AI Mechanics

The methodology, named in full

AI Mechanics is the craft under the agiletoai umbrella — the doctrine, the four redefinitions, and the nine principles that make up what you actually learn to practice. Not a manifesto and not a maturity model: a discipline, tested against the same five changes that have made every other AI framework obsolete inside a quarter. Every claim below carries a receipt — a dated incident it was grounded in — and nothing here names a model, because a principle that names a model expires with that model.

Note I · The doctrine, and what grounded itAI Mechanics, in full

A tachometer, mid-throttle. Read alone it says nothing — paired with load, it says everything about what a system can sustain. The doctrine below is the same instrument for an organization: five gauges, not one throttle, and none of them move because the model changed underneath.

The five change-tests

Model change: Still true if Claude → Mythos → whatever comes next?
Vendor change: Still true if Anthropic → OpenAI → Google → open-source?
Regulator change: Still true if US light-touch → EU AI Act → state patchwork → reverse?
Role change: Still true if “prompt engineer” → “AI engineer” → whatever’s next?
Pattern change: Still true if RAG → agents → orchestration → whatever’s next?

A claim that fails any test is a practice — it belongs in the deck. One that passes all five is a value — it belongs here.

Identification notes

Five values, built to survive the next model

The dominant practice for working with AI changed three times in twenty-four months — prompts, then context, then orchestration. A government revoked access to two production frontier models without warning, then gated the next one before it shipped. Any framework that assumes a stable model, vendor, role, or regulation is at risk of obsolescence within a quarter. These five values survive that — not because they’re clever, but because each one is grounded in a dated incident, and each one still holds when the model, the vendor, the regulator, the role, or the pattern changes underneath it.

01
Human accountability over delegated trust
even when AI is right more often than humans
The agent cannot own outcomes — accountability is a human property, full stop. Receipt:Anthropic’s Mythos 5 withdrawal (June 12, 2026), the model gone in hours while only the humans accountable for the work kept going; the /dashboard parseSort incident the same month, where AI-generated code passed the compiler and failed at runtime anyway. The human was on the hook either way.
02
Verification rituals over output volume
even when verification slows you down
AI generates faster than anyone can review — the bottleneck moved from can we produce enough to can we verify enough. Receipt: the Test-Before-Fix Doctrine, written after that same parseSortincident, flags any bug-fix commit that lands without a coupled test — because “tests pass” was shallow when the tests never covered the change.
03
Boundary craft over either-side mastery
even when one-side specialists are still being hired
Pure-AI work is brittle; pure-human work is slow. The defensible value sits at the interface — where the agent is allowed to act, where a human must decide. Receipt:the Agent Spawn Policy, tightened after a $100 overnight subagent overrun (June 2026) — “the value is an intentional, managed, step-by-step workflow,” not unattended fan-out.
04
Documented intent over implicit context
even when documentation feels slow
In the AI era the documented intent isthe contract the agent acts on — a vague spec produces confidently wrong work, at speed. Receipt:the decisions captured the day they’re made outlive the session that made them; undocumented work dies with the agent’s context window.
05
Trust-paced adoption over capability-paced adoption
even when capability tempts you to move faster
AI capability advances monthly. Trust compounds yearly. The right pace is set by the slowest-moving human in the loop, not the latest model card. Receipt: the June 2026 sequence — Mythos 5 and Fable 5 access pulled June 12, partially restored to roughly a hundred approved parties June 26, GPT‑5.6 gated before its own launch the same day. Teams that had paced adoption to trust kept shipping. Teams that had bet on one model lost weeks.
Five values. Each one survives the model changing, the vendor changing, the regulator changing, the role changing, the pattern changing underneath it — which is the only entry criterion that matters.

Note II · The four redefinitions, in fullAI Mechanics, in full

An inspection stamp — The same stamp as the home page’s naming plate, held longer here. A mark is worthless without the name behind it — which is the whole of the fourth naming, and the reason this page exists.

The three working verbs

Augment: What are we augmenting — and who decided?
Accelerate: What pace can we defend?
Interpret: Who is accountable for the answer?

Identification notes

The same two letters, read four ways

The first reading points at a machine. The last points at a person. Nothing between them is a different technology — only a different reading of the same two letters, and the reading you use decides where your team looks for the answer.

