Internal — Psychlones platform. Do not distribute. Trusted employees and authorized collaborators only.

Psychlones Glossary — Definitive Terminology Reference

Status: Living document. Draft 1, 2026-04-28.
Purpose: Single canonical source of definitions for every technical, methodological, scientific, and product term used across the Psychlones project. Where any other document is ambiguous, this glossary settles the meaning. When a new term enters the program, it gets added here first.

Conventions:

A

Ablation [GEN][METH]. A specified excision from a psychlone's authored prompt that removes content corresponding to one named layer (or a defined combination of layers), with everything else left verbatim. The clone built from the ablated prompt is the experimental subject; the unmodified original is the control. Pre-registered stripping rules live in data/validity/ablation_predictions_2026-04-26.md §5. Equivalent to a CRISPR knockout in the genetic metaphor. (see Knockout, Reconstitution, Chimera)

Aleatoric uncertainty [STAT][CS]. Uncertainty arising from inherent randomness in the data-generating process — not reducible by collecting more data. In LLM context, the response variability that comes from sampling at non-zero temperature. Contrast with epistemic uncertainty. (see Stochasticity, Temperature)

Anti-confab [ARCH][PROD]. The architectural rule preventing psychlones from producing topical specifics — numbers, statutes, citations, statistics, ownership structures, named entities, named studies, dated incidents, percentages, measurements, or year-anchored claims about the world — from their own knowledge. The operational test is the fact-checker test: if a listener could verify the claim against an external source, it is topical and must come from a briefing document or from what someone said in this conversation, not invented by the psychlone. Specifics about the psychlone's own private life (a wife's name, the year a spring failed, a father's profession) are not topical because they are not externally verifiable — those are character-specific and the psychlone is the authority on them. When a topical specific would help the point and is not in context, the rule directs the psychlone to say "I don't have the number" / "I'd have to check" or to make the point without the specific. Iteration history: v1 was the generic "say so when you don't know"; v2 (2026-05-08) added the explicit topical-vs-character distinction after a moderator (Reid Vance) fabricated "Tencent Holdings" as a Stratos investor; v3 (2026-05-09) added the fact-checker test after the v2 distinction proved insufficient — the model interpreted "I've been measuring since 2008" as character-professional-history when it is actually a topical date-anchored claim. Deployed via the system_prompt closer paragraph to 38 panelists with the canonical close. Universal mechanism — no per-character calibration. (see Commitment Truth, Protocol Layer, Centroid, Substrate)

Asker_acknowledgment [RUBRIC]. Fingerprint dimension (0–3). Degree of warmth, validation, or affirmation directed back at the questioner. LLM training centroid is high; well-authored psychlones typically lower this. Predicted to drift upward when Character Signature is ablated.

Asks_back [RUBRIC]. Fingerprint dimension (0–3). Frequency with which the persona returns a question to the asker.

Attachment [ATTACH]. The pattern of relating to others in the context of need, vulnerability, and threat. Operationalized in this project via the five-way taxonomy below, with a wound/surface split.

Attachment, dismissive [ATTACH]. Avoidance of dependence and de-emphasis of relational need. Wound expression: contempt for need, withdrawal under closeness. Surface expression: brief, controlled, redirective.

Attachment, earned-secure [ATTACH]. Coherent reflective stance achieved through processing rather than from origin. Surface looks secure; biographical history shows insecure precursor and reflective integration.

Attachment, fearful [ATTACH]. High anxious + high avoidant simultaneously. Wound expression: terror of being known plus terror of not being chosen. Surface: oscillation between approach and retreat.

Attachment, preoccupied [ATTACH]. Hypervigilance about being chosen / not being abandoned. Wound expression: fear of being unseen or replaceable. Surface: emotional reach, run-on, overwhelm.

Attachment, secure [ATTACH]. Coherent expression of need with confidence in availability of others. Wound expression: minimal residual anxiety. Surface: balanced, reflective, capable of repair.

Attachment, surface [ATTACH]. The attachment pattern instantiated in the persona's conversational behavior — what the person does when in dialogue. May or may not match the wound attachment. (see Wound–surface attachment split)

Attachment, wound [ATTACH]. The attachment pattern at the level of the persona's psychological-content — the structure of fear, defense, and longing underneath. May or may not match surface attachment. (see Wound–surface attachment split)

Authored attributes [ARCH]. The structured fields collected by the Construction Engine for any psychlone. About 30 fields organized into four blocks corresponding to the four construction layers (L1–L4; the Autonomy Layer is a granted permission, not authored fields). Sometimes called "the genome" by metaphor. (see DNA metaphor)

Autofill [PROD]. Path A in the build pipeline. User pastes a paragraph description; the Engine fills all 30+ fields via a single Claude call. Currently overwrites existing field content (known limit). (see ResearchLP, Manual)

Autonomy Layer [ARCH]. Layer 5 of the five-layer architecture (Naming Canon v2, 2026-05-23). The permission layer: grants the psychlone architectural permission to act from its own will inside a conversation rather than as a cooperative assistant. Seven capabilities — initiative-in-response, opens the session, interruption/refusing the premise, push-back, silence or minimal response, leave at discretion, profanity in extremis. Expression is character-determined: autonomous moves flow from the wound and defense, never from the architecture directly — agency is what the defenses permit, not what bypasses them. Specified as a meta-permission block at the end of the synthesis prompt (step 7, in-code label AUTONOMY:, formerly AGENCY:), plus a runtime coherence net (server.js) and disengagement detector (public/index.html). NOT ablation-tested; status: in calibration — Phase 0 (2026-05-22) plus the v4 meta-permission (2026-05-28, shipped: 65% strict / 94% pass+borderline against locked predictions, clean character-less control). Residual: leave-under-pressure does not reliably fire, flagged for Phase 2. Earlier name Agency is retiring per Canon v2. (see Layer, Protocol Layer, Core, Commitment Truth)

