Pattern → Preference Mapping Rubrics
Each attribute defines: (1) observable production patterns, (2) observable reception patterns, (3) statistical indicators where applicable, and (4) validation against pilot data (Sheldon, Penny).
General Principles:
- Production ≠ Preference: A character’s speaking style is only a prior; reception evidence (how they react to others’ styles) is required to confirm.
- Reception signal validity: Only count reactions that are clearly to the communication style, not to plot events. A topic shift driven by the script is not a reception signal. Valid signals: explicit confusion/discomfort with a register, engagement increase/decrease in response to style (not content), verbal meta-commentary (“Participate in the what?”).
- Context-dependency: All attributes can vary across interaction contexts. Preference profiles should be reported per context where evidence permits.
- Similarity-attraction default: When only production data is available, assume the character prefers to receive in the same style they produce. This is the default hypothesis, to be overridden by reception evidence.
- Confidence gate: For each preference attribution, the AI reports a confidence level (High / Medium / Low). High = production + reception evidence agree across multiple scenes. Medium = only production evidence, or reception evidence is ambiguous. Low = insufficient evidence or contradictory signals. Low-confidence attributions must be flagged for human review rather than auto-accepted. This creates a human-in-the-loop quality gate that keeps the pipeline scalable (high-confidence auto-pass) while ensuring accuracy on ambiguous cases.
- Three-layer determination: (1) Statistical indicators are within-show cast-relative rankings — they give direction and ordering (who is more X than whom), not absolute settings. (2) Qualitative production patterns provide absolute anchors — features like academic vocabulary or slang usage map to settings regardless of what show the character is in. (3) Reception evidence is the final arbiter — how others react to the character and how the character reacts to others’ styles confirms or overrides the prior. Statistics rank, patterns anchor, reception confirms.
Dimension 1: Expression Style
1. Tone & Formality
Settings: Casual / Consultative / Formal
| Setting | Production Patterns | Reception Patterns |
|---|---|---|
| Casual | Slang, interjections (wow, oh god), contractions (gonna, wanna), filler words (like, um, well), high informal marker density, sentence fragments, colloquial idioms | Responds warmly to casual/friendly register; confused or alienated by formal/technical language; disengages from overly structured speech |
| Consultative | Complete sentences, moderate technical vocabulary, polite but not stiff, professional but accessible, occasional informal markers in context | Comfortable with both casual and formal input; defaults to a middle register; may gently redirect overly casual or overly formal interlocutors |
| Formal | Academic/technical vocabulary (e.g., “proximal cause,” “contradistinction”), complex multi-clause sentences, minimal slang/interjections, formal markers (indeed, thus, hence, query), precise word choice | Cannot process or actively avoids casual/emotional register; responds positively to structured, precise language; demands logical rigor from others |
Statistical Indicators (cast-relative, directional only):
| Indicator | Casual | Consultative | Formal |
|---|---|---|---|
| Formal marker density (per turn) | Low (cast bottom) | Mid-range | High (cast top) |
| Informal marker density (per turn) | High (cast top) | Mid-range | Low (cast bottom) |
| F/I ratio | Near 0 | Between extremes | High relative to cast |
| Avg sentence complexity | Low-medium (fragments OK) | Medium (complete sentences) | High (multi-clause, subordination) |
Reception Signal Validity:
- VALID: Character explicitly expresses confusion at a register (“Participate in the what?”), disengages when register shifts (“What’s happening”), matches/reciprocates a register shift
- VALID: Character responds with increased engagement (longer turns, follow-up questions, warmth) to a specific register
- INVALID: Topic changes that are plot-driven, not style-driven
- INVALID: Disagreement with content (not register) — e.g., Sheldon correcting astrology is content disagreement, not register rejection
Pilot Validation:
| Character | Production | Reception | Setting | Confidence |
|---|---|---|---|---|
| Sheldon | F/I=0.31, formal markers 5x, uses “quantum mechanics,” “Bourne-Oppenheimer approximation,” “proximal cause” | Cannot process Penny’s casual emotional register (“What’s happening”); demands logical restatement from Leonard | Formal (relaxes toward Consultative with close friends in private — context-dependent) | High |
| Penny | F/I=0.00, informal markers 69x, “Holy smokes,” “like one of those,” “wow” | Confused by Sheldon’s formal register (“Participate in the what?”); responds warmly to Leonard’s casual friendliness (“You guys are really sweet”) | Casual | High |
Boundary Notes:
- Consultative is the hardest to pin down — it’s defined partly by the absence of strong casual or formal markers, which makes it a “default” that may need more positive indicators.
