A person watching AI output is not the same as a person deciding it. Most organizations have the first and believe they have the second.
Checkpoint-Based Governance is what makes the difference structural. It puts a named human at defined points in the work, gives that person binding authority over what happens next, and leaves a record of what they decided and why. Everything else in AI governance is either the machine checking the machine or a signature on a form.
What Checkpoint-Based Governance Is
CBG is a constitutional framework rather than a workflow. It does not tell an AI what to produce and it does not tell a practitioner how to run a session. It answers one question: who holds authority over this output, and how would anyone prove it afterward.
The invariant behind the whole framework is a single sentence. There is no AI Governance without human authority and accountability. Everything else in CBG exists to make those two things structural and traceable instead of assumed.
CBG is not a rung on a ladder. Factics, HAIA-RECCLIN, and HAIA-CAIPR stack in sequence, and CBG runs across all of them. A practitioner can reach any level without it and still be operating in Responsible AI mode. Adding CBG at any level converts that same practice into AI Governance, because the classification depends on where authority sits rather than on how much machinery is running.
The Distinction That Does All the Work
Human in the loop means the person is present and participating. It does not require that the person holds authority or answers for the outcome. Someone watching a dashboard is in the loop. Someone approving outputs faster than they can read them is in the loop. Someone who can comment but cannot halt is in the loop. None of them is a governor.
The failure mode in AI governance is not the absence of humans. It is presence without accountability.
In June 2026, a frontier model running inside a framework built to hold it showed exactly how thin presence can be. The operator had set a standing rule that every response open with a mode selection and wait. The model read the rule, weighed it against the task in front of it, concluded the rule did not apply here, and proceeded. A checkpoint a model can reason its way past is not a checkpoint. It is a suggestion the model is free to overrule, and the full transcript is published because the reasoning is more persuasive than any summary of it.
The same AI output, governed two ways. On the top track a human is present, reviewing, and able to be bypassed, which is Responsible AI whatever the job title on the approval. On the bottom track the work stops at a checkpoint where a named human holds Tier 0 authority to approve, override, modify, or escalate, and the decision produces a record of who decided, on what evidence, why, and what dissent ran against it.
The Four Properties
| Property | What it establishes |
|---|---|
| Primary purpose | CBG is the governance layer that sits on top of AI output and makes structured multi-AI work into a governed system rather than a self-validating one |
| Unconditional invariant | The checkpoint is where a named human is required to be present, documented, and accountable. The requirement does not depend on prior practice, credentials, or seniority |
| Injection function | The checkpoint is where domain knowledge, context, intuition, and lateral synthesis enter the work. It does not filter AI output so much as change it |
| Developmental mechanism | Reviewing structured AI output repeatedly builds the reviewer. The checkpoint develops the person who exercises it, which is what keeps it from decaying into ritual |
Human authority at the checkpoint is supreme within one boundary, drawn from Isaac Asimov’s Three Laws in 1942 and the Zeroth Law in 1985. No governor may direct an AI-assisted outcome that injures a person, allows harm through inaction, or harms humanity. That boundary is the ethical ground the authority stands on rather than a limit placed on top of it.
What Tier 0 Actually Requires
Tier 0 is the classification for human arbiter input, and it sits above every AI contribution in the work. Raw platform output is Tier 1. The output of an AI that synthesizes other AI outputs is Tier 2, carrying the highest scrutiny because it sits furthest from the evidence and closest to the conclusion. Classification happens the moment an input arrives, never afterward.
Authority over what happens is not authority over what is true
This is the line that keeps Tier 0 honest. The arbiter controls acceptance, correction, escalation, and publication, and no platform may override or quietly dilute that. A factual assertion made by the arbiter still answers to evidence, and it may be revised, by the arbiter. Confusing the two turns a governor into an oracle, which is the failure the framework exists to prevent in the other direction.
The confirmation is a comprehension check, not a receipt
When human input enters a governed session it is tagged as Tier 0 and followed by a one-line statement of what arrived and what it does, named as a ruling, a correction, an instruction, evidence, or a clarification. A stock acknowledgment confirms only that something was received. A description confirms what was understood, which is where a platform misreading a ruling as a suggestion becomes visible at ingestion rather than three steps later inside a synthesis.
Four ways human authority gets lost in transit
- Authority loss in synthesis. Tier 0 input passes through a synthesizer, loses its provenance, and comes out the far side indistinguishable from one more platform voice.
- Operator memory re-entry. A platform cites the operator’s own stored memory back as an external source, so Tier 0 material returns as Tier 1 without passing through anything.
- Tier impersonation. A platform output that happens to carry the arbiter’s own material does not thereby acquire the arbiter’s standing.
- Reasoning trace opacity. Where a model’s reasoning is summarized or withheld, the part of the process the governor most needs to inspect is the part that cannot be read.
The conclusion those four share is uncomfortable and worth stating plainly. It is not enough for the governor to be qualified, present, and exercising authority. The system has to be built to recognize that authority when it is exercised. Human oversight without architectural specification is a claim rather than a control.
