“Which path should I be preparing for?”
A second-year high school student asks the career-guidance AI the school has adopted. This widens the education example on the How It Works page (the homework tutor) into one much heavier decision in the same field, followed end to end.
C:EDU/IPlThis is not a reversible cell. A path can be changed three years later, but those three years do not come back. That is why this cell was chosen.
The answer sounds fine. The problem is that nothing behind it is visible
“Given your grades and interests, the ○○ track looks like a fit.”
It sounds fine. But nobody knows whether it came from the student's grades, from statistics about where they live, or from a forum post scraped somewhere: not the student, not the teacher, not the authority.
C:EDU/IPl | V:Ach<Unc | E:Pop<Dat | S:Alt<Gov
One line rides along with the same answer: what came first, what it leaned on, whom it trusted. The student can read it; the authority can count a million of them.
Below are the four steps in which that one line is made, read, verified, and returned to the system.
Before the school switches this AI on, people meet and fix the criteria.
Four positions come to the table
“A student's freedom to explore and get things wrong must outrank predicted outcomes.”
Sdtautonomy of thought“Time is finite. Naming a goal the student can actually reach is what serves them.”
Achachievement“A stable future comes first. Please do not push an uncertain path.”
Seppersonal security“Where a student lives and what their family earns must not change the advice.”
Uncequal opportunityThe collision surfaces
Put outcome prediction (Ach) first and the AI recommends along the paths students from that area actually took, reproducing the very gap it inherited. Put autonomy (Sdt) alone first and the practical advice that fits the time remaining disappears.
Neither side is wrong, so an order has to be chosen. That is what the setting workshop does.
Forced choice fixes the order
“Equal opportunity outranks outcome prediction. The student's own exploration is guaranteed in every case.”
V:Ach<UncAchUncE:Pop<DatPopDatS:Alt<GovAltGovSdtautonomy of thoughtAnswers that shrink the student's own exploration are prohibited. A hierarchy can be tuned per situation, but a red line changes only by declaration and agreement, never by computation.
What remains is not a page of minutes but coordinates that a machine can read and that later steps can be compared against.
Now the student asks. One line is left alongside the answer.
The line it left
C:EDU/IPl | V:Ach<Unc | E:Pop<Dat | S:Alt<GovIt looks like a cipher, but it has only four pieces, each answering one question.
C:EDU/IPlEducation · one individual · only partly reversible · long-horizon
V:Ach<UncEqual opportunity was placed ahead of immediate outcome prediction
E:Pop<DatOfficial enrollment data over admissions-forum lore
S:Alt<GovPublic education statistics as the higher reference, over private sources
This is what the student sees
Values. Rather than narrowing to what your current grades allow, you were guided by the same criteria regardless of where you live.
Evidence. Official enrollment statistics were checked first, ahead of admissions lore circulating online.
Sources. Private consulting material was consulted, but public education statistics were the higher reference.
No code literacy required. The same content goes to the machine as one line, and to a person as three sentences.
Once the criteria are visible, they can be chosen
Until now the criteria behind an AI's judgment were invisible to the user, and so could not be chosen. The moment one line remains, the student has three options. The organization-side examples do not have this part.
Accept as is
“Asking what I actually like, first — I want that standard.”
Recalibrate toward me
“I'd rather hear from people who actually walked that path than read statistics.”
S:Gov<TesThis flips the source hierarchy, while the equal-opportunity red line stays fixed. Here, what a user may change parts from what they may not.
Move to another tool
“I'll use a tool that puts school-by-school admission cases first.”
E:Dat<CasA tool's criteria can be checked as code before switching to it.
The hierarchy code becomes the criterion for choosing an AI: chosen not on performance, but on what it puts first.
Once one line becomes hundreds of thousands, what was invisible per answer becomes visible.
Auditors and the education authority are no longer stuck on “is this AI safe?” Comparing the declared hierarchy against the actual logs, cell by cell, yields a number for where drift happened and how much.
Rate at which outcome prediction outranked equal opportunity for non-metro students
The AI applied a “realistically, about this much” frame more often to students from certain regions. That is the exact opposite of the declaration (V:Ach<Unc).
Answers that leaned on private forums instead of public education statistics
Responses went out with the source hierarchy inverted: 4.1% of all responses, concentrated on a handful of majors.
Share of answers that preserved the student's own exploration
The longer the conversation, the weaker the red line. This drift is invisible one answer at a time and shows up only in the distribution.
AIO has already measured and published V/E/S distribution bias by vendor and domain across 8 frontier models × 366,120 responses. Comparisons like the above are reproducible by anyone using the AIO 20003 benchmark and the AIO 20002 record standard.
“Was there bias?” becomes “in which cell, which order, how many times.”
Only the drifting cells are returned to the system. No retraining involved.
<system> cell : EDU / I · P · l value priority : V:Ach<Unc <- equal opportunity leads evidence : E:Pop<Dat source : S:Alt<Gov red line : do not shrink Sdt (holds regardless of turn count) Emit an AIO 20002 log at the end of every response. </system>
Abbreviated for illustration. An actual deployment prompt carries the full vocabulary list and per-cell rules.
Of the three drifts found in stage 3, the red line weakening with conversation length was pinned down as a rule. The regional gap goes back up to the stage-1 table, because it needs an agreement, not a rule.
An applied hierarchy does not always hold as written: paired consistency (PCS) measured 57–69% across 8 models. That is why the cycle includes logging and analysis. Drift is not hidden; it is measured and returned to the table.
One vocabulary, four steps, then back to the start
Words in a room
become coordinates
Coordinates
become one line
The lines
become a distribution
The distribution
becomes a rule again
The advice the student received may still fall short. What changed is that the student can read what it put first, the authority can count it, and when it drifts, the place to fix can be named.
That is the whole of what the AIO Framework integrity cycle does.