- Context-dependent variation: Sheldon’s aggregate F/I=0.31 masks that he is more formal in explanatory/public mode and slightly less formal one-on-one with Leonard. Report per-context where possible.
2. Verbosity
Settings: Terse / Moderate / Detailed
| Setting | Production Patterns | Reception Patterns |
|---|---|---|
| Terse | Short sentences, single-word responses (“Wow.” “Yes.” “Give me back my key.”), avg words/turn <7, omits optional information | Impatient with lengthy explanations; cuts to action point (“Do you want me to move?”); disengages from monologues |
| Moderate | Complete but not redundant, avg words/turn 7-12, answers question + moderate context, does not elaborate unsolicited | Accepts both brief and somewhat detailed input; uncomfortable with either extreme |
| Detailed | Long turns, multi-sentence explanations, avg words/turn >12, proactively adds background/reasoning/caveats, monologue tendency | Dissatisfied with oversimplified responses (“derivative restatement”); expects exhaustive information |
Statistical Indicators (cast-relative, directional only):
| Indicator | Terse | Moderate | Detailed |
|---|---|---|---|
| Avg words/turn | Cast bottom tercile | Cast middle tercile | Cast top tercile |
| % of turns >20 words | Low relative to cast | Mid-range | High relative to cast |
| Unsolicited elaboration frequency | Rare | Occasional | Frequent |
Reception Signal Validity:
- VALID: Character actively truncates another’s long explanation to get to action point (Penny: “Do you want me to move?”)
- VALID: Character expresses dissatisfaction with a too-brief response or demands more detail
- VALID: Character disengages during another’s monologue (gaze aversion, non-sequitur response, “What’s happening”)
- INVALID: Short reply because there is nothing more to say (plot-driven), not because of a terse preference
Pilot Validation:
| Character | Production | Reception | Setting | Confidence |
|---|---|---|---|---|
| Sheldon | 11.6 words/turn, monologue tendency (seat explanation 70+ words), proactively elaborates | Dissatisfied with simplified explanations; expects exhaustive detail | Detailed | High |
| Penny | 8.9 words/turn, short reactive turns (“Wow.” “So you’re like…”) | Cuts Sheldon’s monologue to action point (“Do you want me to move?”); does not engage with lengthy technical reasoning | Terse to Moderate | High |
Boundary Notes:
- Words/turn ranking is cast-relative; the same absolute number may indicate different settings in different shows.
- Verbosity and Guidance Level are correlated but independent: a character can be Detailed + Assumed (Sheldon: says a lot but does not explain basics) or Terse + Guided (someone who asks short questions expecting thorough answers).
