Glossary

A working vocabulary, defined for the room you are in.

The terms behind the practice, defined three ways. Use the register that fits who is asking.

See the approach ->Discuss your context ->

House rule

The expert definition is not always the useful one. The buyer's definition is not always the precise one. Both belong in the room.

Plain-language summary

Three definitions per term. The first is for peers in AI governance, safeguards, and industrial-safety practice. The second is for a smart non-specialist who has never heard the term before. The third is for someone deciding whether to engage the practice for a specific decision already in front of them. Pick the register that matches the situation; the term means the same thing in all three.

Term

Industrial safety for algorithmic systems

For specialists

A governance posture that treats high-stakes AI-enabled workflows the way industrial safety treats consequential physical systems: with named hazards, evidence trails, stop-work authority, escalation routes, independent inspection, and a refusal to let smooth output stand in for verified control. The posture is not restriction-by-default; it also challenges safety measures that remove support, context, or agency without evidence that the withdrawal helps. Distinct from generic AI ethics by its insistence on operational primitives rather than principles.

In plain language

We treat AI in serious workflows the way the safety inspector treats a factory floor. There are hazards. There are stop buttons. There is paperwork. The point is not to prevent the machine from working. It is to make sure a person can pause it when something is wrong, and also to make sure nobody calls a shutdown 'safety' without checking who it helps and who it hurts.

For buyers

If your team is being asked to trust an AI-enabled decision (a vendor's risk score, an LLM-generated audit summary, a workforce optimisation tool), and there is no clear answer to 'who can pause this and where does the evidence live,' this is the practice that answers those questions before the next meeting.

Term

The Watchdog (Grievance Watchdog Architecture)

For specialists

A governance architecture for high-stakes decision systems that need to hear harm signals from affected people, preserve raw evidence beyond categorisation, route escalation independent of operational pressure, and retain the right to refuse continuation. Distinguishes a sentinel (which can halt) from a sensor (which only transmits). Substrate for product instances such as GrieVoice.

In plain language

A way of designing complaint and reporting systems so that the original story does not disappear into a category before anyone has acted on it. The original voice has to stay locatable. The system has to be able to stop, not just report.

For buyers

If your grievance system is producing low complaint volume and you are not sure whether that means 'no problems' or 'nobody trusts the channel enough to use it,' Watchdog architecture is what we use to find out which one it is and to redesign for the second case.

Term

H∞P (Humans in the Hoop)

For specialists

Training framework for teams carrying live accountability for AI-enabled decisions. Replaces the decorative 'human in the loop' construction with operational requirements: stop-work authority, evidence trails, role clarity under uncertainty, vendor-claim interrogation, and the capacity to challenge a confident automated output without taking personal career risk. Includes the H∞P Challenge Lab for live scenario pressure-testing.

In plain language

Training for the people who are supposed to be 'in the loop' when AI tools enter their work, designed so that being in the loop actually means something. Not just rubber-stamping the model's output. Being able to slow it down, ask the harder question, and have the paperwork to back the answer.

For buyers

If your audit, ESG, risk, M&E, or compliance team is being asked to sign off on AI-assisted work and you want them to be able to defend that sign-off under scrutiny, H∞P Training is the format that builds the practice rather than the slide deck.

Term

Sonic Cycles

For specialists

Parallel creative channel running alongside the newsletter, generating sonic companions, multi-model arena experiments, and alternate ways into the research. Functions as both distinctiveness layer (a non-fungible signal in a crowded thought-leadership market) and as primary research surface for some of the model-behaviour observations later folded into the analytical writing.

In plain language

The music side of the practice. Some of the research turns into songs. Sometimes the songs are the research. It is a different way of getting at material that does not always survive being argued in paragraphs.

For buyers

Not a deliverable. A signal that the practice is doing original work rather than recycling other people's frameworks. If you found us through the writing, the Sonic Cycles are where the same questions arrive in another register.

Term

Calvin-compliant procurement

For specialists

Procurement posture grounded in the six mechanisms of the Calvin Convention: pre-action constraints, right of override, edge-case registry, right of refusal, provenance preservation, and accountable continuation. Translates AI-governance principles into vendor-contract clause language, evidence-request lists, and operating-control requirements that survive the gap between policy and implementation.

In plain language

A way of writing AI-vendor contracts so that the things your policy says are important (being able to stop the system, knowing where its evidence comes from, having the right to refuse an unsafe recommendation) actually appear in the contract, not just in the values statement.

For buyers

If you are about to sign a vendor agreement for an AI-enabled system and you want the contract itself to encode the controls your team will need when something goes wrong, Calvin-compliant procurement is the clause work and the vendor-interrogation script we use.

Term

GrieVoice

For specialists

Multilingual grievance intake infrastructure for worker voice, community reporting, whistleblower channels, and early harm signals in high-risk operating environments. Voice/WhatsApp/USSD intake; structured record with original-language preservation; identity decoupling for anonymous reports; operational-blackout contact rules; labour-broker firewall. Sold as a scoped pilot rather than a low-ticket subscription.

In plain language

A grievance reporting system that people can actually use. They can talk, in the language they speak at home, on the phone they already have. What they said is kept the way they said it. The supervisor cannot see when they are being followed up with. They keep a reference number that travels instead of their name.

For buyers

If you run a large distributed workforce (mining, agriculture, construction, manufacturing, logistics) and your current grievance system is showing low complaint volume that you do not entirely believe, GrieVoice is the scoped pilot that tests whether the channel itself is capable of hearing the workers most at risk.

Term

The Primer Hypothesis

For specialists

A hypothesis that AI companions may provide stabilizing relational support for isolated users whose human support systems are absent, inconsistent, or unsafe. The hypothesis does not clear companion platforms of product risk. It requires any withdrawal or dampening intervention to measure subgroup outcomes, displacement, and crisis-channel effects before treating visible refusal as settled safety.

In plain language

Some people may be using AI companions as a bridge when no person is available. If a platform suddenly removes that support, the change itself may hurt some users. The claim needs measurement, especially for isolated young people.

For buyers

Use this when a safety proposal removes relational features from a companion, support, or care-adjacent system. It helps frame the evidence question: who is protected, who is abandoned, and how will the organisation know the difference?

Terms here will grow as the practice does. If a term you ran into on the site or in a session is missing, use the structured enquiry form and name it, and it will land here in the next pass.