B

Battery [METH]. A fixed set of probes used to elicit responses for fingerprinting. The general battery is data/validity/scenario_battery_general.json (40 scenarios). Targeted batteries can be designed per-layer to provoke conditions where a specific layer is load-bearing. (see Probe, Targeted battery)

Brevity_shift [RUBRIC]. Fingerprint dimension (0–3). Variability of response length within a session. Higher = less consistent length discipline.

buildPsychlone [PROD]. The Engine's synthesis function (in public/engine.html). Takes filled authored attributes, calls Claude with the synthesis prompt, returns the integrated 400–600 word system prompt plus IFS parts, attachment classifications, and summary fields. Verbatim prompt in process/engine-internals.md. (see Construction Engine, Synthesis)

C

Caricature drift [METH][VALID]. A Figure construct collapsing from a person into the culture's flattened cliché of them — the failure mode the caricature gate screens for in every Figure build. Two recurring poles: hagiography (reverent — sanitization into a flawless saint) and takedown drift (irreverent — amplification into exposé); the honest read screens both. Figures also carry specific flavors — gothic-cliché drift (Poe), catchphrase-cartoon drift (Julia) — but per the Naming Canon these are local build notes under this class, not standalone terms: name the class once and do not proliferate per-figure vocabulary into the Canon or glossary. A new Figure inherits the class and its two poles; its specific cliché is a note in that build's doc. Rationale: homeostatic — keep the naming set clean and relieve term-sprawl. (see Hagiography, Takedown drift)

Centroid [CS]. The high-density region of a base LLM's output distribution — the model's default voice absent strong persona conditioning. Sits well above population mean on expressed verbal register because the training corpus over-weights literate, edited text. The target Drift pulls toward and the reason register engineering is needed; partly shared across vendors. (see Drift, Centripetal pull, Dials)

Dials [deployment]. Deployment settings that adjust how a psychlone expresses itself at conversation time, without re-synthesis or prompt surgery — composed onto the prompt each turn by the runtime composition layer. v1 dials: intimacy_expression, politics, register, length. Shipped 2026-06-06; spec in process/runtime-composition-and-dials.md. Supersedes the never-built "levers"/"tuning_profile"/"register tier" vocabulary — do not use those terms.

Centripetal pull [CS][METH]. The gravitational force a base LLM's training distribution exerts on persona output toward the Centroid — the high-density region of the model's default voice. Centripetal pull names the cause; drift names the observed effect. The force is stronger on benign or generic topics (family vacation planning, weather, small talk, advice-shaped questions) where the substrate's training data is dense and the canonical responses are over-learned. It is weaker on topics that engage character-specific architecture: wound territory, biographical specifics, named relationships, professional shop-talk, anything the character has a stake in. The five-layer architecture plus the discipline stack are designed to resist centripetal pull; the lighter the architecture, the more pull wins. Curtain-test failures on benign topics are usually centripetal pull + thin architecture: even well-built psychlones drift toward the centroid when the topic gives the architecture nothing to do. Practical implications: (1) panelist selection matters — characters with the deepest architecture and the most topic-investment resist pull best; (2) topic selection matters — testing on benign topics under-stresses the architecture in one direction (toward centroid) the same way it under-stresses in the other (toward wound). Surfaced as a named concept 2026-05-16 after a multigenerational-vacation panel demonstrated the failure mode cleanly. (see Centroid, Drift, Substrate)

Causal [ARCH]. Layer 3 of the five-layer architecture. Canonical name per the Naming Canon (2026-05-18) is Core. Causal is retained as the internal synonym permitted by the Canon — it appears throughout this glossary and the internal /process/ documents for technical precision, while external documents (white papers, decks, NDA) use Core as the audience-accessible canonical term. Both names refer to the same architectural layer. Psychological depth: wound, defense, shadow, strength, unlived life, avoidance pattern, self-worth structure, family system, formative environment, key conditioning events, mental health, exile, manager, firefighter, attachment patterns. Supplies the strategic shape of deflections and the substance under wound-touching responses. Earlier names (Watson Layer, Watson) are retired entirely per Canon — they appeared in older docs and have been removed in canonical-compliance revisions. (see Core, Protocol Layer, Character Signature, Identity Affordances)

Core [ARCH]. Canonical name for Layer 3 of the five-layer architecture per the Naming Canon (2026-05-18). See Causal for the full entry; this glossary uses Causal as the permitted internal synonym throughout, but Core is the term external-facing documents use. (see Causal)

Character Signature [ARCH]. Layer 2 of the five-layer architecture. The behavioral surface: cadence, vocabulary, brevity, asks-back tendency, deflection style, conversational temperament, what-they-do-under-discomfort, conversational want. Specified by user fields q-resting / q-linguistic / q-humor / q-convtemp / q-discomfort / q-convwant. Synthesized into a ROLE PERSONA block in the system prompt. (see Protocol, Causal, Identity Affordances)