01
Artificial Intelligence
the framing you were handed
The framing everyone was handed. Notice what it does: it points the leader at the technology — licenses, tools, a procurement problem. It is not wrong so much as aimed at the wrong desk.
02
Augmented Intelligence
the question of what you amplify
Your people and systems are already intelligent. The question this reading actually asks is what you amplify — not what you replace. It moves the decision back onto a desk a leader can sit at.
03
Accelerated Inference
probability, at a speed you have never had
The language model itself doesn’t calculate — it infers. “There is a high probability this is true,” at a speed you have never had. Speed is real. Certainty is not what it sounds like.
04
Accountable Interpretation
the naming that has a name attached
An accelerated inference has no meaning until a person interprets it — and that person has a name. Risk lives there. Ethics lives there. Value lives there. Accountability never transfers to the tool.
Same two letters, four times. The first pointed at a machine. This one points at a person — and it is the reading everything else on this page argues from.

Note III · Nine principles, scored on this deviceAI Mechanics, in full

The core asset — live

Three verbs. People, process, tools. Not one of them names a model.

That is not an accident — it is the whole design. A principle that names a model expires with that model. These nine are built to outlast the next release, and every one of them is scored the same way: honestly, at the row that is actually true, not the row that is impressive to say out loud. Score all three and the pace-setting row highlights itself. Your scores stay on this device, and the same scorecard on the FedAgile companion sheet (/resources) reads them back.

	People	Process	Tools
AUGMENTWhat are we augmenting — and who decided?	Name the human whose judgment is amplified, not the headcount being saved.	What context goes in is a management decision, not an individual’s.	Curated context beats more context. Plan for what it silently forgets.
1 · more noise, faster — / 5higher-order problems · 5
ACCELERATEWhat pace can we defend?	Your verification capacity — not your build capacity — sets your real pace.	Diagnose every miss by cause: augmentation, acceleration, or interpretation.	Measure what ships and survives, not what gets produced.
1 · the old loop, faster — / 5shrinking loops · 5
INTERPRETWho is accountable for the answer?	A named expert signs off. “The AI said so” is not a sign-off.	Source → Authority → Own-Eyes, before anything moves.	Pair the language model with things that actually compute.
1 · blind trust — / 5practised judgment · 5

The hinge: your lowest row isn’t a weakness — it’s your pace. Score all three rows to read yours.

Not saved yet · scores this device only, nothing sent anywhere

Context, and how it is quietly lost — the bucket, the three-stage failure sequence, and access is not attention — is on the home page.

Note V · Faster is not betterAI Mechanics, in full

Two pressure gauges on one line — Two gauges, plumbed off the same line. 1 reads high; 2reads moderate. Same pressure, read differently — because pace was never one number. It is whichever gauge you decide to trust when the room is watching.

Keeper lines

Three lines, same grammar, printed on the reverse

Delivered roughly twenty minutes apart in a live room, the three keeper lines are the method compressed to its studs. Read together, they are the whole argument in nine words.

Keeper · 01Access is not attention.

Keeper · 02A system is not a model.

Keeper · 03Faster is not better.

The machine sets the speed it CAN go. You set the speed it SHOULD go.

A document is the output of thinking. It is not the thinking.

Note VI · The vocabularyAI Mechanics, in full

Reading key

Eight terms, used the same way in every lane

The talk, the deck, the shelf, and this page draw from one glossary. When every surface says “Doctrine” and means the same five values, the brand feels coherent. When one says “AI safety check” and another says “verification process,” it doesn’t — so nothing here gets a synonym.

Canonical vocabulary

AI Mechanics: The methodology brand and the methodology IP — the doctrine, the redefinitions, and the nine principles on this page. Not “the AI Mechanics framework.”
Doctrine: The five enduring values named in Note I. Not “principles” or “values” used loosely — those words are reserved for this canonical five.
Discipline: The five supporting practices that operationalize the doctrine day to day. Not “best practices” or “habits.”
Concept: One of the twenty-eight named mechanics of AI. Not a “topic” or an “issue.”
The Shelf: The artifact lane — prompts, GPTs, Gems, Projects, IDE configs. Not a “toolkit” or a “library.”
The 3-Check: The verification ritual: Source → Authority → Own-Eyes, three minutes per document, before anything moves. Not “triple-check.”
The 80% Problem: AI gets to almost-right fast; the remaining twenty percent is human judgment — not a limitation to route around.
The Ownership Principle: The agent is never accountable. Accountability is human, full stop. Not “AI ethics” or “AI responsibility.”

Read the system →Work with usThe method is free, in full. This is the whole of it, minus the room to run it in.