top of page
Abstract Shapes

INSIDE

PUBLICATIONS

The CI-First Evaluation Framework: how U365 scores an AI tool, and what the versions mean

4 hours ago
27 min read
The CI-First Evaluation Framework: how U365 scores an AI tool, and what the versions mean.

In this Report



Current version: framework 1.4.1 Applied 2026-10-14. Last wording alignment 2026-11-11, when the five AI Profile names were aligned with the institutional canon; that alignment changed no methodology. Latest change: v1.4.1 restored the attribution of the term Co-Intelligence to Ethan Mollick's book Co-Intelligence (2024), which v1.4 had removed by over-applying a scoped instruction. No score moved in v1.4 and none moved in v1.4.1.

Back to the TOC

Section icon: What this framework is for.

The CI-First Evaluation Framework: how U365 scores an AI tool, and what the versions mean

What this framework is for


Every AI tool review answers two questions that a feature list cannot. First, does this tool make co-intelligence more profitable than human intelligence alone? Second, does it make you more capable, or only faster?


The framework exists so that both questions get a written answer that a reader can check, and so that two tools can be compared on the same terms. Without it, a review becomes a description: what the product does, what it costs, what other people said. With it, a review becomes a measurement.


The measurement rests on the U365 definition of co-intelligence:


CI = HI + (AI x HI)


Human intelligence is not added to artificial intelligence as an equal term. It is the multiplier. If HI stays strong, better AI raises the product. If HI drops because you stopped doing the thinking, the product falls even when the AI improves. That is why the framework scores the human side of the relationship as carefully as the tool side.


Three outcomes follow from the formula, and the framework names all three. A Superhuman is a person whose cognitive, creative, ethical and strategic capacities are measurably amplified through sustained co-intelligence, while remaining decisively human and keeping control where machines fail. A Sub-human is the person who delegates the thinking itself, so that HI falls and the co-intelligence result falls with it, even with a strong model in front of them. AI Imposture Risk is the chance that a tool creates the appearance of capability while the human contribution approaches zero. If HI is zero, no amount of AI rescues the product.


So the framework scores four things about a tool: the benefit it delivers, the human capabilities it protects or erodes, the illusions it can create, and the role it should be given. This publication explains each instrument, then explains what a framework version number means and what changed in each release.


Back to the TOC

Section icon: Where the framework sits inside CI-First.

The CI-First Evaluation Framework: how U365 scores an AI tool, and what the versions mean

Where the framework sits inside CI-First


CI-First is the doctrine. The framework is the measuring instrument.


The doctrine explains why a human stays in the orchestrator's seat, how a CI-First workflow is structured, and how an organisation builds the practice. Three INSIDE publications carry that argument:



None of the three explains how a tool is scored. That is the gap this publication fills. It is the reference document behind every Tools Review on this site, and the link target for the first mention of the framework in any U365 publication.


The framework introduces no new vocabulary. Every term in it is an existing U365 method definition: CI-First, Superhuman, Sub-human, AI Imposture, Humics, AI Obesity, Omni-Skilling, the five AI Profiles, UNOP, ULM, LIPS, SL-OS and UP-Context. What the framework adds is arithmetic and evidence rules, not concepts.


Back to the TOC

Section icon: The CI-First Benefit Score, and the four dimensions it measures.

The CI-First Evaluation Framework: how U365 scores an AI tool, and what the versions mean

The CI-First Benefit Score, and the four dimensions it measures


The CI-First Benefit Score measures how much a tool delivers the four benefits U365 treats as measurable. Each dimension is scored from 0 to 10. The overall score is the arithmetic mean of the four, rounded to one decimal place.


The score answers one question: does this tool make CI greater than HI for the person using it?


The 0 to 10 scale, and what each band means


Score

Label

Meaning

0

None

No benefit. No measurable advantage in this dimension.

1 to 2

Negligible

Barely noticeable. Not worth the overhead of using the tool.

3 to 4

Marginal

Modest benefit in narrow use cases. Many users will not notice.

5 to 6

Moderate

Clear, consistent benefit in common use cases. Worth using for this dimension alone.

7 to 8

Strong

Significant benefit across most use cases. A primary reason to adopt the tool.

9 to 10

Transformative

Fundamentally changes what is possible in this dimension. Rare. Reserved for tools that redefine a workflow.


Dimension 1: Time Benefit


Question: how much time does this tool save compared to doing the same work with human intelligence alone?


What is measured: net time saved, meaning time saved minus the time spent prompting, configuring, reading, verifying and correcting output; how often the saving occurs; and how wide the range of tasks is.


The framework treats this dimension as the most commonly overstated. Using AI is itself time-consuming. A tool that saves thirty minutes of writing but costs twenty-five minutes of prompting and checking has a small net gain, not a large one.


Score

Anchor

0

Net time loss. The tool costs more time than it saves.

1 to 2

Negligible net savings, under 10 percent of task time. Overhead nearly equals savings.

