Glossary.

The terms behind keeping AI agents in hand, defined plainly and the way we use them. Each one links to the instrument or note where it matters.

AI agent: Software that takes actions on someone’s behalf, usually with tools and credentials it was given: charging a card, shipping code, sending an email. The controls Deimos AI LLC builds keep a person in charge of what agents do.See What Deimos AI does
Agent oversight: Seeing what your AI agents are doing, approving what matters and being able to stop them. Clockford™ is built to give you that from your phone, whichever company built the agents.See Clockford™
Human-in-the-loop approval: A step where an agent must get a person’s yes before an action that matters. Our first Flight Rule puts it plainly: an agent can propose anything, a person decides what can’t be undone.See Flight Rules
Effect: An action that leaves your system and can’t be taken back by simply trying again: a charge, a deploy, a published package, a sent email.See Ratchet
Effect gate: A checkpoint in front of an effect. The agent asks before acting, the gate keeps a durable record, and every retry gets the recorded answer instead of a second attempt. Ratchet is an effect gate: one call before acting, one after.See Ratchet
At most once: The promise that the same real-world action is started no more than once, however many times it is retried, restarted or raced. It is the strongest promise that can honestly be kept for an effect that leaves your system, so it is the one Ratchet makes.See Field Note 01
Idempotency key: A unique key attached to an action so that repeats of it can be recognised as the same action. Ratchet keys every effect this way; only the first ask about a key is metered.See Ratchet
Indeterminate effect: An effect whose outcome nobody can know, because the agent crashed after acting and before reporting. Ratchet marks it as a known unknown rather than guessing, and your policy decides: block, retry or check first.See Ratchet
Decision gate: A gate that makes a decision from a rule and facts, never from an agent’s own conclusion. Crimp is an open-source decision gate: the agent brings the rule and the facts, and Crimp works out the outcome.See Crimp
Sealed rule: A rule fixed before anyone knows which way the outcome will go, so it can’t be tuned after the fact. A sealed rule can be re-run against the same facts later and give the same answer.See Crimp
Unknown never seals: Crimp’s rule that a missing fact leaves a decision open. “We don’t know” never quietly becomes a no, or a yes.See Field Note 02
Model Context Protocol (MCP): An open protocol that lets AI agents call tools and data sources through servers. Ratchet speaks it, and Marklode counts the public registry of servers that implement it.See Marklode
MCP server: A program that exposes tools or data to AI agents over the Model Context Protocol. Marklode’s October 2026 recount found 45,156 distinct servers behind 156,111 public registry entries.See Field Note 03
MCP registry census: Marklode’s monthly count of the public MCP registry: every listing, deduplicated into the servers behind it. Metadata only; no server is ever contacted, probed or pinged.See Marklode
Startup due diligence: Checking an idea before you raise money on it: who else does it, what got there first, the strongest attacks on it and what survives. Marklode delivers it in 24 to 72 hours, with every finding checked by a person.See Marklode
Prior art: Patents, papers and products that already disclose an idea. Marklode reports a prior-art landscape; it is not legal advice or a freedom-to-operate opinion.See Marklode
Behavioural baseline: What normal looks like for one person, set from the opening minutes of a recording. M.A.X measures change against that person’s own baseline, never against a stereotype, and it does not detect deception.See Field Note 04
Public-source early warning: Surfacing public conversations that are escalating toward harm so a person can review them, with a link to every source post. It is how Project Vessel is described; it reads public posts only and profiles no one.See Project Vessel
Protected-attributes firewall: A safeguard in Vessel that keeps a person’s characteristics out of the evidence. Who someone is never counts toward a flag.See Project Vessel
NOT ASSESSED: The answer Vessel gives when it can’t read content reliably, instead of a false all-clear.See Project Vessel
Self-check gate: In NeuroPhonics Module 1, a check that refuses to report when the model doesn’t fit or the signal-to-noise ratio is too low. Module 1 is a simulation on synthetic signals, for research only.See NeuroPhonics Module 1
Rainbow Series: A set of US computer-security guidelines from the 1980s, known by the colours of their covers. The Pastel Series is our unpublished research revisiting those criteria for a world with agents inside.See The Pastel Series
OpenSSF Best Practices badge: A badge from the Open Source Security Foundation for open-source projects that meet its published criteria, at passing, silver and gold levels. Ratchet and Crimp hold Silver (projects 14440 and 14613).See Security posture
Signed release: A software release published with a signature and a checksum, so you can confirm it came from its maker and wasn’t altered. Crimp’s v0.1.0 release publishes a SHA-256 of a76f21d53944…38a6.See Security posture
Patent pending: A patent application has been filed and not yet granted. Clockford™ and Crimp are patent pending under US provisional filings; we don’t publish application numbers.See The Record
Ephemeris: A table of where a body in the sky is at given times. The live position of the moon Deimos on every page comes from NASA/JPL Horizons, body 402.See Tonight’s Deimos
Flight Rules: The four rules every Deimos system answers to. FR-1: people stay in charge of what matters. FR-2: assume something is already inside. FR-3: if it matters, make it provable. FR-4: bend, don’t break.See Flight Rules
Claims ledger: The public list of every number and dated fact on this site, each with a status, a source and the date it was last checked.See The Record