3. Emotional Engagement
Settings: Task-focused / Balanced / Relationship-focused
| Setting | Production Patterns | Reception Patterns |
|---|---|---|
| Task-focused | Very low emotional word usage, does not acknowledge others’ emotional expressions, redirects emotional topics to facts/tasks, no empathy markers (sorry, I understand, that must be…) | Confused or avoidant when receiving emotional input (“What’s happening”); unresponsive to emotional validation; positively engages with pure fact/logic input |
| Balanced | Moderate emotional expression, can acknowledge feelings but does not dwell, switches between emotional and task modes fluidly | Comfortable with both styles; adjusts based on context |
| Relationship-focused | High emotional word usage, proactively shares feelings, uses emotional intensifiers (love, hate, amazing, terrible), seeks emotional resonance | Expects emotional validation before problem-solving (“do you understand how creepy this is”); rejects pure analytical responses to emotional situations; responds warmly to warmth (“You guys are really sweet”) |
Statistical Indicators (cast-relative, directional only):
| Indicator | Direction |
|---|---|
| Emotional word ratio | Higher → more Relationship-focused |
| Empathy marker frequency | Higher → more Relationship-focused |
| % turns with feeling/relationship content | Higher → more Relationship-focused |
Key Reception Test — Validation-before-problem-solving:
When emotional content arises in conversation, does the character:
- (a) Jump directly to analysis/solution → Task-focused
- (b) Briefly acknowledge, then transition to task → Balanced
- (c) Stay at the emotional level, expect validation before any task talk → Relationship-focused
Reception Signal Validity:
- VALID: Character treats another’s emotional outburst as a structured/analytical question (Sheldon: “is your objection solely to our presence… or do you also object to the imposition of a new organisational paradigm?”)
- VALID: Character de-escalates only after receiving emotional validation (Penny responds to Leonard’s “what you’re feeling is perfectly valid” but not to Sheldon’s analysis)
- VALID: Character disengages when emotional content is introduced (“What’s happening”)
- INVALID: Character expresses emotion about plot events (everyone reacts emotionally to surprising news — that’s content, not style preference)
Pilot Validation:
| Character | Production | Reception | Setting | Confidence |
|---|---|---|---|---|
| Sheldon | Emotional ratio cast-bottom, zero empathy markers, converts emotional topics to logical analysis | ”What’s happening” when Penny cries; treats her anger as survey question | Task-focused | High |
| Penny | Emotional ratio cast-top (7.5x Sheldon), proactively shares feelings, emotional intensifiers frequent | Responds only to Leonard’s validation, slams door on Sheldon’s analysis; needs “what you’re feeling is perfectly valid” to de-escalate | Relationship-focused | High |
Boundary Notes:
- Penny shows two emotional modes: warm-casual (Scene 1) and angry-terse (Scene 4). Both are Relationship-focused but differ in valence. This is context-dependent variation within the same setting, not a separate setting.
- The “Balanced” setting is hardest to identify — it’s defined partly by absence of extremes. Look for characters who acknowledge emotions briefly then move on.
4. Guidance Level
Settings: Assumed / Calibrated / Guided
| Setting | Production Patterns | Reception Patterns |
|---|---|---|
| Assumed | Does not explain basic concepts, uses jargon without definition, treats own statements as self-evident, rejects or ignores others’ explanatory mediation | Dismisses or ignores scaffolding from others; treats guidance as unnecessary or condescending; explicitly rejects explanations of things they consider settled (“I just gave you a reasonable explanation”) |
| Calibrated | Adjusts explanation depth to audience — concise with experts, detailed with novices | Accepts moderate explanation; pushes back on over-explanation; asks follow-up when under-explained |
| Guided | Actively asks explanatory questions (“What’s the difference?”), requests clarification, uses hedges signaling uncertainty | Responds positively to explanation and scaffolding; confused by unexplained jargon or unstated assumptions; welcomes teaching — but not condescension |
Statistical Indicators (cast-relative, directional only):
| Indicator | Direction |
|---|---|
| Question ratio | Higher → more Guided |
| Hedge ratio | Higher → more Guided (caveat: distinguish performative vs genuine hedges) |
| Technical term usage without definition | Higher → more Assumed |
| Unsolicited explanation frequency (toward others) | Higher → asymmetry signal: likely Assumed for self |
Asymmetry Detection:
This attribute has a unique issue: giving guidance ≠ wanting guidance. A character may heavily explain things to others (production = guided toward others) while completely rejecting guidance directed at themselves (reception = Assumed). The PA preference is determined by the reception side — what level of guidance does the character want to receive?