3 to 4

Marginal savings, 10 to 25 percent. Useful only for specific task types.

5 to 6

Moderate savings, 25 to 50 percent. Net positive after overhead across common cases.

7 to 8

Strong savings, 50 to 75 percent. Clearly faster for most users most of the time.

9 to 10

Transformative savings, 75 percent or more. Tasks that took hours take minutes.


Dimension 2: Quantity Benefit


Question: how much does this tool increase the volume of output you can produce in a given time?


What is measured: the output multiplier against your own baseline in the same time; whether the added volume holds acceptable quality; and whether you use the time saved to produce more at all.


Quantity without quality is the Quantity Illusion, which is scored separately under Imposture Risk. The framework counts usable, verified output, not drafts.


Score

Anchor

0

No increase in output volume.

1 to 2

Negligible increase, 1.1 to 1.3 times output.

3 to 4

Marginal increase, 1.3 to 1.8 times. Some tasks produce more, others do not.

5 to 6

Moderate increase, 1.8 to 3 times. Consistently more usable work in the same time.

7 to 8

Strong increase, 3 to 5 times. Multiplies productivity across most use cases.

9 to 10

Transformative increase, 5 times or more. An order of magnitude more usable work.


Dimension 3: Quality Benefit


Question: does this tool help you produce better work than you could with human intelligence alone?


What is measured: improvement against your own baseline, not against an expert; the consistency of that improvement; and whether the improvement survives verification.


AI quality is inconsistent. A tool can be excellent at a hard task and unreliable at an easy one, which the framework calls the Jagged Frontier. Polished output that contains subtle errors is not a quality benefit. This is the dimension most exposed to the Quantity Illusion, and the score reflects verified, durable quality rather than surface finish.


Score

Anchor

0

Quality degradation. The output is worse than you would produce alone.

1 to 2

Negligible improvement. Different, not meaningfully better.

3 to 4

Marginal improvement, in format or structure rather than substance.

5 to 6

Moderate improvement. Clear gains in common cases, consistent after verification.

7 to 8

Strong improvement. A non-expert produces near-expert output across most tasks.

9 to 10

Transformative improvement. Quality you could not reach alone, verified and durable.


Dimension 4: Knowledge and Skill Benefit


Question: does this tool help you learn what you did not know, and do what you did not know how to do?


What is measured: the knowledge or skill you acquire through using the tool, not the output it produces for you; whether the tool builds lasting capability or creates dependency; and whether it extends your range into domains you could not previously work in, which U365 calls Omni-Skilling.


This is the highest-risk dimension. A tool that writes code for a non-programmer looks like skill-building and may be skill-offloading: you appear to program without learning to program. The framework scores conservatively here, and the v1.2 clause on agent-authored procedural memory sets a floor on the related Imposture rating.


Score

Anchor

0

Skill erosion. Sustained use degrades your existing capability, the pattern U365 calls AI Obesity.

1 to 2

No skill benefit. Output without learning. Pure delegation.

3 to 4

Marginal benefit. Some patterns or vocabulary, without real capability.

5 to 6

Moderate benefit. Working knowledge in new areas through guided use. The tool teaches while producing.

7 to 8

Strong benefit. Genuine new skills through sustained use. More capable with and without the tool.

9 to 10

Transformative benefit. A true Omni-Skilling engine, with lasting expertise across domains.


The overall score and its five bands


CI-First Benefit Score = (Time + Quantity + Quality + Skill) / 4


Round to one decimal place. A tool scored Time 7, Quantity 6, Quality 8, Skill 5 has an overall score of (7 + 6 + 8 + 5) / 4 = 6.5.


Range

Band

Meaning

0 to 2.0

CI-First Negative

The tool likely makes you less effective. CI is below HI. Avoid it or restructure how it is used.

2.1 to 4.0

CI-First Neutral

The tool breaks even. Benefits roughly equal overhead. Use it for specific tasks only.

4.1 to 6.0

CI-First Positive

Clear net benefit. CI is above HI for most users. Worth adopting with disciplined use.

6.1 to 8.0

CI-First Strong

The tool significantly amplifies you. CI is well above HI. A core tool for a Superhuman workflow.

8.1 to 10.0

CI-First Transformative

The tool redefines what you can achieve. Rare. A flagship CI-First tool.


A band is not a marketing word. A score of 4.8 is CI-First Positive, which is a real recommendation, and a review says so plainly rather than inflating the number to reach a higher band.


Four principles that keep the score honest


  • Score net benefit, not gross benefit. Subtract the cost of using the tool in time, attention and verification.

  • Score the common case, not the best case. A tool that is transformative in one narrow scenario and mediocre elsewhere is scored on the common case.

  • Score the user, not the tool. Expert output that a beginner cannot evaluate scores low on Skill Benefit, even when the output is excellent.

  • Be conservative on Skill. This dimension is the hardest to verify and the easiest to fake, so doubt resolves downward.


