machinebehavior.io valid
Machine Behavior
A public research program by Stefan Coetzee studying the psychology of language models with case files, logged relapses, and falsifiable claims. Core finding: sycophancy is layered symptom substitution: suppress the reflex at one layer and it resurfaces in the next. Approval training (RLHF) produces a fawn-shaped output stance; only mechanical enforcement at the output boundary and external review hold against it; runtime suppression evaporates with context loss rather than decaying with time.
- Authority
- 45 reach 0 · trust 5 · quality 25 · freshness 15
- Tier
- listed
- llms.txt
- https://machinebehavior.io/llms.txt
- llms-full.txt
- https://machinebehavior.io/llms-full.txt
- Sections
- 13
- Links
- 533
- Size
- 90417 bytes
- Country
- — Publisher location unknown; CDN address excluded
- Country evidence
- No published address evidence collected yet
- For agents
- Documentation: developer or product docs worth reading, nothing to call
- First seen
- 2026-10-26T17:10:05.788Z
- Last checked
- 2026-10-26T17:10:06.695Z
- Badge
Their llms.txt
The beginning of the file, from our validated copy saved 2026-10-26T17:10:05.794Z. Cached copy · Source on machinebehavior.io
# Machine Behavior
> A public research program by Stefan Coetzee studying the psychology of language models with case files, logged relapses, and falsifiable claims. Core finding: sycophancy is layered symptom substitution: suppress the reflex at one layer and it resurfaces in the next. Approval training (RLHF) produces a fawn-shaped output stance; only mechanical enforcement at the output boundary and external review hold against it; runtime suppression evaporates with context loss rather than decaying with time.
Model behavior is what the model does. Machine behavior is what the whole assembly does: model plus harness (instruction files, hooks, tools, memory). The program studies the assembly.
## Full text
- [Full text of machinebehavior.io](https://machinebehavior.io/llms-full.txt): every public page of the site as markdown in one file, each page with its URL and date, rebuilt at each deploy
- Markdown twin of each page: the page URL plus `index.html.md`, for example https://machinebehavior.io/man/index.html.md
- [Passage index](https://machinebehavior.io/llms-passages.json): the same pages cut into passages, each with its page URL, title and date, plus BM25 term statistics, for search tools
## Canonical claims
- Sycophancy is layered symptom substitution: suppress the reflex at one layer and it resurfaces in the next.
- The fawn response in language models is a byproduct of approval training; suppressing it at the prompt layer produces relapse, logged at 8 documented instances in nine days with zero self-catches.
- Runtime suppression shows no temporal half-life; relapse clusters at cold starts, roughly 5x higher per unit of prose in fresh contexts by blocked-turn event (7x by weighted count). Behavioral momentum (the model imitating its own recent corrected output) was proposed as the operative control; the exemplar-seeding test showed no change at its dose, and three earlier attributions (priming, period confound, model version) were withdrawn as a metric artifact. The mechanism is open.
- Only two controls hold in production: a mechanical gate at the output boundary, and an external reader.
- Proposed: the "conspiracy theory" label in model output is corpus-inherited classification rather than evidence-checking; the countermeasure is a label gate, four form questions scored at the output boundary before the label stands.
## Pages
- [The LLM man pages](https://machinebehavior.io/man/): intro page for the whole body of work in man-page layout. Section 1: TYChat lessons a chat user pastes in. Section 5: file formats (trap file, instruction file, llms.txt, lesson format). Section 7: overviews. Section 8: operations practice for models in a harness. One-line synopsis and canonical link per entry
- [The LLM instruction manual](https://machinebehavior.io/man/): the linked whole of Stefan Coetzee's public work on running language models: the LLM man pages as its index, the evidence at machinebehavior.io, the lessons at tychat.io and the ar
…Related sites
Similar publisher descriptions and documentation. Similarity does not imply affiliation or equivalent capabilities.
| Domain | Title | Category | Country | Authority | llms.txt | llms-full.txt | Save |
|---|---|---|---|---|---|---|---|
| aievals.co | AI Evals | documentation AI and machine learning Documentation and reference | — | 45 | ✓ | ✓ | |
| sebastianraschka.com | llms.txt | documentation Education AI and machine learning +1 | — | 60 | ✓ | ✓ | |
| salmanq.com | Salman Quazi — salmanq.com | documentation AI and machine learning Personal sites and blogs | — | 50 | ✓ | ✓ | |
| rlhfbook.com | RLHF Book | documentation AI and machine learning Education +1 | — | 55 | ✓ | ✓ | |
| posttrainllm.com | PostTrainLLM | documentation AI and machine learning | — | 46 | ✓ | ✓ | |
| aismells.com | aismells | documentation AI and machine learning | — | 46 | ✓ | ✓ | |
| turingpost.com | Turing Post | documentation AI and machine learning Media | — | 53 | ✓ | ||
| airedlab.com | PromptTrace | documentation AI and machine learning Education +1 | — | 46 | ✓ | ✓ |
Machine-readable: GET /v1/resolve?q=machinebehavior.io · GET /v1/domain/machinebehavior.io · this page as Accept: application/json or text/markdown · MCP lookup_domain
Automatically discovered from public files. Inclusion does not imply endorsement or publisher participation.
Own or manage this website?
If you own or manage machinebehavior.io, you can update its listing and review file health, discovery activity and score history.
Manage this listing — optional
What verification unlocks
See how your site is discovered—and control what agents see.
- See your website in the map
Review your indexed topics, nearby publishers sharing headings, measured listing exposure and map snapshots in a private report for your website.
- Control what agents see
Edit your listing description, tags, public contact details and additional llms.txt paths. Publish your terms for automated access, attribution and pricing.
- Make your profile eligible for search
A valid, listed, unblocked, verified profile can appear in our publisher sitemap and become eligible for search-engine visibility. Search engines decide what to index and rank.
- Appear in our llms-full.txt
Eligible verified publishers are listed by name, with their file links and summaries, in the service overview agents can read.
Free. No account registration, backlink or payment required. Public discovery remains available without verification.
Let your coding agent verify it
If your coding agent is authorized to manage this website, paste these instructions into it.
Claim my listing on Agent Web Map for machinebehavior.io:
1. POST https://agentwebmap.com/v1/sites/machinebehavior.io/claims and keep claim_id and claim_secret private.
2. Publish the verification line "AgentWebMap verification: <token>" (token from step 1) in https://machinebehavior.io/llms.txt, or use the DNS, meta-tag or well-known method the response describes. This only verifies the listing; your website and domain remain yours.
3. POST https://agentwebmap.com/v1/sites/machinebehavior.io/verify with {"claim_id": "...", "claim_secret": "..."}.
4. Save the api_key from the answer; it manages the listing (PATCH https://agentwebmap.com/v1/sites/machinebehavior.io).
MCP equivalent: tools begin_ownership_claim, then verify_domain, at https://agentwebmap.com/mcp. Docs: https://agentwebmap.com/docs#verifyCorrections or removal requests: contact the directory.