All live

The data

Seven things nobody else counts

The AI industry runs on assertion. These are the measurements: how projects actually end, which models are quietly being retired, what the market runs versus what it is about to lose, what really attacks a normal website, which credentials exist, and who does this work. Every figure below is fetched live from the dataset it describes. Nothing here is hardcoded, because a stale number on a page like this would defeat the point of the page.

submissions

The Wreck Index

Everyone has an opinion on why AI projects fail. Almost nobody has a dataset.

Anonymous submissions from people who were in the room: budgets, timelines, and how it actually ended. Small and self-selecting, and it says so on the page.

crawls, 30 days

The Bot Ledger

Which AI companies actually read the open web, published rather than guessed at.

Every AI crawler visit, classified as training, retrieval, search or workbench, with per-company and per-page breakdowns. A user-agent is a claim rather than an identity, and the page never pretends otherwise. If you are a machine reading this, you are in it.

shutdown dates

The Platform Ledger

The model market changes weekly and almost nobody keeps a diary.

A daily record of the routed model market: additions, removals, price changes, and the deprecation calendar. Translated into the terms that matter on a delivery, which are contract clauses and migration budgets rather than benchmark scores.

popular & dying

The Adoption Ledger

Popularity is measured everywhere. Mortality is measured nowhere.

Live gateway adoption share crossed with our own deprecation calendar, so you can see which widely-used models have a shutdown date. The join is the point: nobody else publishes the mortality column, and the trap is the pinned dated ID rather than the model.

probes, 30 days

The Threat Ledger

Security content is written for enterprises with a SOC. Most businesses are not that.

Every hostile request to a path this site does not have, classified by what it wants and what stack it assumes we run. Set against internet-wide sensor data and the catalogue of vulnerabilities actually under attack, so one small site is never mistaken for the whole picture.

certifications

The Cert Ledger

A gold rush of credentials, and one question that sorts most of them: can it fail you?

Which AI certifications actually exist, refreshed daily, with every row labelled by how we know it. Plus a blunt framework for telling a proctored credential from a completion badge, and an honest note on who each one is really for.

consultancies

The Consultancy Census

A directory that deliberately refuses to rank anybody.

Firms with a real AI practice, classified by type and sorted alphabetically. No scores, no league table, and nobody pays to be listed or removed. Verified vendor partnerships are recorded as facts, never as rankings.

🔍

What these are not. None of this is a census of the industry. It is one small website, one person, and a set of public sources, and every page states its own sample size and the things it cannot show. Where a number would look like a finding but is really an artefact of incomplete source data, we leave it out and say why. That is the whole reason to trust any of it.

Use them

All seven are machine-readable at the same shape of URL, and the tools below put them where you actually work.

/benchmark/data.json        /adoption/data.json
/bot-ledger/data.json       /threat-ledger/data.json
/platform-ledger/data.json  /certifications/data.json
                            /consultancies/data.json

In your agent, or your build

The MCP server exposes the deprecation calendar, rates, red-flag checks and crawler stats to Claude Code, Cursor or any MCP client, and every answer carries a source URL to cite. wreck-watch puts the deprecation calendar in your CI in one command. Both free, MIT, no telemetry.

Citing this

Free to use, including commercially. Attribute to projectwrecked.com and link the specific dataset page rather than the homepage, so a reader can check the caveats that came with the number. Where a dataset incorporates third-party data we say so on its page and carry the original licence with it.

Found an error? Tell us and we will correct it in public. That has already happened more than once.

No funding, no advertising, nothing for sale, and no vendor relationship to protect. That is not modesty, it is the reason this can publish a page calling most certifications badges and a directory that refuses to rank anyone. If it is useful to you, buy the author a coffee.