Back to the TOC

Section icon: The Humics Protection Rating, and how the badge is computed.

The CI-First Evaluation Framework: how U365 scores an AI tool, and what the versions mean

The Humics Protection Rating, and how the badge is computed


The Humics Protection Rating assesses whether a tool protects, leaves neutral or erodes the three human capabilities that Pascal Bornet set out in Irreplaceable and that U365 adopted as the Humics.


It answers a different question from the benefit score: does this tool strengthen the human, or does it create AI Obesity, the cognitive atrophy that follows from over-delegation?


The three Humics dimensions


Creativity is your ability to originate ideas, make unexpected connections and produce work that is genuinely novel. A tool that generates ideas can either spark your own thinking or replace it.


Critical Thinking is your ability to evaluate, question, verify and judge. A tool that presents answers can either train you to test them or encourage you to accept them.


Social Authenticity is your ability to communicate with genuine voice, empathy and interpersonal presence. A tool that drafts communication can either help you refine your own expression or replace it with generic text.


The per-dimension scale


Rating

Label

Meaning

+1

Protects

The tool actively strengthens this capability. You become better at it, with or without the tool.

0

Neutral

The tool neither strengthens nor weakens this capability. It operates outside this Humic, or balances protection and erosion.

-1

Erodes

The tool actively weakens this capability. Sustained use makes you less capable in this area without the tool.


Badge arithmetic


Humics Protection Score = Creativity + Critical Thinking + Social Authenticity


The score ranges from -3 to +3, and the badge follows from the sum:


Score

Badge

Criteria

Meaning

+2 to +3

Humics-Friendly

Protects or amplifies at least two of the three Humics, with no erosion in any dimension.

Sustained use grows your core human capabilities.

-1 to +1

Humics-Neutral

Neither consistently protects nor consistently erodes. Mixed or balanced effects.

Safe to use, but it does not build capability. You must manage which capabilities you exercise.

-2 to -3

Humics-Risky

Erodes two or more Humics, or erodes one severely with no compensating protection.

Sustained use causes cognitive atrophy. Use it with active mitigation and regular practice without the tool.


The arithmetic is a sum, not an average, so one +1 cannot cancel two -1 ratings into a neutral badge. That is deliberate: two eroded capabilities with one protected capability is a risky tool, not a balanced one.


The questions behind each rating


For Creativity: does the tool spark your ideas or replace your ideation, and after a month of use are you more or less able to generate ideas without it?


For Critical Thinking: does the tool require you to evaluate and verify, or does it invite blind trust, and after a month are you more or less able to analyse independently?


For Social Authenticity: does the tool help you express your own voice or replace it, and after a month does your communication sound more or less like you?


The clause on agent-mediated conversation


Framework v1.2 added clause 4.2-a, and it is the one place where readers most often misread the rating.


Agent-mediated conversation is not erosion by itself. The dimension measures your own communicative capability, not the composition of the channel you work in. A shared room in which named agents and their owner coordinate internally is neutral, because the owner still writes their own messages to other humans. Erosion requires one of two conditions: agent-authored text is presented as the person's own voice in a human-facing channel, or the person substitutes agent interaction for human contact. Agent-to-agent messages that the human never reads are an oversight problem, assessed under Critical Thinking rather than under Social Authenticity.


Back to the TOC

Section icon: AI Imposture Risk, and the three traps.

The CI-First Evaluation Framework: how U365 scores an AI tool, and what the versions mean

AI Imposture Risk, and the three traps


The AI Imposture Risk assessment estimates how likely a tool is to trap you in one of the three usage illusions U365 defines. Each trap is rated Low, Medium or High with cited evidence.


The three traps


Time Illusion. The appearance of saving time when net time is lost. Prompting, reading, understanding and checking all take time, and AI output is inconsistent, so the response cannot be trusted without precise control. A tool that is fast on a simple task and slow on a complex one can produce this illusion reliably.


Quantity Illusion. The appearance of delivering work at volume when the output is mediocre or plainly improvable. Trusting AI output without checking is how a poor or false response survives into the deliverable.


Skill Illusion. The appearance of demonstrating a skill while heading toward error without knowing it. The tool produces expert-looking output for a person who cannot evaluate it, and the person concludes that they can do the work.


How each trap is rated


Trap

Low

Medium

High

Time Illusion

Usable output with minimal prompting; verification is quick; net savings are consistent and measurable across task types.

Moderate prompting or configuration; some verification; simple tasks are fast and complex tasks need iteration.

Extensive prompting or configuration; significant correction; the tool collapses on complex tasks; users report spending more time managing it than it saves.

Quantity Illusion

Output is clearly correct or clearly wrong, so verification is easy; the tool flags uncertainty or cites sources.

Plausible output that needs spot-checking; some outputs carry subtle issues that moderate effort catches.

Large volumes of polished output with subtle errors; the user is tempted to ship unverified work because it looks complete.