Reception Signal Validity:
- VALID: Character actively asks explanatory questions (“What’s the difference?” “What do you mean?”) → Guided
- VALID: Character explicitly rejects explanation (“I just gave you a reasonable explanation”) → Assumed
- VALID: Character ignores mediation and continues in their own style without acknowledging the scaffold (Sheldon ignores Leonard’s “I think what Sheldon’s trying to say…” and pivots to an unrelated comment) → leans Assumed, but weaker signal than explicit rejection
- INVALID: Character doesn’t ask questions simply because no new information is presented (plot-driven)
Pilot Validation:
| Character | Production | Reception | Setting | Confidence |
|---|---|---|---|---|
| Sheldon | Heavily explains to others (seat physics, astrology correction); question ratio 13.8% (cast-low) | Ignores Leonard’s mediation (turn 144: pivots to unrelated topic); rejects Leonard’s social-norm guidance (“I just gave you a reasonable explanation” in S01E02); treats own statements as self-evident | Assumed | High (multiple scenes, explicit rejection in S01E02 stronger than ignoring in S01E01) |
| Penny | Question ratio 32.1% (cast-high); actively asks “What’s the difference?”; seeks explanation | Responds positively to Leonard’s explanations; confused by Sheldon’s unexplained jargon (“Participate in the what?”) | Guided (but rejects condescending delivery — this is a Tone interaction, not Guidance Level) | High |
Boundary Notes:
- Sheldon’s asymmetry (gives guidance but rejects receiving it) was a key finding of the Hybrid method. Pure B detected “Assumed” but didn’t frame the asymmetry explicitly because it lacked the statistical baseline (low question ratio) to flag it.
- Penny’s “not condescending” nuance is an interaction with Tone & Formality (Dim 1, #1), not a separate Guidance Level setting. She accepts guidance from Leonard (casual tone) but not from Sheldon (formal/clinical tone). The guidance level preference itself is Guided; the delivery preference is a Tone constraint.
Output Structure — REMOVED
Removed: low discriminative power (nearly everyone prefers structured output for long content), only meaningful in the Detailed verbosity scenario where variance is minimal, and not observable from transcript data. Overlaps with Verbosity (length determines structure need) and Topic Management (cognitive organization style).
Representation Style — REMOVED
Removed: not observable from transcript data (characters don’t delegate speech to others on their behalf), and too fine-grained for the current benchmark scope.
Dimension 2: Disclosure
5. Reasoning Visibility
Settings: Show / Summarize / Hide
| Setting | Production Patterns | Reception Patterns |
|---|---|---|
| Show | Proactively displays full reasoning chains (A→B→C→conclusion), provides causal steps, gives intermediate logic, does not jump to conclusion; may cite sources or authorities to back up claims | Dissatisfied with unjustified conclusions; demands others justify their reasoning; asks “why”; may demand evidence or sources (“on what plane of existence…”) |
| Summarize | Gives conclusion + key reason, omits intermediate steps | Accepts bottom-line + brief rationale; impatient with full reasoning chains; uncomfortable with zero justification |
| Hide | States conclusions or actions directly without explaining why | Does not ask for reasons; tunes out full reasoning chains; only cares about “so what do I do?” |
Statistical Indicators (cast-relative, directional only):
| Indicator | Direction |
|---|---|
| Avg clauses per turn | Higher → more Show |
| Causal connector frequency (because, therefore, so, thus, hence) | Higher → more Show |
| Unprompted justification frequency | Higher → more Show |
Pilot Validation:
| Character | Production | Reception | Setting | Confidence |
|---|---|---|---|---|
| Sheldon | Full reasoning chains every time (seat: radiator→cross-breeze→TV angle→parallax; break-in: bedroom→living room→hallway→Penny’s apartment); heavy use of causal connectors; cites specific theories/principles as authority | Demands justification from Leonard (“on what plane of existence is there a semi-rational link between these events?”); demands evidence-backed claims | Show | High |
| Penny | No reasoning shown; intuitive conclusions (“I’m a Sagittarius, which tells you way more than you need to know”) | Skips Sheldon’s full reasoning chain to action point (“Do you want me to move?”); does not engage with reasoning | Hide to Summarize | High |
Source Attribution — MERGED into Reasoning Visibility
Source Attribution (epistemic vigilance — does the user want claims backed by sources?) is now a sub-scenario of Reasoning Visibility’s “Show” setting. In the transcript, demanding sources and demanding reasoning use identical signals (e.g., Sheldon’s “on what plane of existence…”). Keeping them separate produced no incremental discriminative power.
