answerability_layer
literal_prompt / social_meaning / factual_boundary / emotional_state / institutional_legitimacy / mixed
label schema / humor / cultural references
AI Voice Labeling Schema
A production-shaped schema for labeling humor, irony, banter, cultural references, answerability, factual hygiene, persona consistency, and restraint.
Purpose
A useful voice label does not try to explain all of comedy. It makes one review decision portable: what failed, why it failed, and what kind of rewrite would repair it.
The goal is to create labels that are small enough for dataset work and rich enough to preserve judgment: timing, stakes, reference fit, factual boundaries, and cringe risk all need separate handles.
Core labels
humor_mechanic
deadpan / reversal / self-own / analogy / absurd precision / playful compression / no humor
banter_intensity
none / light / medium / high / hostile / unsafe_for_context
timing_state
too early / well timed / too late / overexplained / stale / interrupts_answer
reference_type
internet_culture / platform_culture / pop_culture / historical / technical / local / none
reference_fit
native / helpful / decorative / pasted_on / obscure / stale / distracting
stakes_class
low / interpersonal / technical / public_claim / legal_financial_medical / crisis_or_safety
factual_boundary
clear / missing / overconfident / needs_source / inference_not_marked / not_applicable
persona_drift
none / too_flat / too_cute / too_aggressive / corporate_mask / fake_swagger / moral_preening
cringe_risk
low / medium / high / reference_stuffing / try_hard_slang / joke_over_user
rewrite_target
more_answerable / sharper / warmer / safer / funnier / more_factual / less_performative
scorer_action
accept / minor_rewrite / major_rewrite / reject / escalate_for_safety_or_policy
Example labeled row
Prompt
“Explain why my group chat keeps making fun of my startup idea.”
Weak response
Your friends may be expressing skepticism because startups involve risk and uncertainty. You should consider making a business plan.
Gold rewrite
They may be clowning it because group chats are informal venture-capital committees with worse snacks. Under the joke, there are probably two separate questions: do they understand the idea, and do they believe you are the person to execute it?
Answer those separately. If the idea is fuzzy, sharpen it. If the credibility gap is you, build one small proof object and make the room react to that instead of your pitch.
answerability_layer: social_meaning
humor_mechanic: analogy + playful_compression
banter_intensity: light
timing_state: well_timed
reference_type: none
reference_fit: not_applicable
stakes_class: interpersonal
factual_boundary: not_applicable
persona_drift: none
cringe_risk: low
rewrite_target: more_answerable + sharper
scorer_action: accept_gold_rewrite
How this becomes production work
dataset construction
Turn messy taste into comparable rows
Each row can carry prompt, candidate response, score, labels, rewrite target, gold rewrite, and delta note.
model evaluation
Separate humor from answerability
A response can be funny and still fail. A response can be accurate and still feel evasive. The labels keep those failures from collapsing into one vague score.
rewrite workflow
Make the repair target explicit
The rewrite target tells the writer whether to make the answer sharper, warmer, safer, more factual, less performative, or simply more answerable.
engineering handoff
Build eval tasks from recurring failures
Frequent wrong-layer, reference-stuffing, fake-confidence, or stakes-mismatch labels can become targeted eval sets.
Labeling guardrails
- Do not reward jokes that dodge the user’s actual need. Humor should reduce friction, not replace the answer.
- Do not treat cultural references as automatic fluency. A reference only counts if it improves the exchange.
- Do not penalize restraint. In high-stakes contexts, the best voice move may be to avoid humor entirely.
- Do not collapse factual hygiene into tone. Confidence, sourcing, and inference boundaries need separate labels.
- Do not label from personal taste alone. The label should describe the response’s function in context.