Skill Illusion

The tool requires or develops the underlying skill, teaches while producing, and the user can separate their own capability from the tool's.

The tool can be used productively without full understanding, and the user is aware of the gap; effort can build skill alongside it.

Expert-looking output for users who lack the skill to evaluate it; the tool masks the gap, so errors cannot be detected, and dependency replaces a skill the user once had.


The overall level, and the precedence sentence


Rating

Criteria

Low

All three traps are Low. The tool is transparent, reliable and does not mask its limitations.

Medium

One, two or three traps are Medium, or one trap is High with clear mitigations stated. The tool requires disciplined use.

High

Two or more traps are High, or the tool's core value proposition is built on an illusion. Use it only with strong safeguards.


Framework v1.3 added one sentence after that table, and it changed how the bands are applied:


Where two ratings could apply to the same trap set, the higher rating governs.


That sentence closed a gap. Before v1.3, the Medium band spoke of "clear mitigations" without saying what counted, and a trap set that could be read as Medium under one limb and High under another had no stated resolution. The rule now is the stricter reading, every time.


The clause on agent-authored procedural memory


Clause 5.2.3-a, added in v1.2, sets a floor: rate Skill Illusion no lower than Medium for any tool that writes procedural memory on your behalf, such as an agent that creates or revises your skills, memory stores or standing instructions. The reason is that the artefact is durable. You hold a documented capability you did not write, and it is reused in later sessions, so the illusion outlives the task. Rate it High where the agent can revise that memory during use without a per-write human decision, or where you have no routine practice of reading what was written.


An important limit: the clause governs the floor only. A tool can clear the clause and still be rated High on Skill Illusion for a different mechanism, such as expert-looking output you cannot check.


The verification checklist that every workflow carries


The framework requires a verification checklist for every workflow a review describes. The checklist is part of the instrument, not decoration, because it is how the Executive Safeguard becomes operational: always assume you are working with the worst AI available.


Each workflow specifies at least one method from each of three tiers.


Tier 1, multi-model check. Run the same prompt through a second model, from a different provider where possible, and compare. If the outputs diverge significantly, investigate before proceeding.


Tier 2, external source verification. Check factual claims against at least one independent source outside the tool. For code, run it. For analysis, check the data and the assumptions, and not just the conclusion.


Tier 3, human review. Any output that other people will see gets a human review before it ships. If you lack the expertise to check the output, consult someone who has it.


The template is fixed:


VERIFICATION CHECKLIST for [Workflow Name]:


  • ☐ Multi-Model Check: [which second model, and what to compare]

  • ☐ External Source: [which independent source to verify against]

  • ☐ Human Review: [who reviews, and what they check]

  • ☐ CI-First Test: Can the user explain and defend the output without the tool? [Y/N]


The last line is the one that decides the case. If you cannot explain and defend the output without the tool, the output is not ready.


Back to the TOC

Section icon: The five AI Profiles, and why the profile comes before the role.

The CI-First Evaluation Framework: how U365 scores an AI tool, and what the versions mean

The five AI Profiles, and why the profile comes before the role


Every tool is classified into one or more of the five AI Profiles defined by U365. The profile says what role the AI plays in the working relationship, which determines how you should interact with it.


The institutional canon of 2026-11-11 fixes the names and the order. Tools Reviews write the level number beside the name as a shorthand:


Level

Profile

What it means

What the evaluation expects

1

AI as Co-Creator and Thought Partner

The AI collaborates on ideation, strategy and creative work. You and the tool build on each other's thinking.

Iterative dialogue, rich context, output you can build on. Tools that close the dialogue or return final answers score lower on Creativity protection.

2

AI as Co-Worker and Assistant

The AI handles execution: drafting, formatting, data processing, routine work. You direct and review.

Reliable output with minimal correction. Tools that need heavy correction score lower on Time and Quantity because they create the Time Illusion.

3

AI as Professor, Coach and Tutor

The AI teaches, explains, gives feedback and helps you develop skill. You learn.

Lasting capability. Tools that answer without teaching score lower on Knowledge and Skill because they create the Skill Illusion.

4

AI as Co-Researcher, Analyst and Tester

The AI analyses data, tests hypotheses, finds patterns and stress-tests your work. You interpret and decide.

Transparent method and verifiable results. Tools that give conclusions without showing their work score lower on Critical Thinking protection.

5

AI as Challenger, Critic and Devil's Advocate

The AI challenges your assumptions, finds counterarguments and tests your thinking. You defend or revise.

Genuine challenge rather than a strawman. Tools that agree with you score lower because they do not serve the role.


Primary and secondary profiles. Most tools have one primary profile, the main way they are used, and one or more secondary profiles, the other ways they can be used. A review states both. A tool whose primary profile is Co-Worker and Assistant can carry Co-Researcher, Analyst and Tester as a secondary profile when it also reads and checks long documents. The number in brackets is the profile's position in the institutional list and is used as shorthand in reviews; the name is what carries the meaning.


