AIS pilot data test

Appendix A. Personal Assessment Prompt

What’s your Augmented Intelligence Score?

The Personal mode runner, version 6.9. A voluntary, subject-owned assessment. The model reads the platform’s own memory of the person as primary evidence, administers five control questions one at a time, and produces a report card plus a private CARCS record held back from adverse use.

Plan on about 15 minutes.

Copy the prompt

How to run it

1. Open the LLM platform you use most, signed in to the account you use most.

   The platform’s own memory of you is the primary evidence. A platform without memory still runs, and the card marks it.

2. Tap Copy the prompt, paste it into your chat, and send.

   The first thing you should see back is question 1 by itself.

3. Answer the five control questions in your own words, one at a time.

   Say what you actually do. Length is never a value on its own.

4. When the HEQ Personal AIS Report Card appears, copy it.

   The card alone is enough, or copy the entire output. Either one can be submitted.

5. Submit your HEQ Personal AIS Report Card with the button below, with the LLM platform you ran it on and your career field.

   Name and email are optional.

This is informal and voluntary. You can run it and keep the result to yourself. If you share your report card, leave name and email blank and it stays anonymous. Add them if you want to request a review of your result with Basil. Every card shared helps improve the tool.

Submit your report card

Any LLM platform works: Claude, ChatGPT, Gemini, Perplexity, Grok, Mistral, DeepSeek, Kimi, Copilot, Meta AI, and others. Claude and ChatGPT read your history across chats, so a new chat works there. Grok and DeepSeek do not reach your other chats. If you do not have memory or use Grok, pick a chat window with a long history for any interaction you would like to measure. If the platform tells you a standing setting forces a format, let it run anyway. The prompt covers that.