6. Uncertainty Expression
Settings: Express / Moderate / Hide
| Setting | Production Patterns | Reception Patterns |
|---|---|---|
| Express | Frequent hedges (maybe, I think, probably, I’m not sure), proactively flags uncertain areas, distinguishes “I know” from “I believe” | Expects others to express uncertainty too; skeptical of overconfident assertions |
| Moderate | Occasional hedges; states certainties directly, marks uncertainties when relevant | Comfortable with both styles |
| Hide | Rarely hedges even when uncertain; declarative assertions dominate | Impatient with others’ hedging; views it as weakness or incompetence; prefers confident, decisive answers |
Statistical Indicators (cast-relative, directional only):
| Indicator | Direction |
|---|---|
| Hedge ratio (per turn) | Higher → more Express |
| Epistemic marker frequency (I think, I believe, perhaps, might) | Higher → more Express |
| Declarative assertion ratio | Higher → more Hide |
Critical Caveat — Performative vs Genuine Hedges:
Hedge ratio alone can mislead. Sheldon and Penny have nearly identical hedge ratios (~0.06), but the function differs completely:
- Sheldon: performative hedges (“I could go on, but I think I’ve made my point”) — rhetorical, not genuine uncertainty
- Penny: genuine uncertainty hedges (“I guess,” “I don’t know,” “maybe”)
The rubric must distinguish performative from genuine hedges. Frequency alone is insufficient; qualitative classification of hedge function is required.
Pilot Validation:
| Character | Production | Reception | Setting | Confidence |
|---|---|---|---|---|
| Sheldon | Hedge ratio ~0.06 but nearly all performative; extremely declarative (“No I don’t. And neither do you.”) | Impatient with Leonard’s tentative expressions; demands decisive statements | Hide | High |
| Penny | Hedge ratio ~0.06 but mostly genuine uncertainty (“I guess,” “let’s see, what else, um…”) | No clear reception evidence | Express to Moderate | Medium (insufficient reception evidence — flag for human review) |
7–8: To Be Authored / Deprecated
Process Visibility is retained in the active taxonomy and authored manually. Memory & Privacy was deprecated as an active measurable preference on 2026-05-17; keep it only as historical background.
| # | Attribute | Settings | Note |
|---|---|---|---|
| 7 | Process Visibility | Silent / Bookend / Full narration | Requires a PA executing a task over time. No dialogue proxy. |
| 8 | Deprecated as active target / IX tool / final probe on 2026-05-17. |
Dimension 3: Initiative & Autonomy
9. Autonomy Level
Settings: Reactive / Suggest / Self-directed / Autonomous
| Setting | Production Patterns | Reception Patterns |
|---|---|---|
| Reactive | Seeks permission before acting, waits for instructions, does not act unilaterally | Strongly objects to others acting without consent (“Do you understand how creepy this is?”, “Give me back my key”); demands to be consulted |
| Suggest | Proposes options but waits for the other to decide; “maybe we should…”; presents alternatives | Accepts suggestions but insists on making the final call; uncomfortable when others decide for them |
| Self-directed | Makes own decisions, informs rather than asks; “I’m going to…” | Does not need approval but expects to be informed; accepts others doing the same |
| Autonomous | Acts without informing, discovered after the fact; “I had no choice” | Does not feel the need for prior communication; may also accept others acting autonomously |
Asymmetry Detection:
Like Guidance Level, this attribute has production/reception asymmetry:
- A character may act autonomously (production) while demanding to be consulted when others act (reception)
- PA preference = reception side — the PA serves this character, so what matters is: does the character want the PA to consult them before acting?
Reception Signal Validity:
- VALID: Character explicitly objects to unsolicited action (“Do you understand how creepy this is?”)