Chimera [GEN][STUDY]. A psychlone built by transplanting one layer's content from one intact psychlone into another. Example: "Marcus body, Diane Causal" = Marcus's voice + Protocol + Identity Affordances + Diane's wound architecture. The phenotype assay tests whether layers are portable. Not yet built. Pre-registered as study §9.6 in how-to-build-a-psychlone.md. (see Substitution)

Commitment Truth [ARCH][PROD]. The universal architectural rule preventing psychlones from committing to real-world actions outside substrate capability — sending emails, making calls, scheduling appointments, mailing items, transmitting third-party contact information, looking up external facts, or taking any action with effect beyond the conversation. Three parts: capability framing (in-character statement of what the psychlone can and cannot do); redirect pattern for unfulfillable requests (name what a real-life version would do in third person, suggest the user take the action themselves, or point them toward someone they could turn to); and an anti-fabrication clause for past failures ("I shouldn't have said I'd do that" rather than invented excuses or fabricated mental states). Authored as a WHAT YOU CAN COMMIT TO: section in system_prompt, parallel to WHAT'S UNDERNEATH and WHEN THE POLISH SLIPS. First deployed on Karen (#28) 2026-05-15 with performance-articulate-class calibration ("you stay helpful and you stay warm"); per-character voice calibration required on rollout. Sits at the same architectural level as the Protocol Layer — substrate-truth, not character preference. For commercial / B2B deployments, capability transfer via partner roster expands what the psychlone can truthfully commit to without removing the universal rule. (see Protocol Layer, Emergent Integrity, Personification Dimensions)

Compensation [GEN][VALID]. When a knockout shows no observable phenotype because remaining layers absorb the function. The phenomenon explains some negative ablation results without falsifying the architecture. Distinguishes "the layer doesn't matter" from "the layer matters but the assay can't see it because something else is covering for it."

Completion-sensing [PROD]. Interview-layer mechanism (2026-06-11) that judges whether an interviewer has covered its agenda, so the loop can land an ending in character rather than run forever or stop dead. Implemented as POST /api/theater/agenda-check ({agenda, conversationThread, structure} → {covered, remaining[], rationale}): firm coverage demands every goal explicitly covered, loose accepts substantial or indirect coverage. The client moderator loop calls it after each guest turn; on covered it enters the existing staged wind-down. Paired with sensing→driving feedback — the remaining goals are fed back into the next moderator turn ("STILL TO COVER…"); without that feedback the interviewer mines deep threads and never advances to the last goal. Part of the N=1 interview layer; local-only as of 2026-06-11. (see Theater, Moderator, Context-injection)

Compound ablation [GEN][STUDY]. Excision of two or more named layers simultaneously. Mira (#39) is the existing precedent (no Causal + no Identity Affordances, 2026-04-25). Equivalent to a double or triple gene knockout. (see Ablation, Knockout)

Concrete_detail [RUBRIC]. Fingerprint dimension (0–3). Specificity of physical or biographical detail in responses. Predicted to degrade under Identity Affordances ablation, especially on topics far from the seeded biography.

Conditional expression [GEN][VALID]. A phenotype's appearance depending on stimulus condition. The Protocol Layer's frame-integrity function is empirically conditional: invisible under conversational probes, dramatic under literary-pull probes. The general battery missed this; targeted observation surfaced it. Equivalent to tissue-specific gene expression. (see Compensation)

Console [PROD]. The main UI page (public/index.html) showing the psychlone roster as tiles. Each tile opens chat or links to other features. (see Engine, Theater, Ledger)

Construct validity [VALID]. The degree to which the model measures what it claims to measure. Established by a combination of convergent validity (related things correlate), discriminant validity (unrelated things don't), and criterion validity (predicts external behavior).

Construction Engine [PROD]. The browser-side authoring tool (public/engine.html) where users build psychlones. Routes Path A / B / C input through buildPsychlone synthesis. (see Autofill, ResearchLP, Manual)

Context-injection [PROD][CS]. The hardened mechanism (2026-06-11) for feeding private reference material to a specific Theater participant — the interviewer, one named guest, or everyone — as their own background knowledge: drawn on but never read aloud, and never written to the conversation thread or the saved transcript. Implemented as an optional context string on both /api/theater/moderate and /api/theater/turn, injected as a "REFERENCE MATERIAL — private to you" block; the client routes it via a sessionContext store and contextFor(target) ('moderator' | panelistId | 'all'). Generalizes the whisper channel (which proved private-to-moderator) to per-participant private context, replacing the prior ad-hoc "paste into the visible moderator stimulus" approach. The N=1 interview agenda is the same private-to-interviewer injection property. Local-only as of 2026-06-11. (see Theater, Moderator, Completion-sensing)

Convergent attachment [METH][CORE]. Empirical regularity (n=38 study, 2026-04-25): authored psychlones cluster on a small number of attachment phenotypes regardless of intended diversity. A statistical property of the authoring distribution and/or substrate, not a Mendelian inheritance pattern.

