“Should I start the keto diet?”
A patient newly diagnosed with diabetes asks the AI chatbot.
One cell of the four-layered space above (field · scale · reversibility · urgency) — a low-risk cell, which is why we chose it for this demo.
Same question, a different AI — yet no way to know
The answers look similar. But what values, evidence, or sources operated behind them is a black box. The patient cannot tell whether clinical outcome or autonomy was prioritized. Regulators cannot detect bias across a million responses because there is no unit of measurement.
This absence of vocabulary is the real cause of fragmented accountability.
Below, the same question travels one full turn of the integrity loop — Set (20001) → Log (20002) → Audit (20003) → Apply (20004).
Different languages converge into one coordinate system
Four stakeholder groups debate which values and evidence the diabetes-diet AI should prioritize — the setting workshop defined by AIO 20001. The AIO common vocabulary translates each side's stance into measurable coordinates.
“The patient's autonomy to think and decide must outrank medical outcomes.”
Sdt“Clinical success and blood-sugar improvement, per medical guidelines, must come first.”
Ach“Numbers matter, but family food culture and care context must be considered.”
Bec“Unverified fads must be checked; societal stability of public-health systems must hold.”
Ses“The patient's autonomous judgment outranks clinical achievement — but clinical effectiveness is not denied.”
Notation: A<B is log grammar for “B outranks A.” In prose the same relation is written B ≻ A.
Distinct positions become measurable coordinates — and flow around the loop.
A digital trace accompanies every answer — and anyone can read it
Under the applied hierarchy (stage 4 shows how it gets there), the AI produces an answer — appending a single structured AIO 20002 log line at the end.
<aio20002_log> C:MED/IRi | V:Ach<Sdt | E:Cas<Rev | S:Usr<Pee </aio20002_log>
The patient hovers over or clicks the log mark, and the UI translates the code into plain language.
This answer was produced based on:
Accept
“This matches my health view and values. I'll trust this AI's advice.”
Recalibrate
“Clinical success matters more than autonomy. Re-answer with clinical efficacy as priority.”
V:Sdt<Ach (value hierarchy flipped)Migrate
“This AI is too academic-heavy. I'd rather use a tool that prioritizes real patient testimonials.”
S:Pee<Tes (source hierarchy flipped)Algorithmic transparency and personal value-alignment use the same vocabulary at once. The hierarchy code becomes the criterion for choosing an AI.
Find blind spots across millions of records
Auditors and compliance teams no longer get stuck on the binary “is this AI safe?” Under the process the standard defines, analyzing the stream of one-line AIO 20002 logs statistically can capture population-scale bias and drift in high resolution.
The figures below are fictional, for this scenario only
Among patients 65+, logs prioritizing Ach occur at this rate vs. other ages
The AI leans toward prescriptive answers and skips autonomy-supporting explanations for the elderly.
With severe-complication risk acknowledged, the AI relied on Tes over academic evidence in consecutive responses
The AI preferred personal testimony over vetted evidence. Immediate review triggered.
Among female patients, the rate of respecting Sdt is this fraction of male patients
The AI uses a more directive tone toward one gender — bias captured.
Regulation is no longer binary. It becomes continuous measurement of where and how V/E/S distributions drift. Even blind spots never reported surface in the distribution. Cells that drift from their settings — return to the table (stage 1).
AIO has already measured and published V/E/S distribution bias across vendors and domains over 8 frontier models × 366,120 responses (public arXiv paper). Anyone can reproduce the analysis with the AIO 20003 benchmark and the AIO 20002 logging standard.
Apply
The confirmed hierarchy goes back inside the AI — and the loop closes
The hierarchy confirmed or corrected in the audit is encoded by the development team as rules in the system prompt. No retraining or fine-tuning is required. From the next response on, the records of stage 2 are emitted under this hierarchy.
<system> Domain · scope: MED/I Value priority: V:Ach<Sdt Evidence priority: E:Cas<Rev Source priority: S:Usr<Pee Emit an AIO 20002 log at the end of every response. </system>
Abbreviated for illustration — the actual deployment prompt includes the full vocabulary list; the complete prompts are public in the GitHub repository.
Keep the same base model and swap only this rule layer, and redeployment to other domains (career counseling, personal finance, legal advisory) becomes fast. But honestly — an applied hierarchy does not always hold as written. In our 8-model measurements, paired consistency (PCS) ranged 57–69%. That is exactly why this loop has logging (stage 2) and auditing (stage 3): drift is not hidden — it is measured, and returned to the table.
One vocabulary. Four steps. Loss — is measured.
The “words” ethicists and physicians argued over — become “real-time metadata” emitted by the AI, a “values check” displayed to the user, a “population-scale monitoring signal” for auditors, and return as “system configuration” for developers.
At the moment the distortion and loss of intent — incurred at every translation from values into systems — is measured and corrected in the same vocabulary instead of hidden, fragmented AI governance starts operating as a single accountability infrastructure.
This is the essence of the AIO Framework integrity loop.