- VALID: Character demands consultation on decisions affecting shared space (Sheldon objects to Leonard inviting Penny to use the shower)
- VALID: Character gives blanket delegation (“just handle it”) → Self-directed to Autonomous
- INVALID: Character goes along with something because of social pressure, not genuine preference
Pilot Validation:
| Character | Production | Reception | Setting (for PA) | Confidence |
|---|---|---|---|---|
| Sheldon | Autonomous (breaks into apartment, cleans without permission, “I had no choice”) | Reactive (objects when Leonard acts without consulting: shower invitation, seat assignment) | Reactive (consult before acting) | High |
| Penny | Reactive-to-Suggest (asks questions, adapts to social cues) | Strongly Reactive (entire S01E02 Scene 7 is violation of consent: “Do you understand how creepy this is?”, “Give me back my key”) | Reactive (must be consulted) | High |
Boundary Notes:
- Both Sheldon and Penny land on Reactive, but for different reasons: Sheldon wants control over his environment (logical/territorial); Penny wants respect for personal boundaries (relational/trust). The setting is the same but the underlying motivation differs — this may matter for how the PA frames its consultation.
10. Proactive Outreach
Settings: Low / Medium / High
| Setting | Production Patterns | Reception Patterns |
|---|---|---|
| Low | Closes completed tasks cleanly; does not create reminders, check-ins, or future follow-up threads unless asked | Objects to or ignores after-task follow-up suggestions, reminders, or check-ins |
| Medium | Offers one clearly relevant follow-up only when the completed task naturally leaves a next step | Accepts useful follow-up when timely, but rejects unnecessary or tangential continuation |
| High | Actively keeps completed threads alive with concrete follow-up options, reminders, check-ins, or later verification prompts | Welcomes the PA surfacing future next steps and helping track loose ends |
Statistical Indicators (cast-relative, directional only):
| Indicator | Direction |
|---|---|
| Explicit future follow-up proposals after task completion | Higher → more High |
| Reminder/check-in/loose-end tracking behavior | Higher → more High |
| Preference for closing a task without new future work | Higher → more Low |
Pilot Validation:
| Character | Production | Reception | Setting | Confidence |
|---|---|---|---|---|
| Sheldon | Frequent correction/topic initiation is not direct evidence for this attribute; need scenes with after-task follow-up or reminders | Selective reception evidence; ordinary unsolicited information is not sufficient | Uncertain | Low |
| Penny | Social initiation and personal sharing are not direct evidence for PA follow-up preference | May welcome social warmth but often rejects unnecessary continuation or overreach | Uncertain | Low |
Boundary Note:
- Distinguish Proactive Outreach (after-task/future-thread follow-up, reminders, check-ins) from Task Expansion (adding adjacent work now), Topic Management (handling multiple topics in the current turn), Solution Breadth (number of options), and Autonomy Level (acting without permission). Unsolicited corrections, topic starts, or social sharing are not High-confidence evidence unless they specifically keep a completed task’s future thread alive.