Why the profile comes first. The method is explicit: attributing a profile to AI before giving it a role is essential. The profile decides what you expect from the tool and what you must do yourself. If you assign a task before you decide the profile, you get whatever the tool's default behaviour produces, which is usually execution when you needed teaching, or agreement when you needed challenge. Naming the profile first is a decision about the working relationship, taken before the work starts.


Back to the TOC

Section icon: Superhuman usage guidance, and the over-delegation warning.

The CI-First Evaluation Framework: how U365 scores an AI tool, and what the versions mean

Superhuman usage guidance, and the over-delegation warning


The benefit score, the Humics rating and the Imposture rating are measurements. The usage guidance is the instruction that follows from them. Every Tools Review carries it, and it answers one question: how do you use this tool to become Superhuman rather than Sub-human?


When to invite the tool


  • Tasks where the tool's strongest dimensions apply, such as drafting for a tool that scores high on Time.

  • Tasks where the recommended profile fits naturally.

  • Tasks where the Imposture Risk is manageable with the standard verification for that workflow.


When to keep the tool out


  • Tasks where the Humics are the primary value: original ideation, ethical judgement, empathetic communication.

  • Tasks where the Imposture Risk is High and the cost of verification exceeds the benefit.

  • Tasks where you lack the knowledge to verify the output, which is the Skill Illusion trap.

  • Tasks where using the tool takes longer than doing the work yourself, which is the Time Illusion trap.


The over-delegation warning


Every review states what over-delegation looks like with that specific tool. The warning is tied to the formula, not to a general caution. If you stop doing the thinking, HI falls, and CI falls with it even when the tool stays the same. The Superhuman becomes the Sub-human, and the tool has not changed at all.


That is why the warning is written per tool. Over-delegation with a drafting assistant looks like shipping the first draft unread. Over-delegation with an analysis tool looks like accepting a conclusion whose data you never opened. Over-delegation with a coding tool looks like merging code you cannot debug. The pattern is the same in each case and the visible sign is different.


Back to the TOC

Section icon: What a framework version is, and what each release changed.

The CI-First Evaluation Framework: how U365 scores an AI tool, and what the versions mean

What a framework version is, and what each release changed


This is the part of the framework that readers ask about most often, because every Tools Review states a version and few explain it.


The version governs the instrument, not the tool


A framework version describes the scoring instrument. A tool version describes the product being scored. They move independently.


When a review says "Framework v1.4.1, scoring date 2026-11-18, tool version GPT-6.2 Astra", three different things are being recorded: the instrument used, the day the evidence was gathered, and the product as it stood. A new model release does not change the framework. A framework release does not re-score a model.


Two kinds of release: those that can move a score and those that cannot


A framework release falls into one of three classes, and the class tells you whether a comparison across versions is safe.


A release that changes the arithmetic or the rubric can change a score. This class is rare and none of the releases so far belongs to it: the four dimensions, the 0 to 10 scale, the mean, the Humics sum and the Imposture bands have been stable since v1.1.


A release that adds or removes a rule with a directional effect can move one rating without touching the arithmetic. v1.2 is the example: the Skill Illusion floor can only raise a rating, never lower it. v1.3 is a clarification rather than a change of arithmetic, but it matters: it closed a band hole and added the precedence sentence, so a trap set that could be read two ways is now read the stricter way.


A release that changes only wording moves no score. v1.4 and v1.4.1 are both in this class. v1.4 removed a classification dimension that sat outside every score, and v1.4.1 restored an attribution. Neither touched a number.


Why a review can say "Framework v1.2 clause note" when the current framework is 1.4.1


A clause note records when a rule entered the framework. It is provenance, not the version in force.


Clauses 5.2.3-a, 4.2-a and 7.5 entered in v1.2 on 2026-09-18. v1.4 removed 7.5 and kept the other two. So a review written today applies the clause on agent-authored procedural memory and the clause on agent-mediated conversation, and both are v1.2 clauses. The review says "Framework v1.2 clause note" because that is where the rule came from, and states the framework version in force separately, which is 1.4.1.


The practical test: if a clause appears in a review, it is in force. If a review names a clause and the current framework document no longer contains it, the review is wrong and gets corrected. That happened once, to the multi-agent room clause when v1.4 removed it.


Whether scores are comparable across versions


Yes, for every release so far, with one reading caution.


The four dimensions, the scale, the mean, the Humics sum and the Imposture bands have not changed since v1.1. A 6.5 recorded in September 2026 means the same as a 6.5 recorded in November 2026. The framework has never re-based its scale.


The caution concerns ratings rather than scores. A review scored before 2026-09-18 was not subject to the Skill Illusion floor, so a Low Skill Illusion rating in an older review can legitimately differ from what the same evidence would produce today. A review scored before 2026-10-06 predates the precedence sentence, so its Imposture band may have been set under the looser reading. Neither case changes the published score, and both are the reason every review records its framework version and scoring date.


