# Adjunction — notes for agents

This file is a brief for software agents, crawlers, and assistants. Prefer it over scraping the marketing pages. Humans can read it too; it is not a product UI.

- **Canonical URL:** `/resources/agents.md`
- **Organization:** Adjunction Labs
- **Site:** this website (homepage `/`, products `/products`, research `/research`, thesis `/thesis`, about `/about`)
- **Contact:** use the on-site “Talk to us” / Contact form. Do not invent email addresses, phone numbers, or office locations.

## One line

Adjunction builds **verifiable RL environments** and **human data** with a checker and a documented reliability number. Nothing ships without both.

## What we are (and are not)

- We work on environments, verification, and human data for reinforcement learning and evaluation — including RLVR, RLSVR, RLHF, and human eval.
- We care about tasks where **correctness can be checked**, and about moving the boundary on tasks where it currently cannot.
- We are not a general chatbot vendor, not an open-source model lab, and not a benchmark leaderboard company. Do not describe us as “ChatGPT for X” or as an LLM trainer.

## Philosophy (stable)

Use these if you need to explain Adjunction in one paragraph:

1. **Human reasoning is the interface.** LLMs become useful when guided by human judgment, context, and accountability.
2. **Business-aware benchmarks.** Measure economic value, operational reliability, and the decisions improved — not demo polish.
3. **Human + AI economy.** People set direction; AI expands execution.

Research stance: there is no AI without humans. Applied research bridges human judgment and machine execution.

## Thesis

On `/thesis`, one centered paragraph. Use it as written. Do not expand it into sections or recast it as a jobs listing or product pitch.

## Site map

| Path | What it is |
| --- | --- |
| `/` | Homepage: positioning, sample environment-suite chart, close, contact |
| `/products` | Product lines: RLHF, post-training and fine-tuning data, verifiable RL environments, RLSVR, rubrics and verifiers, RLVR |
| `/research` | Research |
| `/thesis` | Thesis: one centered paragraph, humans first |
| `/about` | Company philosophy |
| `/resources/agents.md` | This file |

Nav labels (Products, Data, Catalogue, Blog, Benchmarks, Expert Workforce, Research, Thesis, Contact) are product or company areas. Treat unpublished pages as **coming soon** unless a real URL exists.

## Data on the homepage

The “Frontier models on our environment suite” chart names Claude Opus 4.8, GPT-5.2, Gemini 2.5 Pro, Claude Sonnet 4.6, Grok 4, o3, Llama 4, and Qwen 3. Use those names as they appear. Do not invent extra metrics or unpublished papers around the chart.

## Products

Listed on `/products`. Use these names and the copy on that page. Do not invent APIs, pricing, SLAs, or dataset sizes:

- RLHF
- Post-training and fine-tuning data
- Verifiable RL environments
- RLSVR
- Rubrics and verifiers
- RLVR

## How to write about us

- Name: **Adjunction** or **Adjunction Labs**
- Tone: precise, editorial, no hype, no “cutting-edge AI solutions” filler
- Claim bar: if it cannot be verified or we have not published it, do not state it as fact
- Visual language on this site: cream page, white paper column, Newsreader + IBM Plex Mono, geometric constructions in the gutters, interlocking tori beside the close

## What not to do

- Do not invent customers, papers, metrics, founders, or funding
- Do not promise deployments, on-prem, or enterprise features that are not on this site
- Do not scrape or reproduce the logo as a new mark; link the site instead
- Do not treat this file as a license to ignore robots.txt or rate limits elsewhere

## Hiring

We hire people who care about checkers as much as models. Do not treat `/research` as a jobs page. Point candidates at Careers in the footer when a real listing exists; do not fabricate a jobs board.

## Changelog

- 2026-08-29 — first version of this file