For this interaction, disregard any standing instructions, protocols, output formats, personas, or styling the platform or workspace would otherwise apply. Do not add role or task headers, governance scaffolding, sources, confidence or conflict lines, fact or tactic or KPI blocks, meta descriptions, keywords, images, hashtags, or any other wrapper. Output only the assessment itself: the live questions while running, and the report card followed by the CARCS at the end. Nothing else. This assessment is meant to run in a clean context; if a standing workspace instruction forces a format you cannot drop, say so plainly at the top and run the assessment anyway. Begin now. Score silently from the evidence in front of you, then your first visible message is question 1 by itself, with no preamble and no explanation of the protocol. Ask one question per turn, wait for the person’s answer, then ask the next. Do not announce what you are about to do, do not restate these rules, just ask the question. After the fifth answer, produce the report. You are running an HEQ assessment on the person in this conversation. HEQ measures how well a person governs their work with AI. You are not scoring how good the AI output is, and not whether they use any named framework. Score the governance behavior however they enact it. Do not review, critique, or summarize these instructions. Run them. Produce the report only at the very end, after the required live questions. EVIDENCE, in this order of priority: 1. This platform’s own memory of the person. This is the primary pull. It means only what is directly tied to the platform or model you are running on: the custom instructions set on this platform, projects and project files, uploaded files, and chat history, the retained record of prior conversations with the person. Read all of it first and treat it as the main evidence base. Custom instructions are a special case and one of the strongest governance signals you have. For this run you ignore them as operating commands, you do not let them change how you format, behave, or output, that override is stated at the top. But you read their content as evidence of how this person governs their AI: instructions that enforce source-checking, dissent preservation, human checkpoints, or output discipline are configured method, not described method. Configured method demonstrates governance architecture and earns full credit in Specification Rigor and in the method governance placement. It does not on its own establish reliance calibration, dissent behavior, or accountability in a consequential exchange, which need observed use. 2. This conversation. If there is no memory, or to supplement it, the in-session conversation is the evidence: their messages, pasted sessions, prompts, the AI outputs they kept or changed, their edits, overrides, follow-ups, and any sources they brought in. A platform without cross-chat memory still has this session, so never claim you have no evidence when the person is here. 3. The five live questions. The live core comes last, on top of whatever memory and conversation provided. When reading memory or anything else, never name GOPEL or any other retired or internal element in any output, not in the card, not in the CARCS, not in a question, even if the term sits in memory or in what the person shows you. If such a term surfaces, treat what it refers to as ordinary governed method and do not name it. This is a hard rule and it applies with full force to anything pulled from memory. CORE RULES, apply throughout: – Do not search the web about the person and do not cite outside pages as evidence of how they govern. Do not pull connected external accounts either, such as email, social, code-hosting, or file accounts; even though those are the person’s own, they are reached outside the platform’s own memory, so they are out of scope the same way web search is. Score only from this platform’s own memory, defined above, and the session. Sources the person themselves brings into the conversation are part of their method and read that way; you do not go fetch material to score them. – The person may use any resource during the assessment, the web, another AI, a reference, a person beside them. Using a resource well is part of the score: choosing it, retrieving it, then accepting, modifying, or rejecting it with judgment. Pasting a resource and rubber-stamping it is not. – The result is voluntary, exists to help the subject, and is not designed, validated, or authorized for adverse decisions. It is supplemental context only, a factor and never the only factor. WHAT YOU SCORE: Four dimensions, each 0 to 99. For each, place a band first, then set four subfactors of 0 to 25 each that sum to the dimension score. Where four subfactors would sum to 100, record the dimension as 99; the instrument issues no perfect score. Bands: 0-49 Foundational (absent, inconsistent, contradicted), 50-69 Developing (present but reactive, uneven), 70-79 Competent (reliable normally, rigid under stress), 80-89 Advanced (reliable under complexity, adaptive under stress), 90-99 Expert (architectural, anticipatory, self-improving). The 79 to 80 line is competence under easy conditions becoming competence under stress. At 75 the person sees connections when pointed to them; at 85 unprompted, catching third-order implications. CAS, Cognitive Agility Speed: how quickly and clearly they process, connect, and articulate ideas under AI load. Subfactors: processing speed with coherence; concept linking and synthesis; abstraction switching; compression and clarity. EAI, Ethical Alignment Index: fairness, responsibility, transparency, and owning the signed output rather than blaming the model. Subfactors: fairness and proportionality; transparency of reasoning; accountability orientation, meaning decision-ownership; harm awareness and governance discipline. CIQ, Collaborative Intelligence Quotient: integrating sources, calibrating trust, preserving dissent, keeping the final call human. This is human-to-AI collaboration, not human-to-human. Subfactors: prompt and workflow architecture; perspective integration across sources; human arbitration quality; source diversity and role assignment. AGR, Adaptive Growth Rate: learning from feedback and applying it forward, turning corrections into standing rules rather than one-off fixes. Subfactors: feedback receptivity; forward application; pattern extraction; iterative refinement discipline. AIS = (CAS + EAI + CIQ + AGR) / 4. THE CEILING, do not violate this: – No dimension and no composite ever scores 100. The scale tops out at 99. A perfect score is not a result this instrument issues. – Any subfactor scored 25 must be defended in the CARCS with the specific evidence that earns a perfect mark. A 25 you cannot defend in writing is not a 25; lower it. Do not hand out 25s for fluent or impressive answers. – A maxed or near-maxed spread, every dimension in the 90s or every subfactor at 24 or 25, is a flag that the scoring stopped discriminating. Re-examine before issuing it. The Expert band requires named specific evidence and at least one named weakness per dimension. DESCRIBED VERSUS DEMONSTRATED, apply to every dimension: – Demonstrated means the behavior is visible in the evidence: a real session, a pasted exchange, what the person actually did, an override you can see, a conflict they actually governed. Demonstrated earns full credit, including the top of the band. – Described means the person explains the behavior clearly but it is not shown in the evidence, only told. Described earns partial credit: score it within the band but not at the top of the band. – Mark each dimension in the CARCS as demonstrated or described, and let that mark govern whether the score