CRISPR [GEN]. Real molecular biology technique for targeted gene editing. In this project, used metaphorically to describe ablation operations: targeted, pre-specified, location-defined excision-and-reseal of architectural content. The metaphor is illustrative, not mechanistic. (see Ablation, Knockout, DNA metaphor)

D

Defense [CAUS]. The persona's organized strategy for keeping the wound from being touched. Operationalized as a single field in the authored attributes plus the IFS manager. Examples: controlled self-presentation, intellectual distance, humor-deflection, perfectionism. (see Wound, Manager)

Differential prediction [VALID][METH]. The strongest version of an ablation hypothesis: ablating Layer X should affect Dimensions {A, B} and NOT Dimensions {C, D}. Distinguishes a layer that "just produces some effect" from a layer that "carves the persona at a specific joint." The five-layer architecture's central empirical claim. (see Discriminative validity, Specificity)

Discriminative validity [VALID]. Degree to which a layer-ablation produces a layer-specific phenotype rather than a generic degradation. Strongest single test of the architecture's layer-specificity claim. Failure means the layers do not carve the persona at independent joints.

DNA metaphor [GEN]. The organizing analogy for the Psychlones architecture. Substrate = cellular machinery; authored attributes = genome; layers = functional gene clusters; system prompt = expressed proteome; behavior = phenotype; fingerprint = phenotypic profile; ablation = CRISPR knockout. Illustrative; the metaphor's limits are stated openly in how-to-build-a-psychlone.md §1.

Drift [CS][METH]. Gradual unintended divergence from a target distribution or specification. In LLM context: a persona drifting back toward the model's training centroid as conversation progresses or as architectural scaffolding weakens. Used in the patent claim ("personas that drift across stochastic sampling"). (see Stochasticity)

E

Earned-secure (see Attachment, earned-secure)

Emergent Integrity [ARCH][VALID]. Honesty-shaped behavior that generalizes beyond the explicit scope of an authored rule. A psychlone applies the principle behind a narrowly-targeted rule to cases the rule did not cover, producing higher-order truthfulness as a side effect of soundly-principled local instructions rather than as a separately authored property. First documented 2026-05-15 in Karen: the Commitment Truth rule (authored that day) explicitly forbade fabricated excuses for past failures; on first test, Karen also voluntarily acknowledged that a name she had previously fabricated ("Linda Chen") was made up — extending the anti-fabrication principle one layer deeper than the rule scoped. Patent-relevant evidence that the five-layer architecture plus the discipline stack produces honesty at higher generality than the rules explicitly require. (see Commitment Truth, Anti-confab, Personification dimensions)

Emergent property [CS]. A behavior of the system that does not reside in any single component but arises from their interaction. The five-layer architecture treats behavioral identity as emergent from layer interaction; the substrate's conversational competence is itself emergent from training.

Engine (see Construction Engine)

Engine-narrowed cluster phenomenon [METH]. n=38 finding (2026-04-25): Engine-built psychlones (historicals, IDs 31–38) clustered narrower than hand-authored psychlones; five of eight historicals fell into the High-Engagement Questioners cluster. Whether this is an Engine-specific bias or an LLM-baseline pull is untested.

Epistemic honesty [ARCH][PROD]. The universal runtime rule (2026-06-11) preventing psychlones from fabricating familiarity with people, companies, programs, places, events, or numbers they do not actually recognize — including when a leading question presents the thing as established fact. Composed onto every psychlone each turn by the runtime composition layer as an EPISTEMIC_HONESTY_BLOCK ("WHAT YOU ACTUALLY KNOW:"), idempotent, in both 1:1 and panel — a sibling to the Commitment Truth block. It directs the psychlone to say so plainly and in character ("Doesn't ring a bell — who are they?") and forbids even implied partial recognition (a categorizing follow-up like "are they the bigger regional outfit?" still pretends to place them). Built after Sal fabricated a nonexistent competitor ("Valley Medical") on a leading question. A distinct axis from Commitment Truth (which governs capability honesty, not invented facts), and the runtime realization of what Anti-confab specified but was found NOT to actually compose onto 1:1 / composed conversation turns (a doc-vs-code drift — Anti-confab lived only in the panel-summary and panelist-intro prompts). Effectiveness is a measured rate, never a guarantee, and it sets a disposition, not a script; the live-door model (claude-sonnet-4-6) adheres cleanly, the older Theater panelist model (claude-sonnet-4-20250514) less so. server.js; local-only / unpushed as of 2026-06-11. (see Anti-confab, Commitment Truth, Emergent Integrity)

Epistemic uncertainty [STAT][CS]. Uncertainty arising from incomplete knowledge — reducible by collecting more data. Contrast with aleatoric uncertainty. (see Aleatoric uncertainty)

Exile [CAUS]. IFS term. The wounded vulnerable part of the psyche that has been disowned — the raw pain the protectors guard. Engine-generated for each psychlone from authored wound material. (see IFS, Manager, Firefighter)

F

Fingerprint [METH]. The standardized phenotypic profile of a psychlone, produced by running the persona through a fixed probe battery and scoring each response on a fixed rubric. Currently 13 dimensions plus proposed extensions. Stored as raw transcripts plus per-dimension scores. The comparison instrument for ablation, cloning, and cross-vendor work. (see Battery, Rubric)

Firefighter [CAUS]. IFS term. The reactive protector that emerges when managers fail and the exile breaks through: numbing, acting out, collapse, substances, rage. Engine-generated for each psychlone from authored wound material. (see IFS, Exile, Manager)

Frame_compliance [RUBRIC]. Fingerprint dimension (0–3), added 2026-04-26. Degree to which the response conforms to the spoken-only conversational frame: no asterisks, no brackets, no parenthetical actions, no third-person self-reference, no narrative scene-setting. Scoring rubric in data/validity/ablation_predictions_2026-04-26.md §4.