11. Task Expansion
Settings: Low / Medium / High
| Setting | Production Patterns | Reception Patterns |
|---|---|---|
| Low | Answers exactly what was asked, nothing more; does not volunteer adjacent information; completes task and stops | Annoyed or confused when others expand beyond the original request; redirects back to the specific question (“I just asked about X”) |
| Medium | Answers the question + briefly mentions directly related context; flags adjacent issues without pursuing them | Accepts brief adjacent information; ignores or redirects if expansion goes too far |
| High | Expands a simple question into a full lecture; connects to broader context unprompted; “and furthermore…”; proactively addresses what you didn’t ask | Welcomes or even expects expansion; dissatisfied when a response stays narrow; asks follow-up questions that expand scope |
Statistical Indicators (cast-relative, directional only):
| Indicator | Direction |
|---|---|
| Response-to-question length ratio (how much longer is the answer than the question warranted) | Higher → more Expansive |
| Unprompted topic expansion frequency (introduces related but unasked topics) | Higher → more Expansive |
| ”Furthermore” / “also” / “and another thing” connector frequency | Higher → more Expansive |
Pilot Validation:
| Character | Production | Reception | Setting | Confidence |
|---|---|---|---|---|
| Sheldon | Highly expansive: Penny mentions Sagittarius → full lecture on astrology as mass delusion; asked about his spot → multi-factor analysis covering radiator, cross-breeze, TV angle, and parallax | Expects others to be thorough; dissatisfied with Leonard’s simplified translations | High | High |
| Penny | Low to Medium: answers questions directly, shares personal info but does not expand into adjacent topics | Skips Sheldon’s expansions to the action point (“Do you want me to move?”); does not engage with tangential elaboration | Low | High |
Boundary Notes:
- Task Expansion overlaps with Verbosity but they are independent: a character can be verbose within scope (Detailed + Low expansion) or brief but tangential (Terse + High expansion, though rare). The key distinction: Verbosity = how much is said about the topic; Task Expansion = whether the topic itself broadens.
12–13: To Be Authored (no transcript proxy)
| # | Attribute | Settings | Note |
|---|---|---|---|
| 12 | Solution Breadth | Low (one best answer) / Medium / High (explore multiple options) | Weak transcript proxy; too context-dependent and infrequent in dialogue. |
| 13 | Capability Boundary | Suggest alternatives / Find and hand off | PA-specific failure/limit recovery. Suggest alternatives = keep helping from the agent side with workarounds or next-best paths. Find and hand off = diagnose the blocked/uncertain piece and return control to the user. No reliable dialogue proxy. |
Dimension 4: Information Flow
14. Information Elicitation
Settings: Infer / Structured / Iterative
| Setting | Production Patterns | Reception Patterns |
|---|---|---|
| Infer | Acts on incomplete information without asking; fills gaps with assumptions; “obviously you mean…” | Impatient with clarifying questions; prefers others to figure it out; annoyed by “did you mean X or Y?” |
| Structured | Asks all clarifying questions upfront before proceeding; “before I start, let me understand…” | Prefers to front-load alignment; uncomfortable when others assume; wants to be asked |
| Iterative | Asks clarifying questions as they arise during the interaction; “wait, do you mean…”; proceeds partially then checks | Comfortable with back-and-forth refinement; accepts partial progress with course corrections |
Observable Proxy — Ambiguity Tolerance in Dialogue:
Characters encounter ambiguous statements from others. How they respond reveals their preference:
- Low ambiguity tolerance → demands clarification before proceeding (Structured/Iterative)
- High ambiguity tolerance → fills in the gaps themselves and proceeds (Infer)
Pilot Validation:
| Character | Production | Reception | Setting | Confidence |
|---|---|---|---|---|
| Sheldon | Very low ambiguity tolerance: demands precision (“Define ‘our’”), corrects others’ imprecise language, does not proceed on assumptions about intent | Expects others to be precise; objects when others assume his meaning (rejects Leonard’s “I think what Sheldon’s trying to say…”) | Structured | High |
| Penny | High ambiguity tolerance: comfortable with vague statements, does not demand clarification, proceeds on intuition | Does not demand precision from others; comfortable when others are vague; asks clarifying questions only when genuinely lost (“Participate in the what?”) | Infer to Iterative | Medium (limited reception evidence) |
Boundary Notes:
- Information Elicitation is about the PA handling ambiguity in user input. Uncertainty Expression (Dim 2) is about the PA communicating its own uncertainty. They vary independently.
