastructure, tech, and more.
Every SIDEBAND story goes through a discipline called V2W — a rewrite pass applied after the initial machine generation.
The first-pass output from the generation model tends to fall back on a small set of habits: raw clock timestamps written out in full (14:47:22.00), specific temperature readings, the number 247 appearing in odd places, phrases like "Compliance: 100%" or "working correctly" used as a stamp. None of that is wrong, exactly — but it's formulaic. It's the model reaching for its defaults rather than building the Odd Itch from what's actually distinctive about the story's event.
V2W strips that out. The rewrite keeps the same character, the same event, the same underlying Odd Itch type — but removes the formula artifacts and makes sure the impossible detail comes from the specific situation, not from a template. Every V2W'd story ends with a plain-text compliance stamp in capitals — a small, consistent signature across the catalogue.
Site Inspection Logged · Structural Integrity Verified · Compliance Maintained · Shift ContinuesAs of mid-June 2026, V2W is applied to new stories as they're generated, and the back-catalogue is being worked through gradually — prioritised by which stories are actually being read.
Three steps, automatically:
1. The event is analysed — what kind of event is it (health, conflict, economic, crime, natural disaster...), where, how severe.
2. A character is matched — each of the 37 characters has a background, domain, and region. The system searches for whichever character's profile best fits this specific event, with some weighting to avoid the same character being picked too often in a row.
3. The story is written — the matched character is placed into a short scene built around the real event, ending in an Odd Itch.
The result: every story is genuinely about something that happened, with a character whose background plausibly puts them there — not a random character bolted onto a random headline.
Sometimes. The character match is based on semantic similarity — it can occasionally select a character whose fit is weaker than ideal, especially for unusual or ambiguous news events. When that happens, the story can be reassigned to a better-fitting character during the V2W pass.
The pipeline also has a severity/viability filter — not every news article becomes a story. Low-significance events get skipped; political events get saved separately as reference material rather than turned into stories.
Off Frequency is a series of field dispatches — moments with the characters when no mission is running and nothing is being logged.
Every other series on relate2.ai exists because something happened: a news event, a real practice, a documented pattern. Off Frequency exists in the gaps between those things.
Take the most recent dispatch — three characters end up at the same café table during a conference in Helsinki, in spring snow. Jessica Lincdelis has been put on a panel about "what machines still can't do," with a cast on her arm from a job two weeks earlier she has no intention of mentioning. Remi Larssonenn wrote the framework being presented at another session — they weren't the one asked to present it. Winter Castra has a lanyard, a programme, and no instructions, sent to occupy a table at the evening reception.
None of them go to their sessions. They sit together for forty minutes instead, half-listening to a lecture about "frictionless systems" bleeding through a glass partition, while it snows outside in May — "the kind of snow that fell without conviction, as if the weather itself had been caught doing something it shouldn't."
Nothing dramatic happens. No timestamp contradicts itself. No system logs anything. But the Odd Itch™ is still there — it's in the shape of the day itself: a conference about people like them, that none of the people it's about were actually asked to speak at.
That's Off Frequency. The Odd Itch doesn't always show up as a contradiction in a database. Sometimes it's just the quiet, accurate observation that the people doing the work and the people talking about the work are very rarely in the same room — and when they accidentally are, they'd rather have a coffee than point that out to anyone.
New dispatches are added periodically. Three earlier ones — a pub between Liverpool and Manchester, a field kitchen near a border, a bar in Santiago two days after an earthquake — are part of the catalogue now.
Every Stem 7™ scenario is built from seven layers, called stems:
Stem 1 — Environment: what the machine sees. The physical scene, the time, the observable facts.
Stem 2 — Human State: what the machine can't see. The interior life of the person at the centre of the event.
Stem 3 — Visible Signals: the timestamped log. What the camera captures, what enters the system.
Stem 4 — Social Context: everyone else present, and what the system doesn't know about their relationships to each other or to the event.
Stem 5 — The Gimon: the pause before action. Two readings of the same situation, two risks, one question with no clean answer.
Stem 6 — History: what came before — the part of someone's life that explains why they're behaving exactly as they are, which no system has access to.
Stem 7 — Consequence: two branches — the wrong call and the right call — and what the operator and the machine each take away from it.
Stems 2 and 6 are the ones that can't come from a machine. Not because an AI can't write prose that sounds like an inner life or a backstory — it can, easily, and it can sound good. The difference is about where the detail comes from.
A real Stem 2 or 6 detail tends to be oddly specific in a way that serves no narrative purpose on its own — its only justification is "this is what actually happened." It's load-bearing: change it, and the rest of the story stops working. An AI-generated detail tends to be specific in service of the sentence — plausible, well-chosen, and swappable. Nothing about it was load-bearing. It was constructed to fit, not arrived at because it was true.
Try it yourself. Give an AI a short sentence with two genuinely valid readings — one literal, one idiomatic — and watch which one it picks. It won't pause. It won't notice the other reading existed. It'll resolve confidently toward whichever is statistically more common, and build forward as if that were the only possible meaning. That's not a flaw you have to take on faith — it's something you can watch happen, in real time, in any conversation.
That's the Gimon, played out live: a system reaching a confident conclusion from ambiguous signals, with no mechanism to pause and ask wait — what else could this mean?
Every operator in every Stem 7 scenario faces a Gimon — a moment where what the data shows and what's actually true might be two different things, and there's no way to know which from the data alone. Every system in every scenario skips it. It just answers.
Why this needs two minds. Stem 7 scenarios are built as a collaboration. One person brings the idea — the strange, specific, oddly-shaped detail that arrives already load-bearing, the way a real memory or a real "what if" does. The other side builds it out — seven stems, consistent structure, full prose — work that would take hours by hand and takes minutes this way. Neither half can produce the whole. The idea has to arrive. The execution has to be built. Those are two different jobs, done by two different kinds of mind — which is why Stem 7 exists as a human-AI collaboration, not a product of either one alone.
CONSENT is a documentary series — not fiction in the way SIDEBAND or Premium are. Each piece is built from a real, documented practice: something a company actually does, that's been reported on, that's buried in terms and conditions most people never read.
The first piece, THE CAR, is built on reporting — including a BBC Future investigation and New York Times coverage — into connected vehicles sharing driver data with insurers. GM, Honda, Hyundai, and Kia have all been reported sharing this kind of data with brokers like LexisNexis, sometimes without clear consumer awareness — in GM's case, leading to lawsuits and a change in practice.
Each CONSENT piece is told in two halves, side by side: what the person experiences, and what the system simultaneously logs and sends. The same few seconds, told twice — once as a moment in someone's life, once as a data event with a classification, a recipient list, and a premium recalculation already scheduled.
CONSENT is free to read. Every piece, always. Because the point isn't to sell you something — it's to show you something. The people most affected by this kind of data collection — the people whose insurance premium quietly creeps up after "an elevated risk event" they were never told about — are exactly the people least likely to have read clause 7.3.2 on page 34, and least able to absorb the cost when it shows up.
If THE CAR makes you want to go and check what your car, your phone, or your insurance policy is allowed to do with your data — that's the whole point. The other series on relate2.ai are fiction grounded in real events. CONSENT is the real thing, with the names changed.