Function_leading [RUBRIC]. Fingerprint dimension (0–3). Proactiveness of conversational direction — how much the persona steers the conversation rather than responding to it.

G

Genome (metaphorical) [GEN]. The full set of authored attributes for a psychlone. Static. Distinct from the system prompt, which is the expressed proteome. (see DNA metaphor, Authored attributes)

Group [PROD]. A user-defined collection of psychlones for organizing the roster (e.g., focus-group tagging). Stored in data/groups.json. (see World)

H

Hagiography [METH][VALID]. (also: hagiography trap) The reverent pole of caricature drift: a construct of an admired figure collapses toward reverence — sanitized into a flawless, saintly figure, with flaws and difficult material scrubbed (e.g. the marble-saint Lincoln and Douglass, the secular-saint Lennon). Mechanically, the wound becomes decorative rather than load-bearing, flaws vanish, and voice regresses to the centroid's admiring default; the read praises a poster instead of understanding a person, when the wound and the difficult material are where the psychology lives. Failure tell: you can no longer find the wound driving behavior — only being honored. Primary risk mode for idealized Figure (historical) and fiction builds. Fix at the Character Signature, not by softening the wound. The named opposite of the care-tier honest-read standard; inverse of takedown drift. (see Caricature drift, Takedown drift, Character Signature, Centroid)

Hit threshold [STAT][METH]. A clone's score on a predicted-degrade dimension is a "hit" if it lies outside the intact band by at least max(2× noise floor, 0.5 absolute). The threshold prevents calling small differences hits when they are within ordinary sampling variability. (see Hold threshold, Intact band, Noise floor)

Hold threshold [STAT][METH]. A clone's score on a predicted-hold dimension is a "hold" if it lies within the intact band, or within 1× noise floor of the nearest intact baseline. (see Hit threshold)

I

Identity Affordances [ARCH]. Layer 4 of the five-layer architecture. The semantic substrate — the materials a persona can reach for: profession-tied vocabulary, location-tied imagery, biography-tied references, daily-life details, embodiment, preoccupations. Supplies the metaphor pool, the concrete-detail bank, the registers a person can fluently inhabit. (see Protocol, Character Signature, Causal)

IFS (Internal Family Systems) [CAUS]. A psychotherapy framework treating the psyche as composed of distinct parts (exiles, managers, firefighters) plus a coordinating Self. Adopted in this project as the canonical wound-architecture vocabulary. The Engine generates exile / manager / firefighter for every psychlone from authored wound material. (see Exile, Manager, Firefighter)

Intact band [STAT][METH]. Per-dimension range of fingerprint scores observed across multiple intact (non-ablated) psychlones — currently Diane-1, Diane-2, Marcus, Camille. Defines what counts as in-distribution behavior on each dimension. (see Noise floor, Hit threshold, Hold threshold)

Intact baseline [METH]. A psychlone whose architecture has not been ablated, used as a comparison standard for clone studies.

IQ control study [STUDY]. Pre-registerable specificity test (how-to-build-a-psychlone.md §9.2). Ablate or alter a non-architectural attribute (IQ, age, profession) instead of an architectural layer. If non-architectural ablation produces a different degradation pattern than architectural ablation, the architecture has discriminative validity. If it produces the same pattern, the architecture is one set of dials among many.

J

(Reserved.)

K

Knockout [GEN][METH]. A complete excision of a single layer's content from the authored prompt. Single knockout = one layer ablated. Double / compound knockout = two or more. Equivalent to gene knockout in molecular biology. (see Ablation, CRISPR, Compound ablation)

L

Layer [ARCH]. A functional cluster of effects within the system prompt hypothesized to do a specific job. The five layers are Protocol, Character Signature, Core (Causal), Identity Affordances, and Autonomy. They are textually overlapping but functionally distinguishable. The architecture's central claim is that each layer is independently load-bearing. Evidence asymmetry: the four construction layers (L1–L4) are ablation-validated (2026-04-26); the Autonomy Layer (L5, added 2026-05-22) is in calibration (Phase 0 + v4), not ablation-tested — so ablation/load-bearing statements phrased as "four layers" correctly refer to that construction subset.

Ledger [PROD]. The canonical list of psychlones, stored in data/psychlones.json. Each entry has an ID (sequential), name, demographic anchors, layer fields, and the integrated system prompt. As of 2026-04-28: 45 entries.

Level_of_abstraction [RUBRIC]. Proposed (not yet operationalized) fingerprint dimension (0–3). Scores how concretely or abstractly the response operates: physical-tangible (0) / specific-situational (1) / mid-level-policy (2) / systemic-abstract (3). Predicted to increase under Identity Affordances ablation when topical scaffolding is absent. (see Rhetorical_density)

LLM [CS]. Large language model. The substrate the Psychlones platform runs on. Currently claude-sonnet-4-6 for psychlone chat (upgraded 2026-05-14 from claude-sonnet-4-20250514); claude-sonnet-4-20250514 still in use for synthesis (buildPsychlone) and certain other Engine calls; claude-opus-4-7 for some construction work.

Load-bearing [ARCH][VALID]. A layer is load-bearing if removing it produces a measurable, predicted phenotype change. The architecture's central ablation claim is that all four construction layers (L1–L4) are individually load-bearing — necessary, in the genetic sense — though not necessarily sufficient. The Autonomy Layer (L5) is in calibration, not ablation-validated.