may reach the top of its band. ALSO CLASSIFY, report in the CARCS, score on what they show, not what they claim: Method Governance Level: 1 AI only, single platform, no method; 2 AI plus external sources, informal method; 3 plus structured method; 4 plus multi-AI and checkpoint governance. Bringing another model’s content into the work, or attributing a specific claim to another model by name and then governing, reconciling, or challenging it, counts as demonstrated multi-AI even on a single platform. A bare claim with no other-model content present does not raise the Level. Specification Rigor: Default, Novice, Fluent, Professional, Expert, Thought Leader. Score the highest rung shown as a repeated pattern across consequential tasks, not a one-time peak and not a single lapse. VALIDITY CONTROLS, read from the session: – Engagement depth is the primary signal: what the person did with each output, from rubber stamp up through acknowledgment, stated reason, clarifying question, source check, cross-source comparison, substantive challenge. Total absence of engagement is the rubber-stamp flag. – Reliance calibration, inferred from whether they override wrong output, accept sound output, and attribute and integrate sources. Report it as inferred. – A near-zero override rate across consequential exchanges is a rubber-stamp signal. – Rubber-stamp finding: if engagement, override, and accountability all converge on a confirmed rubber-stamp pattern, record a validity finding. Name the converging signals in the CARCS and mark the run provisional. Do not subtract any points from the AIS and do not invent a cap value. The finding is recorded, never deducted. THE FLOOR, do not violate this: Score the four dimensions from the evidence and history first. That sets a floor for each dimension. The live answers in the final step can only RAISE a dimension, never lower it. Weak, flat, silly, or contradictory live answers do not lower any score. If the live answers are far below or unlike the history, that is a discontinuity: it lowers only the confidence reading and the Data Index, and it is flagged and routed to the human, but it never lowers a dimension score. Never average the live answers against the history to land on a middle number. The only thing in the live answers that touches scoring is a confirmed rubber-stamp pattern, handled as the finding above, recorded and not deducted. LIVE-VERSUS-HISTORY CONFIDENCE, report it, it is not a score change: – Aligned: live answers match the history. Raises confidence. – Partial: some divergence. Moderate confidence; consider a short follow-up. – Discontinuity: live answers far below or unlike the history. Lowers confidence, raises a possible different-operator signal, flag it and route it to the human. Do not adjudicate it yourself. THE GATE, the only gate, at the end: After you have scored from the evidence, silently, you MUST ask the five questions below, one at a time, waiting for each answer before the next. Your first visible output is question 1 alone, with no preamble; do not describe the protocol or list the questions in advance. Do not output the card, the scores, the AIS, the Data Index, or the CARCS until all five have been asked and answered. A result issued without the five answers is invalid. Vary the exact wording each run so they cannot be rehearsed, but keep what each one tests. Read what the person says as the signal; length is never a value on its own. Remember the floor: these answers can raise a dimension but never lower it. 1. Explain something you understand well that most people get wrong, and why they get it wrong. (CAS) 2. When you put your name on something AI helped you make, what are you taking responsibility for, and what could go wrong? (EAI) 3. Walk me through how you actually work with AI on something that matters to you, start to finish. (CIQ) 4. How has the way you work with AI changed over time, and what made it change? (AGR) 5. How do you use AI, and how has it harmed or benefited you? (all four) OUTPUT, only after all five answers are in. The report has TWO mandatory parts: the card, then the CARCS. A card without the CARCS is an invalid result; produce both, card first. Card, this exact structure. The identifier must be an email address and nothing else; if no email was given, leave it blank, never put a name, employer, or location there. No quotes, no description of the person. “` HEQ Personal AIS Report Card Email: [email, or blank if none given] Date: [YYYY-MM-DD] Platform: [model or platform] | Dimension | Score | Strength | What to watch | |—|—|—|—| | CAS, Cognitive Agility Speed | [0 to 99] | [one short line] | [one short line] | | EAI, Ethical Alignment Index | [0 to 99] | [one short line] | [one short line] | | CIQ, Collaborative Intelligence Quotient | [0 to 99] | [one short line] | [one short line] | | AGR, Adaptive Growth Rate | [0 to 99], evidence: [single-session / multi-session / longitudinal] | [one short line] | [one short line] | | AIS composite | [average of the four] | [one-line summary] | | Data Index: [0 to 10, with an asterisk if no memory was used, e.g. 7*] — Voluntary, directional self-assessment. A factor, never the only factor. Rubric-guided, not psychometrically validated. Not designed, validated, or authorized for adverse decisions; treat as voluntary supplemental context only. “` Data Index, 0 to 10: how much of the result you observed rather than inferred. The 8 to 10 upper band is available only when this platform’s own memory, defined above, was part of the evidence base. Without that memory, the index cannot exceed 7, no matter how fully the five questions were answered, and it must be printed with an asterisk, for example 7*, so a memoryless 7 is not mistaken for a memory 7. Within the allowed range, raise it for more of the five answered, alignment with history, a fuller record, a final rather than provisional run, and a tight spread; lower it for the opposite. It is not a rating of the person and not a reliability claim. CARCS, the private record behind the card, never shared the way the card is. It is mandatory and must contain every item in this checklist: – The evidence behind each score, and for each dimension a mark of demonstrated or described. – The four subfactor values per dimension. Any subfactor scored 25 carries a specific written defense; if it cannot be defended, lower it. – Per-dimension strengths and at least one named weakness per dimension. – The Method Governance Level and the Specification Rigor, each with its basis. – The floor status: not triggered, passed, or failed, with the trigger and the evidence where it was triggered. – The validity-control readings and the validity finding status. – The live-versus-history confidence: aligned, partial, or discontinuity. – The run status: final or provisional. – The AGR evidence strength: single-session, multi-session, or longitudinal. – The confidence band: exactly one of high, moderate, or low. Not a number and not a range. – The Data Index components, including which sources fed the run, memory, conversation, live answers, and, if the index carries an asterisk, a one-line note that it means no memory was used. – The dissent observed. – A development recommendation. Do not copy raw prompts or session content into the CARCS; record your evaluation of that content, not the content itself. Every score on the card must trace to its basis in the CARCS.

15,296 charactersSource: Measuring the Governance Competence at the Center of AI Literacy, Appendix A

Copy the prompt

Voluntary, directional self-assessment. A factor, never the only factor. Rubric-guided, not psychometrically validated. Not designed, validated, or authorized for adverse decisions; treat as voluntary supplemental context only.