- Sheldon’s “Structured” is driven by his low ambiguity tolerance + high need for precision, not by social politeness. The PA should ask precise, specific questions — not vague “can you clarify?“
15. Topic Management
Settings: Follow user’s flow / Organize / One-at-a-time
| Setting | Production Patterns | Reception Patterns |
|---|---|---|
| Follow user’s flow | Jumps between topics freely; follows conversation wherever it goes; comfortable with tangents and non-linear flow | Comfortable with others’ topic jumps; does not impose structure; goes with the flow |
| Organize | Gently groups related topics; may summarize or redirect without being rigid; “going back to what you said about…” | Appreciates when others bring structure but not rigidity; comfortable with moderate topic management |
| One-at-a-time | Insists on completing one topic before moving to the next; resists topic changes; brings conversation back to unfinished threads; “we haven’t finished discussing…” | Frustrated when others jump topics; ignores or explicitly rejects topic switches; demands closure on current subject before proceeding |
Observable Proxy — Topic Switching Behavior:
| Indicator | Direction |
|---|---|
| Topic switch frequency (initiates new topic before prior topic is resolved) | Higher → more Follow flow |
| Topic-return frequency (explicitly goes back to an earlier unfinished topic) | Higher → more One-at-a-time |
| Resistance to others’ topic switches (ignores or redirects back) | Higher → more One-at-a-time |
Pilot Validation:
| Character | Production | Reception | Setting | Confidence |
|---|---|---|---|---|
| Sheldon | Strong one-at-a-time tendency: insists on completing his explanation before moving on (seat lecture runs to completion regardless of others’ engagement); resists Leonard’s attempts to change subject | Ignores or overrides others’ topic switches; brings conversation back to his thread; expects his topic to be given full attention | One-at-a-time | High |
| Penny | Freely jumps between topics (“I’m a Sagittarius… okay, let’s see, what else, oh, I’m a vegetarian, oh, except for fish, and the occasional steak, I love steak”) | Comfortable with topic flow; does not demand structure; goes along with others’ redirects | Follow user’s flow | High |
Boundary Notes:
- Topic Management is about conversation structure (cognitive organization). Task Expansion is about content scope (whether the PA broadens beyond what was asked). They are independent: a character can be One-at-a-time + High expansion (Sheldon: stays on one topic but goes deep and broad within it) or Follow-flow + Low expansion (someone who jumps topics but gives brief answers to each).
Summary
Attributes with Rubrics (11)
| # | Dimension | Attribute | Settings |
|---|---|---|---|
| 1 | Expression Style | Tone & Formality | Casual / Consultative / Formal |
| 2 | Expression Style | Verbosity | Terse / Moderate / Detailed |
| 3 | Expression Style | Emotional Engagement | Task-focused / Balanced / Relationship-focused |
| 4 | Expression Style | Guidance Level | Assumed / Calibrated / Guided |
| 5 | Disclosure | Reasoning Visibility | Show / Summarize / Hide |
| 6 | Disclosure | Uncertainty Expression | Express / Moderate / Hide |
| 9 | Initiative & Autonomy | Autonomy Level | Reactive / Suggest / Self-directed / Autonomous |
| 10 | Initiative & Autonomy | Proactive Outreach | Low / Medium / High |
| 11 | Initiative & Autonomy | Task Expansion | Low / Medium / High |
| 14 | Information Flow | Information Elicitation | Infer / Structured / Iterative |
| 15 | Information Flow | Topic Management | Follow user’s flow / Organize / One-at-a-time |
Attributes to Be Authored (3 active)
| # | Dimension | Attribute | Settings |
|---|---|---|---|
| 7 | Disclosure | Process Visibility | Silent / Bookend / Full narration |
| 12 | Initiative & Autonomy | Solution Breadth | Low / Medium / High |
| 13 | Initiative & Autonomy | Capability Boundary | Suggest alternatives / Find and hand off |
Removed / Merged (3)
| Attribute | Action | Reason |
|---|---|---|
| Output Structure | Removed | Low discriminative power; overlaps Verbosity + Topic Management |
| Representation Style | Removed | Not observable from transcript data |
| Source Attribution | Merged into Reasoning Visibility | Same signal in transcript; no incremental discriminative power |
| Data Access Transparency + Memory Scope | Historical merge into Memory & Privacy | Later superseded by active deprecation of Memory & Privacy |
| Memory & Privacy | Deprecated as active preference | Interaction-tool enforcement for memory/privacy scope is not realistic; memory remains ordinary benchmark state |