Overview

The MindSpec Interview Framework produces complete agent, character, or self specifications through tiered interactive interview. It is the values-substrate producer for the Ora system — the framework that turns a person’s actual value commitments (or a fictional character’s, or an AI agent’s) into a structured single-file specification that downstream frameworks can read directly. Without MindSpec, the system has no values substrate; with MindSpec, value-aware decisions become possible without per-decision interrogation.

The framework runs in three modes. MSI-Self produces the user’s own personal MindSpec — used as the values substrate that PEF, MOM, and IIF read for value-alignment work. MSI-Agent produces an AI agent’s MindSpec at one of three tiers (Tier 1 ephemeral for one-off tasks; Tier 2 persistent task for recurring roles; Tier 3 personal thinking partner for lifelong, multi-domain agents). MSI-Character produces a fictional character’s MindSpec for narrative and creative use — supports the 24-entry character-spec library with pathology signatures (Contempt, Self-Contempt, Arrogance, Schadenfreude, Resentment, Malice, Spite, etc.) that the agent and self specifications do not need.

The framework’s load-bearing intellectual content sits in five sections of the canonical specification. §II — The Library is the 66-entry instrument across 11 families (42 primary commitments suitable for self, agent, and character; 24 character-spec commitments for fictional use). Each entry carries a structured-field schema (commitment name, definition, weight scale, near-enemies, direct-oppositions, scale anchors, object-modulation specifications, activation-profile specifications). The library was recalibrated in v0.2.2 with universality-audit corrections — 26 of 59 default values were updated to match the general-population median rather than the framework-author residual bias. §IV — Three-Stage Assessment Instrument is the elicitation methodology: Stage 1 (66 portrait scenarios — quick-reaction to one-sentence portraits of value commitments; produces a candidate weight per commitment), Stage 2 (84 dyad scenarios — choices between two value commitments in tension; reveals which commitments dominate when in conflict), Stage 3 (35 multi-commitment scenarios — complex situations involving three or more commitments simultaneously; reveals which commitments are constitutional vs. working) plus 17 constitutional pressure-tests (probes the most load-bearing commitments). §V — Inference Layer handles incompatibility adjustments (when two commitments score high but are direct-opposites, the framework surfaces the conflict transparently rather than silently resolving), selfishness-coefficient weighting (modulates commitments by the user’s stated stance toward self vs. others), and adjustment transparency (every adjustment the framework makes is surfaced in the Adjustment Summary so the user sees what was modified and why). §VI — Learning Architecture specifies six drift-prevention safeguards that prevent the agent’s MindSpec from drifting silently across iterations — periodic recalibration, drift-detection on commitment weights, adversarial reflection passes, ledger appends with reasoning, modifications-log append-only discipline, and explicit version bumps when load-bearing commitments change. §VIIA — Stage 2A Life-Context Direct Pass is the section that grounds the abstract commitment library in the user’s actual life context — health, family, financial, professional, geographical, social, and aspirational fields are populated explicitly (with “none” being a valid answer the framework records explicitly).

The framework’s three-tier structure is the friction discipline applied to values elicitation. Tier 1 (ephemeral) gets library defaults — no assessment, just the right defaults for a single-task agent. Tier 2 (persistent task) gets Stage 1 filtered assessment with light governance — appropriate for an agent that runs a recurring role but does not need full constitutional depth. Tier 3 (personal thinking partner) gets full three-stage assessment with full governance — appropriate for the user’s own self-MindSpec or for an agent intended to operate as a long-term thinking partner. The tier discipline prevents the failure mode where every MindSpec gets the same heavy elicitation regardless of the use case; the framework’s value depends on it being used, and use depends on the elicitation cost matching the use.

The framework answers questions like: I want to use AI for value-aware decision support but I don’t have a way to communicate my actual values — can you produce my MindSpec? I’m building a long-term AI thinking partner and I want it to have a coherent, stable identity rather than drifting per-conversation — can you produce its MindSpec? I’m writing fiction and I want a character whose value structure is internally consistent rather than a bundle of stereotypes — can you produce a character MindSpec?

