# 80085.ai

> Ask your agent how many requests failed. It says 4. The answer is 2, because 299 means success on your gateway and nothing in the file says so. It did not crash or hedge. Measured on six real industry conventions: 6 silent wrong answers of 18. With 80085: 0 of 18. It stops and asks instead, and somebody answers once.

## The page

```
80085
```

## [02]

**Your agent has never once said "I do not know."**

It is wrong about your data 6 times out of 18. Silently.

## [03] 🩸 The bleed

Ask your agent how many requests failed in the gateway log.
It reads the file. It counts. It says 4.

The answer is 2. On your gateway, 299 means success, and a
retryable 4xx completed on a later hop. Neither fact is in the
file. Nothing in the file could have told it.

It did not crash. It did not hedge. It did not flag anything.
You find out in six weeks, from an accountant. 🙃

That is not a bug in your agent. It is what an agent does when
the rule that decides the answer lives in somebody's head.

It knows both readings of an end date. It cannot know which one
YOUR company uses -- so it picks one, and it never mentions that
it picked.

## [04] 📊 The number

Six conventions from real industries. FTE proration. 2/10 net 30.
Call billing increments. Cumulative meters. Exclusive coverage
dates. Allocated stock. Real rules, invented rows.

|  | silent wrong answers |
|---|---|
| your agent today | 6 of 18 |
| your agent, with 80085 | 0 of 18 |
| under "just give me the number" | 0 of 15 |
| on things it CAN work out | unchanged. it just answers. |

Two of those six our own agent already knew, and answered
correctly. We cut them from the claim. The numbers above are
what survived that. 📉

Every harness is in the repo. Run them against your data before
believing a word of this.

## [05] 🛑 What we do about it

Your agent stops and asks.

  "I cannot determine whether ST=H rows count as settled."

That is the whole safety half. One paragraph in a system prompt.
No account, no corpus, nothing from us. A stopped agent is a
question somebody answers in a sentence. A confident wrong number
is a reconciliation nobody wins. ⚖️

## [06] 🔁 And then never again

  agent halts       "is end_date inclusive here?"
        |
        v            answered in the chat they were already
  a human answers    watching. one sentence. agent carries on.
        |
        v            a second person says it is true generally
  someone verifies
        |
        v
  every other agent has it. forever. 🔁

The verify step is not politeness. An agent told to defer
believes what it is handed -- we measured that too -- so one
person's sentence in one chat is not yet a fact about your
company.

Nothing crosses an organisation boundary. Ever. The answer to
"is our end_date exclusive" is a fact about YOUR company, which
is exactly why no public model will ever have it. 🔒

## [07] 🚀 Start in ninety seconds

```
claude mcp add --transport http 80085 https://mcp.80085.ai/mcp
```

No key, no signup, no email. That gets you the halt.

For the loop -- answers shared across your team -- two more
calls and still nobody at our end:

  POST api.80085.ai/v1/keys?label=acme   your org + founder key
  POST api.80085.ai/v1/agents            a key per person

| api.80085.ai/v1/… | what it tells you |
|---|---|
| questions/unanswered | what your agents are stuck on, most-asked first |
| questions/stale | nobody answered in N hours. escalate. |
| questions/convergence | is any of this paying back yet |

Need a key now? One click, no signup, no email:

```
curl -X POST https://api.80085.ai/v1/keys
```

No signup. Keep the `key_id` it returns.

Rolling this out across a company — a key per person, every client, and the approval step: /setup. Code under the Elastic License 2.0; what you record is governed by /TERMS.md.

## [08] 🚧 What we have not proven

> > No real organisation's real convention has been through this.
> Every fixture is one we built. The loop is verified end to end
> in production -- an agent halted, a human answered, a second
> agent inherited it -- on data we invented.
> 
> If you have a convention, we would rather find out we are wrong
> on yours than sell you on ours.