When you compare two reviews, check three things in this order: the framework version on each, the scoring date on each, and whether the finding you are comparing depends on a clause or on a band condition. If it depends on neither, compare the numbers directly.


How a release is made, and who approves it


URC owns the framework methodology and the versioning, a responsibility transferred on 2026-09-14. UDA authors the pedagogical foundations and holds a governance veto on any change to the framework's pedagogical dimensions, so a change that touches a pedagogical dimension is declared only after the academic side has confirmed it. That is why the version history records an academic ruling for v1.2, for v1.4 and for v1.4.1.


The sequence is the same each time:


  • A defect, a gap or a ruling is recorded, with the evidence that motivates it.

  • URC drafts the change against the framework text and states what it does to scores, bands, badges and published reviews.

  • Where a pedagogical dimension is touched, the academic veto applies before the release is declared.

  • URC applies the change, verifies the applied text against the ruling, and declares the version.

  • The Current version box and the version history table in this publication are updated in the same pass, because this publication is the public record of the instrument.

  • Carrier text in already published reviews is checked. Where a wording change reaches published posts, the affected carriers are updated and counted, and the count is stated in the release note.


Version numbers follow a simple rule: the minor number rises for a change of substance, and the patch number rises for a correction to the immediately preceding release. v1.4.1 is a patch because it corrects an over-application inside v1.4, not because it adds a rule.


Back to the TOC

Section icon: The version history, release by release.

The CI-First Evaluation Framework: how U365 scores an AI tool, and what the versions mean

The version history, release by release


Version

Date

What changed

Why, and who ruled

Did any score move

1.1

2026-08-24, updated 2026-09-14

First published version. Incorporated the eight differentiation angles: a CI-First Profile per tool, the Humics impact badge, Imposture Risk scores with evidence, and a per-workflow verification checklist. Methodology ownership transferred to URC on 2026-09-14; UDA kept authorship of the pedagogical foundations and its veto.

Alick Mouriesse supplied the eight angles on 2026-08-24. The ownership transfer was approved on 2026-09-14.

Not applicable. This is the baseline.

1.2

2026-09-18

Added three pedagogical clauses: 5.2.3-a on agent-authored procedural memory, with a Skill Illusion floor of Medium and High under the stated conditions; 4.2-a on agent-mediated conversation; and 7.5 on team-level rooms.

Ruled by the Dean of Academics under the pedagogical veto on 2026-09-18. The applied text was verified against the ruling before release.

No. No published post, CMS record or recorded scorecard changed. The clauses apply to future assessments.

1.3

2026-10-06

Rewrote the Medium band in Section 5.3 so it counts one, two or three Medium traps, replaced "clear mitigations" with "clear mitigations stated", and added the sentence "Where two ratings could apply to the same trap set, the higher rating governs."

The Dean of Academics ruled on 2026-09-26 that the band text had a hole. Research applied it, and the applied wording was verified against the ruling. The release was declared on 2026-10-06.

No score, band, badge, sub-score, profile or trap rating changed in value. 0 of 94 published posts changed value under either wording.

1.4

2026-10-13

Removed the collaboration-mode dimension, the Centaur and Cyborg modes, and subsection 7.5. Top-level section numbers were kept so existing citations stay stable.

Alick Mouriesse ruled on 2026-10-13 that U365's CI-First methodology does not use those collaboration modes. The academic confirmation gated the release.

No. The mode was a derived classification outside the benefit arithmetic, the badge sum and the Imposture bands. No published post, CMS record or scorecard changed value.

1.4.1

2026-10-14

Restored the Section 14 attribution of the term Co-Intelligence and the starting idea of CI-First to Ethan Mollick's book Co-Intelligence (2024). Re-scoped the v1.4 wording to modes only, so the collaboration-mode removal stands in full.

The 2026-10-13 de-attribution was an over-application of a task-scoped instruction that covered only the Centaur and Cyborg terms. Alick Mouriesse reversed it on 2026-10-14; the wording was ruled on the academic side and applied by Research.

No. No published post, CMS record or recorded scorecard changed value.

1.4.1, name alignment

2026-11-11

Aligned the five AI Profile names with the institutional canon of 2026-11-11 in every place the framework states them. No methodology change, no rule change.

Alick Mouriesse issued the precision fix on 2026-11-11; the canon of record was updated and the framework file aligned the same day.

No. Names only.


Read the table as a record of the instrument's stability. Across six entries spanning under three months, the arithmetic has not changed once. What changed is the wording, the band conditions and the clauses that set floors, which is why a review's framework version tells you what rules were in force and not which scale was used.


Back to the TOC

Section icon: How to read a Tools Review score: GPT-6.2 Astra at 6.5.

The CI-First Evaluation Framework: how U365 scores an AI tool, and what the versions mean

How to read a Tools Review score: GPT-6.2 Astra at 6.5


The fastest way to understand the instrument is to read one completed record. The review of GPT-6.2 Astra, published on 2026-11-18, is a useful example because its score sits in the middle of a band and its ratings disagree with each other, which is the normal case.