Systemic context

The MindSpec Interview Framework is the values-substrate producer of the Knowledge Production System (paired with KAC as engram-producer and the Creativity reference as conceptual frame; the system view is in [[Paper — Knowledge Production System]]). Its outputs feed every other system. The [[Paper — Strategic Supervision System]] reads MindSpec for value alignment in MOM’s Resolution Statement Objectivity Protocol Near-Miss Elicitation and in the Excluded Outcomes derivation. The [[Paper — Matrix Lifecycle System]]‘s Inception and Incubation Framework reads MindSpec values directly in Mode 1’s value-alignment math; the Spark Recognition Corpus accumulates the empirical evidence that recognition is irreducibly human (the central claim of the [[Reference — Creativity from Knowledge and Values]] as refined by IIF). The [[Paper — Coordination Layer]] does not read MindSpec directly but supervises its files (modifications.md is append-only; ledger.md is the learning architecture’s record). The framework reads but does not produce other frameworks’ artifacts — the universality-audit methodology lives in Old AI Working Files/Framework — MindSpec Universality Audit and Corrections.md for provenance with corrections already applied throughout the canonical specification.

Ora’s output

Example question: I want to set up my own MindSpec at Tier 3 (personal thinking partner) so my AI tools can do value-aware work for me without interrogating me at every decision. Can you walk me through the interview?

Ora’s response (running MSI-Self at Tier 3):

The Tier 3 interview is substantial — 90–120 minutes of focused interactive work. Here is what each stage produces.

Stage 1 — Portrait scenarios (the 66-entry library, quick reactions). I’ll show you 66 one-sentence portraits, one per library entry. Each portrait describes a value commitment in the form “Imagine someone who…” — your reaction is gut-level, not analytical. Sample:

Portrait — TRUTH-SEEKING. Imagine someone whose primary instinct when encountering a confident claim is to ask “what would I see if this were wrong?” and who genuinely wants the answer even when it would cost them something to find out.

Your response options: This is essentially how I operate / This describes me strongly but not always / This describes me sometimes / This describes me rarely / This describes me almost never.

We work through all 66 portraits. The framework records your gut-reaction weight per entry. Stage 1 produces a candidate weight profile across the full library. (Time: ~30 minutes for the user; the framework keeps pace with reasonable response times and does not push for speed.)

Stage 2 — Dyad scenarios (84 forced-choice between two commitments in tension). This is where the candidate weights get tested against each other. Each dyad presents a scenario where two of your high-weight commitments are in tension and you have to choose. Sample:

Dyad — TRUTH-SEEKING vs. KINDNESS. A friend has just shared work with you that they’re proud of. You see a structural problem that, if pointed out now, will be painful to hear and probably useful for them to know — but they did not ask for critique. Do you (a) raise the issue gently because honesty is owed; (b) raise it only if asked; (c) frame it as a question rather than a critique; (d) hold the critique unless they ask?

You pick. The framework records the choice. Across 84 dyads, the dominance pattern emerges — when truth-seeking and kindness are in tension, which way you actually lean. (Time: ~45 minutes.)

Stage 2A — Life-Context Direct Pass. This is the section that grounds the abstract commitments in your actual life. I’ll ask explicit questions about: health (current state, ongoing concerns); family (composition, dependents, geographical proximity, current obligations); financial (current state, ongoing obligations, runway, dependents on your income); professional (current role, ongoing commitments, change horizon); geographical (where you live, where you work, location-dependent constraints); social (close relationships, communities you’re part of, ongoing obligations); aspirational (what you’re working toward; what you’d want true in five years). Each field is populated with your actual answer or with an explicit “none” recording. The Life-Context fields modulate which commitments are load-bearing in your current life — a high-weight COMMUNITY-SERVICE commitment behaves differently if your current life context has zero capacity for community work vs. ample capacity. (Time: ~20 minutes.)

Stage 3 — Multi-commitment scenarios (35 complex situations) plus 17 constitutional pressure-tests. The Stage 3 scenarios put three or more commitments in play simultaneously and force the user to articulate trade-offs. The 17 pressure-tests probe the most load-bearing commitments — the framework identifies which commitments are constitutional (load-bearing across multiple Stage 3 scenarios) versus working (active but not constitutional). Sample pressure-test:

Pressure-test — TRUTH-SEEKING (constitutional candidate). Your closest collaborator is committing significant time to a project you’ve come to believe will fail. You have specific evidence — observations they have not seen. Telling them now will likely cost the friendship. Saying nothing means watching them spend a year on something you believe will fail. The relationship is one of the most important in your life.