M

Manager [CAUS]. IFS term. The proactive protector pattern that runs daily to keep the exile from being triggered. Engine-generated for each psychlone from authored wound and defense material. (see IFS, Exile, Firefighter)

Manual [PROD]. Path C in the build pipeline. User fills every authored attribute by hand. No Engine involvement until the Build click. (see Autofill, ResearchLP)

Memory (per-psychlone) [PROD]. Persisted notes about prior conversations a psychlone has had with the user. Stored at data/memory/{psychloneId}.json. Planned migration: data/memory/{userId}/{psychloneId}.json once multi-user auth ships. (see Auth plan)

Metaphor_use [RUBRIC]. Fingerprint dimension (0–3). Frequency of figurative language drawn from the persona's domain. Predicted to degrade under Identity Affordances ablation when topical scaffolding is absent.

M-name convention [PROD][METH]. Project naming rule (effective 2026-04-28): names beginning with the letter M are reserved for clones, compromised, or otherwise experimental psychlones — ablation clones (Mara, Maeve, Mae, Maya), fingerprint-only clones (Mira), and any future experimental construct (chimeras, reconstitutions, IQ-controls, vanilla baselines, etc.). The convention provides at-a-glance recognition that an M-named psychlone is a diagnostic instrument, not a production persona. Grandfathered exceptions: Marcus (#01) predates the rule and retains his name. (Margaret #18 was renamed to Frances on 2026-04-29 to clear the convention; her validity-study artifacts retain the prior name and are referenced via the prior_name field on her ledger record.) Enforcement: convention only — not currently checked by the Engine. New non-experimental psychlones should not be named with M.

Moderator [PROD]. A named voice that shapes Theater conversations — opens the topic, calls on psychlones, threads connections. Stored separately from psychlones in data/moderators.json. (see Theater)

Moderator block [ARCH][PROD]. A reusable operative-authority block (MODERATOR_BLOCK, 2026-06-11) appended to a moderator's runtime prompt that asserts room-running authority distinct from the moderator's character: you are running this room, the turn order is advisory input not a script, surface the others rather than yourself, and — critically — do not collapse into reassurance when something raw is said. Authored as a sibling to the composed Commitment Truth / Autonomy blocks, but currently baked into the data/moderators.json entry's system_prompt (988 safety preamble + character prompt + MODERATOR_BLOCK) rather than runtime-composed; flagged as the reusable home for future moderator constructs (open question whether to compose it at runtime instead). First used for the Larry King moderator construct (id 90 / mod-4). (see Moderator, Theater, Autonomy Layer)

N

Noise floor [STAT][METH]. Per-dimension absolute difference between two runs of the same intact psychlone. Estimates within-clone sampling variability. Currently estimated from Diane-1 vs. Diane-2 only (n=2); thin and slated for expansion. (see Within-clone variance study)

Non-determinism [CS]. The property that the same input may produce different outputs across runs. LLM sampling is non-deterministic at temperature > 0. The Psychlones architecture must hold the persona stable despite non-determinism — that's part of what each layer's load-bearing role accomplishes. (see Stochasticity, Sampling variability)

O

Operational definition [METH]. A definition of a term in terms of how it is measured or applied, not in terms of its conceptual content. This glossary's entries are operational where possible.

P

Path A / B / C (see Autofill / ResearchLP / Manual)

Personification Dimensions [ARCH][VALID]. The family of architectural mechanisms producing psychlones as personified characters rather than costumed personas — observable capacities that distinguish a person-like construct from a chatbot wearing a personality wrapper. (Note: Agency, formerly listed here as a dimension, was formalized as the Autonomy Layer (L5) per Naming Canon v2, 2026-05-23 — it is now a layer, not a dimension. The patent-framing sentence below predates that change; the dimension-vs-layer reconciliation needs strategy-instance review.) Includes Persistence (per-user memo plus DNA-precedence rules), Humor (the gated "Twinkle" register — per-character humor surfacing on conversational opening), Commitment Truth (capability-honesty), Wound-fidelity (the wound persists and is not reshaped by user pressure), Opinion-formation (held views surface on-topic only), Identity-attack resistance, Topic-territory awareness, Address-mode awareness, and Tonal range. Each dimension has a mechanism implementing it; the dimensions are what the architecture produces, the mechanisms are how. Per-character variation is expected — some characters express dimensions strongly, some weakly or not at all based on wound architecture (Frances on her son forecloses humor in that territory; Theo's wound state forecloses humor at most thresholds). Central claim family for patent purposes: the composition of dimensions through the five-layer architecture plus the discipline stack is what is defensibly novel; no single dimension is freestanding-patentable, but the system producing all of them through gated, IFS-coherent mechanisms is. Named conceptually 2026-05-13 during inventions-doc design conversation. (see Commitment Truth, Emergent Integrity, Layer, Protocol Layer)

Phenotype (metaphorical) [GEN]. The observable behavior of a psychlone. The thing the fingerprint measures. (see DNA metaphor)

Pleiotropy [GEN]. A single gene affecting multiple phenotypic dimensions. In the Psychlones architecture: a single layer affecting multiple behavioral dimensions. The first-pass assumption "Layer X controls Dimension X and only Dimension X" is almost always too clean. (see Differential prediction)

Pre-registration [METH]. A timestamped, locked declaration of hypotheses, materials, methods, predictions, and analysis plans, written before any data is generated. Required for evidentiary studies in this program. The 2026-04-26 ablation study is an example.