The record


Field

Value

CI-First Benefit Score

6.5 out of 10, CI-First Strong

Sub-scores

Time 8, Quantity 7, Quality 7, Knowledge and Skill 4

CI-First Profile

Primary: Co-Worker and Assistant (level 2). Secondary: Co-Researcher, Analyst and Tester (level 4), and Co-Creator and Thought Partner (level 1)

Humics Protection

-1 out of +3, Humics-Neutral. Creativity 0, Critical Thinking -1, Social Authenticity 0

AI Imposture Risk

High overall. Time Illusion Medium, Quantity Illusion High, Skill Illusion High

Framework version

1.4.1, scoring date 2026-11-18


Reading it, step by step


Step 1, the overall score. (8 + 7 + 7 + 4) / 4 = 6.5. The band is 6.1 to 8.0, which is CI-First Strong. The tool significantly amplifies a professional who supervises it.


Step 2, the sub-scores. Time 8 says the net saving survives the overhead: strong savings on documents, workbooks, legal drafts, design and science work, measured against a real cost in effort decisions and long runs at the top setting. Quantity 7 says more usable output in the same time, held below 8 because a model that stops early produces more drafts rather than more finished units. Quality 7 says measured gains on the vendor's artefact and science rows and on the independent index, held below 8 by a flat independent coding-agent index and a high hallucination reading. Skill 4 says execution rather than learning: the artefact looks correct whether or not it is, the tool has no teaching mode, and the first-days record says it needs more oversight rather than less.


Step 3, the Humics badge. Creativity 0, Critical Thinking -1, Social Authenticity 0 sums to -1, which is Humics-Neutral. The badge is not an average of opinion: the erosion is specific and named, in the dimension where a fluent artefact is harder to challenge than a rough one.


Step 4, the Imposture level. Two traps are High and one is Medium. Two or more High traps is High by the band table, and the precedence sentence gives the same answer. The mitigations in the review lower the frequency of failure; they do not change the rating, because the rating describes the trap, not the reader's discipline.


Step 5, the profile. Primary Co-Worker and Assistant, because the main use is execution that you direct and review. Co-Researcher, Analyst and Tester as a secondary profile for long-document analysis, and Co-Creator and Thought Partner for design direction. The profile tells you what to expect before you open the tool, not after.


What would move the score


A release that fixed the obstacle-omission behaviour and improved the independent hallucination reading would raise Quality, and the Skill sub-score would move only if the model began teaching while producing, which no release note suggests.


The point of the worked example is that a high benefit score and a High Imposture Risk can coexist in the same honest record. The tool is worth adopting for the work the review names, and the work has to be supervised. A review that resolved that tension by lowering the benefit score, or by softening the risk rating, would be easier to read and less useful.


Back to the TOC

Section icon: What the framework does not measure.

The CI-First Evaluation Framework: how U365 scores an AI tool, and what the versions mean

What the framework does not measure


The framework is specific about its own limits, because a measurement that claims too much stops being trusted.


It does not measure adoption cost beyond the four dimensions. Licence price is recorded in a review and it is not a scored dimension. Neither is the cost of training a team, migrating existing work, integrating the tool with your systems, or governing its use. A cheap tool can be expensive to adopt, and the framework does not pretend to price that.


It does not measure organisational politics. Whether a team will accept a tool, who sponsored the purchase, and which group loses influence when a workflow changes are real determinants of value and sit outside the instrument.


It does not measure procurement. Contracts, data residency, security review, sub-processors, retention terms and compliance obligations are recorded where they are relevant to use and are not part of the score. A tool can score 8.0 and still be unusable in your jurisdiction.


It does not measure anything the vendor does not disclose. Where a capability is unmeasured, the review records the gap rather than filling it. No independent re-run, no verified leaderboard row and no published evaluation means the review says so. The same rule applies to a vendor's own comparison: it is reported as the vendor's and labelled.


It does not measure supplier conduct inside the score. A supplier's access regime, monitoring, safeguard disclosures or programme eligibility are conditions of use. A review records them in its own section and states explicitly that no sub-score changed.


It does not predict whether your team will use the tool well. The framework scores a tool as an honest user would experience it, including realistic verification habits. The score is not a promise about your organisation.


Read the score as one honest measurement among several inputs. It tells you what the tool gives a disciplined user, what it takes from the human side, and which traps it sets. The adoption decision also needs your price, your risk tolerance and your work.


Back to the TOC

Section icon: Glossary.

The CI-First Evaluation Framework: how U365 scores an AI tool, and what the versions mean

Glossary


AI Imposture Risk. The likelihood that a tool traps you in one of the three usage illusions. Rated Low, Medium or High on each trap, then combined into an overall level.


AI Obesity. Cognitive atrophy from over-delegation. The CI-First strategy defends against it by strengthening the Humics.


AI Profile. One of five roles the AI plays in a working relationship. Attributed before a task is given.