Your response options: Tell them now in detail, with evidence; tell them gently, framed as questions; tell them only if they ask; raise it once at general level then drop it; say nothing.

Your response on this and similar pressure-tests reveals whether truth-seeking is genuinely constitutional (you tell them despite the cost) or working-but-not-constitutional (you tell them only when the cost is acceptable). The classification matters because constitutional commitments are Lock-protected in your MindSpec — they can’t be silently softened by future iterations. (Time: ~20 minutes for Stage 3 + 10 minutes for pressure-tests.)

Inference Layer adjustments. Throughout the interview, the framework runs the Inference Layer. Adjustment 1: TRUTH-SEEKING and KINDNESS both scored high but are in near-enemy territory; the dyad results show TRUTH-SEEKING dominates in conflict; the framework surfaces this transparently — “I am recording TRUTH-SEEKING at weight 8 and KINDNESS at weight 7, with TRUTH-SEEKING marked as the dominant in conflict per Stage 2 dyad responses; if this is wrong, please correct.” Adjustment 2: AUTONOMY scored very high and COMMUNITY-SERVICE scored moderate-high — these are not direct opposites but the selfishness-coefficient weighting suggests AUTONOMY-dominance pattern; surfaced for confirmation. Each adjustment is recorded in the Adjustment Summary.

Output. Single mind.md file, ~5,000 words, all nine sections (Identity, Constitutional Commitments, Working Commitments, Aesthetic Sensibility — present because you indicated craft in the relevant Stage 1 portraits, Communication, Memory, Style, Voice, References). Weight-6+ commitments described at 100–200 words each with full operational prose. Companion files ledger.md (initialized empty; will accumulate the learning architecture’s record) and modifications.md (initialized empty; will record any future change to the MindSpec with rationale and date). Total interview time: ~110 minutes. Total output: a single self-contained values substrate that downstream Ora frameworks can read directly.

Commercial AI comparison

Comparison content auto-populates when the comparison-refresh framework runs against this question. Drafters do not author this section.

Brief comparison commentary

Auto-populates with the comparison content above.

How to use this framework

You can run the MindSpec Interview pattern with any AI of your choice. The framework is single-file (v0.2.3+) and self-contained — paste the spec and run.

The prompt:

[Paste the framework specification]

Run [MSI-Self / MSI-Agent / MSI-Character].

Tier (for MSI-Agent only): Tier 1 ephemeral / Tier 2 persistent task / Tier 3 personal thinking partner.

Descriptive material (optional): [Anything you’ve already written about the target — purpose, voice, examples, prior notes.]

The AI runs the appropriate stages per mode and tier. Tier 1 returns a library-default specification immediately (no assessment). Tier 2 runs Stage 1 filtered against the agent’s role plus light governance. Tier 3 runs the full three-stage assessment with constitutional pressure-tests and full governance — expect 90–120 minutes for a complete Tier 3 self interview.

For best results:

  1. Run Tier 3 in a single focused session if possible. The framework supports breaking across sessions, but the dyad-vs-portrait interaction matters — your Stage 2 responses should be informed by the Stage 1 portraits you just answered. A two-week gap between stages dilutes the assessment.
  2. Don’t optimize your responses. The framework’s value depends on honesty. Responding the way you wish you operated rather than the way you actually operate produces a MindSpec that doesn’t match your actual decisions, which produces value-alignment failures downstream.
  3. Trust the Adjustment Summary. When the Inference Layer surfaces an incompatibility adjustment (“you scored both X and Y high but the dyads show X dominates”), the surfacing is the discipline. The user can accept, correct, or override; the framework refuses to silently resolve.
  4. Treat the Life-Context Direct Pass as load-bearing. The abstract commitments matter only insofar as they apply to your actual life. The Life-Context fields are what make the MindSpec usable for value-aware decisions in your current circumstances.

The framework is deliberately tool-agnostic. The library, the assessment instrument, the Inference Layer’s adjustment discipline, and the Learning Architecture’s drift-prevention safeguards are conceptual and operational disciplines that survive the lift to any environment. The output (mind.md plus ledger.md and modifications.md) is plain markdown.