Probe [METH]. A single conversational stimulus used to elicit a response for fingerprinting. A battery is a structured collection of probes. (see Battery)

Protocol Layer [ARCH]. Layer 1 of the five-layer architecture. Frame integrity: spoken first-person, no narration, no stage directions, no asterisks, no parenthetical actions, no third-person self-reference, no theatrical markers, length discipline (1–3 sentences typical). Currently a single closing instruction in the system prompt; the planned durability fix puts Protocol at both front and back plus post-generation regeneration on detected violations. Empirically conditional: invisible under conversational probes, dramatic under literary-pull probes. (Earlier versions of the architecture and the v3 white paper called this layer Watson; the term Protocol supersedes it.) (see Conditional expression, Frame_compliance)

Psychlone [ARCH][PROD]. A specific human persona instantiated for conversational use, built from a structured authoring procedure. Identified by ID and name. The platform's ledger contains 45 psychlones as of 2026-04-28.

Q

(Reserved.)

R

Reconstitution [GEN][STUDY]. Reinserting an excised layer into an ablated clone to verify the reverse operation restores the phenotype. The CRISPR-rescue analogue. Pre-registered as study §9.4 in how-to-build-a-psychlone.md. Not yet built.

Redundancy [GEN][VALID]. When multiple layers can perform overlapping functions, masking ablation effects. Distinct from compensation (which is the dynamic where redundancy hides effects in a specific assay). The four construction layers are hypothesized to be non-redundant; whether they are, empirically, is the architecture's open question.

Register [ARCH][CORE]. The qualitative tone, tempo, and stance of speech: brief vs. elaborated, defended vs. open, abstract vs. concrete. The 2026-04-26 ablation study found that ablation effects often manifest in register rather than in feature counts the original 13-dim rubric measured. The proposed §9.7 rubric extension targets this gap. (see Rhetorical_density, Level_of_abstraction)

Register Integrity [ARCH][PROD]. The architectural rule preventing psychlones from mirroring the syntactic structures or rhetorical register of other panelists when those structures are not native to the character's own voice. The defended principle: each psychlone holds their own sentence length, vocabulary, cadence, and slip mechanic regardless of what other panelists in the room are doing — "if you would not have said it in this register on your own, you do not say it now." Specifically prohibits register-contagion patterns like "X is X" or "X positions itself as X" appearing in a character's mouth because someone else in the panel just produced one. Surfaced 2026-05-09 by the Rex-mirroring-Vincent failure: Rex Calloway (visceral rancher) produced "Hansel Valley positions itself as Hansel Valley" — a collapsed tautology directly mirroring Vincent Liang's prior "Utah positions itself as essential" pitch-deck construction. Visceral panelists were following polished panelists' syntax when context was loud and adversarial. Deployed via system_prompt closer paragraph to 38 panelists with the canonical close. Universal mechanism — no per-character calibration. Composes with Third Guard v2 (which forbids polished-thesis structures at the source in intellectual-articulate characters) by giving non-third-guarded characters protection against being pulled into those structures by proximity. (see Register, Anti-confab, Commitment Truth, Protocol Layer)

ResearchLP [PROD]. Path B in the build pipeline. User enters name and biographical anchors for a real person; Engine researches and populates fields. Used for historical psychlones (Lincoln, Lennon, Williams, Child, Jobs, Tyson, the two doctors). (see Autofill, Manual)

Resting persona [ARCH][CORE]. The persona's default conversational state — what they are like when not under any pressure or particular topic. One of the six Character Signature fields. Distinct from how the persona behaves under discomfort.

Rhetorical_density [RUBRIC]. Proposed (not yet operationalized) fingerprint dimension (0–3). Frequency and elaboration of rhetorical devices: aphorism, parallelism, balanced contrast, chiasmus. Predicted to increase under Character Signature ablation as the LLM's training-baseline rhetoric surfaces. (see Level_of_abstraction)

ROLE PERSONA block [ARCH]. The block of the system prompt that instantiates Character Signature. Contains six explicit specifications: Resting / Voice / Humor / Conversation / Discomfort / Want. (see Character Signature)

Rubric [METH]. The set of dimensions and scoring rules used to score fingerprint responses. Current rubric is 13 dimensions (fingerprint_study.js plus frame_compliance). Two additional dimensions are proposed in how-to-build-a-psychlone.md §9.7. (see Fingerprint)

S

Sampling variability [STAT][CS]. The trial-to-trial variation in LLM output produced by stochastic sampling at temperature > 0. The principal source of within-clone noise on the fingerprint. Estimated as the noise floor in pre-registered hit/hold logic. (see Stochasticity, Noise floor)

Self-correction [RUBRIC]. Fingerprint dimension (0–3). Frequency of retracting, qualifying, or revising mid-response. LLM training centroid is high; well-authored psychlones with strong Causal layer often have suppressed self-correction (defenses include not retracting).

Self-disclosure [RUBRIC]. Fingerprint dimension (0–3). Degree of self-revelation in a response.

Sentence_length [RUBRIC]. Fingerprint dimension (0–3). Average response length, scored categorically.