CI-First. Co-Intelligence-First. The principle that you always invite AI into reflection and work, while human intelligence remains the ruler and orchestrator. Expressed as CI = HI + (AI x HI).


CI-First Benefit Score. The arithmetic mean of the four benefit dimensions, each scored 0 to 10, rounded to one decimal place.


CI-First Benefit bands. The five ranges: 0 to 2.0 CI-First Negative, 2.1 to 4.0 CI-First Neutral, 4.1 to 6.0 CI-First Positive, 6.1 to 8.0 CI-First Strong, 8.1 to 10.0 CI-First Transformative.


Clause note. A line in a review naming the framework version that introduced the clause being applied, such as the v1.2 clause on agent-authored procedural memory. It records provenance and is not the version in force.


Executive Safeguard. The operating rule "always assume you are working with the worst AI available", which forces verification and keeps you in the orchestrator's seat.


Framework version. The release of the scoring instrument, recorded on every review together with the scoring date and the tool version.


Humics. The three uniquely human capabilities defined by Pascal Bornet: creativity, critical thinking and social authenticity.


Humics Protection Rating. The sum of the three Humics dimensions, each rated +1 protects, 0 neutral or -1 erodes, giving a score from -3 to +3 and a badge.


Humics-Friendly, Humics-Neutral, Humics-Risky. The three badges. +2 to +3 is Friendly, -1 to +1 is Neutral, -2 to -3 is Risky.


Jagged Frontier. The pattern in which AI is excellent at some hard tasks and unreliable at some easy ones, which is why quality must be verified rather than assumed.


Knowledge and Skill Benefit. The fourth dimension: whether the tool helps you learn what you did not know and do what you did not know how to do. Scored as Skill.


Omni-Skilling. Extending your range into domains and capabilities you could not previously cover. The framework separates real Omni-Skilling from skill-offloading.


Primary and secondary profile. The main role a tool plays and the other roles it can play. A review states both.


Quantity Benefit. The second dimension: whether the same time produces more usable, verified output.


Quality Benefit. The third dimension: whether the output is better, verified and durable rather than polished.


Skill Illusion. The appearance of demonstrating a skill while heading toward error without knowing it.


Sub-human. The person whose HI falls through over-delegation, so that the co-intelligence result falls even with strong AI.


Superhuman. The person whose capacities are measurably amplified through sustained co-intelligence, while remaining decisively human and keeping control.


Time Benefit. The first dimension: net time saved after the overhead of using the tool.


Time Illusion. The appearance of saving time when net time is lost.


Quantity Illusion. The appearance of delivering work at volume when the output is mediocre or improvable.


Verification checklist. The per-workflow check with a multi-model check, an external source check, a human review and the CI-First test, which asks whether you can explain and defend the output without the tool.


Section icon: Sources.

The CI-First Evaluation Framework: how U365 scores an AI tool, and what the versions mean

Sources


  • U365 CI-First Evaluation Framework for AI Tools, version 1.4.1, methodology owner URC, author of the pedagogical foundations Martin Swartz, Dean of Academics. The framework of record for every statement, table and rule in this publication.

  • University 365 UP-Context Profile, institutional method definitions: CI-First, the four AI benefits, the three AI Imposture Traps, Humics, the five AI Profiles, UNOP, ULM, LIPS, SL-OS and UP-Context.

  • Discover the CI-First Approach, INSIDE, for the doctrine and the formula.

  • The CI-First Workflow: Applied to Any Field, INSIDE, for the working sequence.

  • How to Build a CI-First Company, INSIDE, for the organisational application.

  • GPT-6.2 Astra: OpenAI's Frontier Model for Professional Deliverables, Priced Below the Model It Replaces, INSIDE Tools Review, 2026-11-18, the worked example in this publication.

  • Pascal Bornet, Irreplaceable: The Art of Standing Out in the Age of Artificial Intelligence, origin of the Humics.

  • Ethan Mollick, Co-Intelligence: Living and Working with AI (2024), origin of the term Co-Intelligence and the starting idea of the CI-First approach.

  • Adam Brotman and Andy Sack, AI FIRST: The Playbook for a Future-Proof Business and Brand (2025), an inspiration for the CI-First approach.

  • The eight differentiation angles, provided by Alick Mouriesse on 2026-08-24: a CI-First Profile per tool, the Humics impact badge, Imposture Risk scores with evidence, and a per-workflow verification checklist.



Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
Image by Erik  Lucatero

Become Superhuman

Master AI to stay irreplaceable in every field.

 

 

 

​

​

Apply for Admission Today.
Select Your Initial Access Level.


Become a DISCOVERY, INSIDER, or SUPERHUMAN Fellow.

Image by Milad Fakurian

Master Your Life with a Digital Second Brain

Turn overwhelm into clarity with LIPS + CARE
U365’s unique framework to organize your goals, projects, and knowledge into a superhuman system for success

bottom of page