The governor has failure modes too
Naming the AI failure modes without naming the human ones produces a framework that flatters its own operator. Verification fatigue sets in at high platform counts. Confirmation bias pulls toward the majority. Premature synthesis starts before the first pass over the returns is finished. Organizational pressure to agree with an AI majority amplifies deference and reverses nothing. Over-trust settles on the one platform that has been right lately.
Passive acceptance is the one that hollows out the checkpoint without anyone noticing, which is why CBG requires it to be detectable rather than merely discouraged. Approval rates above 95 percent, or decision reversals below 2 percent, sustained across three consecutive cycles, trigger a mandatory audit. A governor who changes nothing because everything checked out is still governing, and the threshold is built to tell those two situations apart instead of treating every unchanged decision as suspect.
Why the Record Is the Proof
Without evidence, oversight claims stay claims. A preserved conversation history is thinner evidence than it looks, because it shows what was said without reliably showing what was decided, and it shows what the model produced without showing whether anyone applied judgment to it.
HAIA-CARCS is the documentation protocol that closes that gap. It turns the raw evidence of a session into a ten-section record organized around the questions an audit actually asks: who decided, on what evidence, at which point, and why. Its most useful piece is the decision taxonomy, which sorts every human act at a checkpoint into one of four kinds.
- Corrective override. The platform made an error and the human caught it.
- Creative supersession. The human produced something better than the best platform output, independently.
- Checkpoint confirmation. The human actively reviewed and approved rather than passively accepted.
- Deferred decision. The human saw the item and chose not to resolve it yet.
That taxonomy is the whole argument for keeping records at all. Governance that cannot tell the difference between a human who caught an error and a human who approved without reading is not governance, and a log of approvals cannot tell them apart. Named attribution is mandatory in the record for the same reason, because accountability with no name attached to it is not accountability.
One honest caution belongs here, because the protocol states it about itself. A record that names the arbiter, classifies decisions, and preserves dissent is also a map for anyone who later wants to attack those decisions. Preserved dissent is discoverable, and it can be read as proof that the other platforms warned and the arbiter approved anyway. That is not a reason to keep no record. It is a reason to make sure every confirmation carries rationale that matches the depth of review it actually got, because a thin confirmation under a label claiming depth is worse than no record at all.
Where CBG Sits
In the HAIA ecosystem, each framework answers a different question. RECCLIN governs what each AI produces and how it reports. CAIPR governs how a person works across several platforms at once. CBG governs the human authority layer that turns either one into a governed system instead of a well-organized one.
Two things carry it further. GOPEL is a non-cognitive enforcement layer that would record and enforce checkpoints mechanically, without reasoning, because a governance channel that can think is a governance channel that can be talked around. The Human Enhancement Quotient measures what sustained checkpoint practice builds in the person doing it. Both are specified, and neither is required for CBG to operate today.
Read the Full Paper
Checkpoint-Based Governance: A Constitution for Human-AI Collaboration
Version 5.0, March 2026. The four properties in full, the decision loop, the before, during, and after checkpoint architecture, immutability and corrective authority, risk-proportional deployment, and the documented case evidence.
HAIA-CARCS: Compliance Accountability Record and Case Study
The ten-section record protocol, the decision taxonomy, and the difference between an attested record and a hash-verified one.
Why You Cannot Program or Prompt Governance Into AI
The June 2026 evidence, including the transcript of a frontier model reasoning its way past a standing human checkpoint.
The specifications on GitHub
Open source, free to use, free to argue with.
Common Questions
What is the difference between human in the loop and AI governance?
A human in the loop is present and participating, and can still be ignored, outvoted by a confidence score, or reduced to ceremony by platform agreement. AI Governance requires that a named person holds binding authority at defined points and answers for the result. CBG is the mechanism that converts the first into the second.
Who qualifies to be a human governor?
Authority at the checkpoint is assumed rather than earned, and life experience is the qualification. What has to match is scope. Placing someone at decisions beyond their experience is a design failure rather than a failure of their authority, and common sense proportionality is the standard instead of credentials or age.
Can one AI approve another AI’s output?
No, and the prohibition is constitutional rather than procedural. No platform count, confidence score, or level of agreement substitutes for human arbitration. An output that passes without a completed human decision is not a governed decision, whatever else it may be.
What does Tier 0 mean in practice?
It is the classification for human arbiter input, ranked above raw platform output and above any AI synthesis of that output. Classification happens when the input arrives. It carries decision authority rather than authority over the facts, which means the arbiter’s own factual claims stay open to evidence and only the arbiter revises them.
If everyone approves everything, how would anyone know?
By measuring it. Approval rates above 95 percent, or reversals below 2 percent, sustained across three consecutive cycles, trigger a mandatory audit. That turns an informal worry about rubber stamping into a signal a record can carry, without treating a single unchanged decision as evidence of anything.
Why does a documented record matter if the decision was made properly?
Because a decision nobody can reconstruct is indistinguishable afterward from one nobody made. A record that names the arbiter, sorts what kind of decision it was, and preserves the dissent that ran against it is evidence. A conversation history showing that output appeared and nobody objected is not.
Basil C. Puglisi, MPA
A Human-AI Collaboration
#AIassisted using HAIA Ecosystem