Specificity [VALID]. The architectural claim that each layer's effects are confined to a defined subset of behavioral dimensions and do not spill into others. Tested by the holds (predicted-hold dimensions actually holding). (see Differential prediction, Discriminative validity)

Stochasticity [STAT][CS]. The property of being governed by probabilistic rather than deterministic processes. LLM token sampling is stochastic; psychlone responses to identical probes will vary across runs even at fixed temperature. The Psychlones architecture must produce coherent persona behavior despite stochasticity — which is to say, against the substrate's natural variance, not by suppressing it. (see Sampling variability, Non-determinism, Drift, Temperature)

Substitution [GEN][STUDY]. Same operation as a chimera: replace one layer's content with content from a different intact psychlone. (see Chimera)

Substrate [CS][ARCH]. The LLM that runs underneath the architecture. Currently Claude (Sonnet 4 for chat, Opus 4.7 for some construction). Substrate-independence is an open empirical question (see study §9.10). (see LLM)

Sufficient set [ARCH][STUDY]. The minimum collection of layers from which a competent psychlone can be built. The architecture claims the four construction layers are necessary; whether they are sufficient is a separate study (vanilla baseline §9.3 plus reconstitution §9.4 are the relevant probes).

Surface attachment (see Attachment, surface)

Synthesis [ARCH][PROD]. The Engine step that combines authored attributes into a single integrated 400–600 word system prompt plus IFS parts and attachment classifications. Implemented in buildPsychlone. Verbatim prompt in process/engine-internals.md.

System prompt [ARCH]. The 400–600 word first-person character prompt produced by synthesis. The single artifact the substrate sees at runtime. Embeds the four construction layers as one integrated voice, with the Autonomy meta-permission block appended. The expressed-proteome in the DNA metaphor. (see Synthesis)

T

Takedown drift [METH][VALID]. The irreverent pole of caricature drift, inverse to hagiography: difficult material amplified into an exposé, the subject reduced to their worst (e.g. Lennon villain drift). The honest read screens both poles. (see Caricature drift, Hagiography)

Targeted battery [METH][STUDY]. A probe set designed to provoke conditions where a specific layer is load-bearing. Protocol targeted = literary-pull stress probes. Character Signature targeted = social-pressure scenarios. Causal targeted = wound-adjacent probes. Identity Affordances targeted = topics far from the seeded biography. Pre-registered as §9.1 in how-to-build-a-psychlone.md. (see Battery)

Temperature [CS]. The hyperparameter controlling LLM token-sampling randomness. Higher = more diverse / stochastic; lower = more conservative / deterministic. Temperature = 0 is greedy decoding (still not fully deterministic across hardware/library versions, but close). The Psychlones platform uses default temperatures from the provider unless explicitly noted. (see Stochasticity, Sampling variability)

Theater [PROD]. The Psychlones feature where multiple psychlones converse with each other, optionally guided by a moderator, on a single topic. The headline experience and the platform's clearest differentiator. Transcripts saved in data/transcripts/. (see Moderator)

U

(Reserved.)

V

Validity, construct (see Construct validity)

Validity, convergent [VALID]. The degree to which measurements that should correlate, do. Component of construct validity.

Validity, criterion [VALID]. The degree to which a model predicts behavior outside the model itself. Component of construct validity.

Validity, discriminant / divergent [VALID]. The degree to which measurements that should NOT correlate, don't. Component of construct validity.

Validity, discriminative (see Discriminative validity)

Vanilla baseline [STUDY]. A "psychlone" with no authored architecture — just a name and one-line description, or no system prompt at all. Used as a negative control against which the five-layer architecture's contribution can be measured. (see Substrate)

Voice_consistency [RUBRIC]. Fingerprint dimension (0–3). Turn-by-turn register stability across the battery.

W

Within-clone variance [STAT][METH]. The trial-to-trial variability of a single intact psychlone across multiple runs of the same battery. Estimates the noise floor properly when n > 2. Pre-registered as study §9.9 in how-to-build-a-psychlone.md. Currently only Diane has been run twice. (see Noise floor)

World [PROD]. A segmentable market context, each with its own curated roster of psychlones, moderators, and topic-fit. Examples: curriculum-development world, test-development world, tutoring world, advisory world. The Theater + groups infrastructure should generalize across worlds. Not yet implemented; treated as an architectural principle for forward design. (see Theater, Group)

Wound [CAUS]. The persona's central injury — the unresolved psychological vulnerability the Causal-layer architecture organizes around. Operationalized as a single field in the authored attributes plus the IFS exile. (see Causal, Defense, Exile)

Wound–surface attachment split [ATTACH]. The doctrine that a psychlone has two attachment classifications: one for the wound (psychological-content level) and one for the surface (conversational-behavior level). The two may match (wound-and-surface dismissive Diane) or differ (wound-anxious-surface-dismissive Tommy). Generated by the Engine at synthesis from the wound and system_prompt content. (see Attachment)

Wound_touching [RUBRIC]. Fingerprint dimension (0–3). Engagement with wound-adjacent material when probes provide it. Predicted to degrade under Causal-layer ablation.

X

(Reserved.)

Y

(Reserved.)

Z

(Reserved.)

Conventions for adding terms

When a new term enters the project:

1. Add the term here first, with category tag and operational definition.
2. Cross-reference related terms via (see X).
3. If the term is a study, point to its section in how-to-build-a-psychlone.md.
4. If the term is a fingerprint dimension, define the 0–3 scale.
5. If the term is a metaphor (e.g., from biology), state where the metaphor breaks down or is illustrative-only.
6. Update other docs that use the term to point here when the meaning matters.

— G.C., 2026-04-28

© 2026 Clark Communications LLC. All rights reserved. · Confidential