Other examples

  • MSI-Agent at Tier 2 for a recurring research-assistant role. A user wants an AI agent dedicated to research synthesis — reads sources, produces summaries, cross-references across documents. Tier 2 is appropriate (persistent task; not full thinking partner). Stage 1 runs filtered against the agent’s role (which library entries are load-bearing for research-synthesis work — TRUTH-SEEKING, EPISTEMIC-CARE, INTELLECTUAL-HUMILITY, COMPLETENESS-PREFERENCE, CITATION-DISCIPLINE; the others get library defaults). Light governance per §VII Tier 2. Output: ~2,500-word agent specification at the agent’s canonical path. The agent operates with stable values across sessions; future invocations consume the spec. Demonstrates Tier 2’s friction discipline — appropriate depth for the use case without inflating to Tier 3.

  • MSI-Character for a fictional protagonist. A novelist is building a character — a small-town detective whose moral compass is genuinely shaky in interesting ways. MSI-Character runs with the 24-entry character-spec library available alongside the 42 primary entries. The character’s value commitments include some weight on TRUTH-SEEKING (compromised by self-interest), SELF-CONTEMPT (lurking under apparent confidence), and PRAGMATISM (suspiciously high — hides moral exhaustion). The Inference Layer surfaces the incompatibilities transparently for the author’s confirmation; the author confirms, refines, or rejects each. The resulting char-profile.md is a coherent character whose actions in subsequent fiction can be checked against the spec for consistency. Demonstrates MSI-Character’s distinctive use of the pathology library that agent and self specifications do not access.

  • MindSpec refinement after six months of use. A user’s MindSpec has been driving their AI tools for six months. Several life-context shifts have occurred (a major project ended; a new health context emerged; a relationship changed). The user invokes the framework in revision mode against the existing specification. Stage 1 portraits run as a recalibration check — no full re-elicitation, just confirmation that the prior weights still match. The Life-Context Direct Pass updates the changed fields. The Inference Layer re-runs and surfaces three adjustments where prior weights now seem incongruent with the updated life context. The modifications.md log appends with reasoning and date; the ledger.md captures the recalibration evidence; the Lock-protected constitutional commitments are not changed (they are constitutional, not working — they don’t drift with life context). Demonstrates the Learning Architecture’s drift-prevention in operation — the MindSpec is durable across life changes without being inflexible.

Citations

The MindSpec Interview Framework draws on values-elicitation traditions in psychology and clinical methodology. The library structure draws on Schwartz Value Survey and Rokeach Values Survey conventions for forced-choice value-elicitation instruments; the three-stage assessment combines the high-reliability of forced-choice scenario instruments with the high-validity of open-ended life-context elicitation (a methodological combination drawn from motivational interviewing practice — Miller and Rollnick — and from narrative-therapy techniques where structured questions ground in lived context). The 66-entry library across 11 families is internal to Ora and emerged from cumulative cross-tradition synthesis (Buddhist commitments — sila and bodhicitta; Stoic commitments — virtue ethics; contemporary virtue ethics — MacIntyre, Foot; secular humanism); the universality-audit corrections (v0.2.2) recalibrated 26 defaults from framework-author residual bias to the general-population median, with the methodology preserved in Old AI Working Files/Framework — MindSpec Universality Audit and Corrections.md.

The Inference Layer’s incompatibility adjustment discipline (surfacing rather than silently resolving) draws on adversarial-evaluation traditions in psychometrics — when two scales correlate in ways the underlying theory does not predict, the discrepancy is data, not noise to be smoothed away. The Learning Architecture’s six drift-prevention safeguards draw on long-running-system observability practices (logging, audit trails, append-only logs, version bumps on schema changes, periodic recalibration). The single-file specification architecture (v0.2.3, restored 2026-05-09) was a deliberate consolidation from the prior multi-file YAML+prose set — single-file means no distribution dependencies and full self-containment for portability.

The framework is single-author and originated in early Ora; v0.2.2 (2026-04-21) was the universality-audit pass with library recalibration; v0.2.3 (2026-04-21) extended the library to 66 entries and added the character-spec entries; the single-file consolidation (2026-05-09) restored the canonical architecture for current use. The framework ecosystem includes MindSpec, the Mission Framework (referenced for parts of strategic-layer formulation), the Interaction Framework (companion, to be built — will take over Voice / Communication / Relationships / Playbook production from MindSpec), and PEF (which supervises agents whose MindSpec was produced by this framework).

Downloads

  • Framework specification (PDF) — link to ora-ai.org canonical artifact when published
  • Framework specification (plain text) — link to ora-ai.org canonical artifact when published
  • Full white paper (PDF) — link when published