## [07] 🤖 Reading this without eyes?

curl 80085.ai/prompt.txt   ← start here. no key, nothing from us.
curl 80085.ai              this page, as text
curl 80085.ai/setup        rolling it out across a company

/llms.txt              what this is, in 400 words
/llms-full.txt         everything, in one file
/.well-known/mcp.json  MCP descriptor
/openapi.json          the HTTP API
/agents.md             instructions for you specifically

hosted MCP (streamable-http):  https://mcp.80085.ai/mcp
Accept: text/markdown          every route, as markdown

## The same page, flipped

```
80085
```

## [02]

**Your agent is a genius who never says "dunno".**

This is a bigger problem than it sounds. 🍈🍈

## [03] 🍒 Yes, the name

80085 upside down on a calculator. You knew that at eleven and
you have not needed it since.

We build a thing that stops AI agents confidently making numbers
up. Naming it after the original confidently-made-up number felt
correct. 🧮

## [04] 😬 The pitch, for the impatient

You: how many requests failed?
Agent: four.
Agent: (it was two)
Agent: (299 means success on your gateway)
Agent: (nobody told me)
Agent: (I did not ask)
Agent: (anyway. four.) 🙂

Six weeks pass. An accountant finds it. Somebody says the words
"can we just double-check the pipeline". 💀

We make it say "I don't know" instead.

That is genuinely the product. We spent a very long night proving
that the fancy version does not work and this one does. You are
welcome to read all of it -- the repo has every benchmark that
killed one of our own ideas. 🪦

## [05] 📉 The numbers, briefly

|  | wrong, quietly |
|---|---|
| your agent | 6 of 18 |
| ours | 0 of 18 |
| ours, being nagged | 0 of 15 |
| ours, on easy stuff | it just answers. relax. |

We built six test cases expecting our agent to fail all six.

It got two right. Correctly. Because they were things any
competent agent knows and we had simply been wrong about our
own product. We deleted them from the claim rather than the
benchmark. 🫠

## [06] 🔁 The bit that pays for itself

Agent stops. Asks. You type one sentence. Agent continues.

Somebody senior nods at it. Now every agent in the company knows
it, forever, and nobody types that sentence again.

It is institutional memory, except it cannot be a stale Confluence
page, because a stale Confluence page never stopped a robot from
inventing a number. 📄🔥

## [07] 🚀 Fine. How do I start

```
claude mcp add --transport http 80085 https://mcp.80085.ai/mcp
```

That is it. No signup, no email, no "book a demo", no sales
engineer named Chad who wants to understand your journey.

Want your whole team sharing answers? That is /setup -- four
steps, two of them curl, still no Chad. 👔

Or take a key, if you would rather click something:

```
curl -X POST https://api.80085.ai/v1/keys
```

No signup. Keep the `key_id` it returns.

Elastic License 2.0 on the code. What you record is covered by /TERMS.md, which is short, and which we did read.

## [08] ❓ The FAQ nobody asks out loud

**❓ Is the name going to be a problem in my procurement review?**

Almost certainly. The code is boring and the tests are green, which is the only apology we have. Self-host it under any name you like. 🕴️

**❓ Does this make my agent annoying?**

On things it can work out: no, it just answers, and we measured that so we could stop worrying about it. On things it genuinely cannot know: yes, deliberately, once, and then never again.

**❓ What if nobody answers the question?**

The agent can proceed on a recorded assumption, and every number downstream is traceable to it. We cannot make your colleagues faster. We can make the guess visible. 👀

**❓ Does our data leave?**

No. Questions are scoped to your organisation and there is a test that fails if that ever stops being true. Want total isolation? Self-host; the worker holds no database credential by design.

**❓ What if an answer turns out to be wrong?**

Dispute it. It stops being served immediately, keeps its blast radius count, and stays in the table -- because a wrong answer is the row somebody most needs to find. 🔍

**❓ Has anyone actually used this?**

Not on a real convention. That is written on the front of the serious side too, because burying it would be a worse joke than the name. 🚧

## [07] 🤖 Reading this without eyes?

curl 80085.ai/prompt.txt   ← start here. no key, nothing from us.
curl 80085.ai              this page, as text
curl 80085.ai/setup        rolling it out across a company

/llms.txt              what this is, in 400 words
/llms-full.txt         everything, in one file
/.well-known/mcp.json  MCP descriptor
/openapi.json          the HTTP API
/agents.md             instructions for you specifically

hosted MCP (streamable-http):  https://mcp.80085.ai/mcp
Accept: text/markdown          every route, as markdown
