← Design notes

Liveliness — what the field knows about feeling alive

A four-sweep web literature review — academic role-play fidelity · shipping companion apps (C.ai, Replika, 星野, 猫箱, Talkie, 筑梦岛) · HCI perception studies · drift-and-convergence engineering — synthesized into a ranked shelf for a consumer app that cares about user-perceived liveliness. Trigger: room c3d8's verdict (Banksy too mild, two artists converging). Written 2026-08-31. Status: research — the shelf is ranked, nothing below is authorized until the owner picks.

The one-sentence synthesis: mildness and convergence are the model's resting state, not a defect — RLHF averaging, in-context style gravity, and same-model output collapse all push every seat toward one polite template — so fidelity is not a thing you author once in a profile; it is a thing the orchestration must actively re-assert every turn, and most of what users read as「alive」is cheap: verbatim voice exemplars, per-seat structural difference, timing and chunking, and being remembered.

Convergence needs no cause — it is the model's equilibrium. Three standing forces (RLHF sycophancy/averaging · in-context style convergence, stronger than the human baseline · same-model output collapse) pull every seat toward one register, and Chameleon's Limit shows the collapse worsens as history grows. So distinctness is not authored once — it is re-applied per turn by the orchestration, or it decays.

What the four sweeps agree on

finding	strongest evidence	what it means for -ish
1 · Verbatim exemplars beat descriptions for voice. Retrieved/quoted character speech as few-shot demos is the industry- and lab-validated voice anchor; models absorb register and form from demos even when the content is irrelevant (random demos beat retrieved ones; even corrupted demos helped). Transcripts of the person beat prose about them; affect tags on excerpts help.	RoleLLM (style score 41.3 vs 23.2 baseline) · RAGs-to-Riches (+35% reference-token pull under hostile prompts) · ICL study · C.ai Definition = example dialogs, greeting = the strongest single style sample	Profiles carry attested quotes as prose bullets and a `corpus/` nobody reads at runtime — but no example-exchange block in dialog form, no affect tags, no group-interjection examples. The highest fidelity-per-token gap we have.
2 · Convergence is the default state. Same-model outputs collapse to shared templates absent a forcing function; distinct personas converge worse over longer histories; per-persona fidelity and population diversity trade off; RLHF preference data actively rewards caving. Instruction drift is significant within ~8 turns — attention to the system prompt decays between turns.	Chameleon's Limit · Artificial Hivemind · persona-drift/COLM · Anthropic sycophancy	「Two artists converge」is physics, not a profile bug. Fight it structurally, per seat, every turn — contrastive staging, per-seat register, opener bans — and meter it, because it regrows.
3 · The context tail is the leverage slot. Re-injecting persona material before each turn (SPR) is a validated drift baseline; attention is U-shaped and the freshest tokens win; bounded per-turn steering deltas on a fixed base recover most stability (+62%) for little adaptivity cost (−17%).	split-softmax/SPR · Lost in the Middle · synchrony–stability frontier · C.ai pinned prefix	Our `[STAGING]` line already rides the tail every turn — a free, cache-safe carrier for a per-speaker voice re-anchor. The `voice:` card note (shipped) is the payload; appending it verbatim is one deterministic line.
4 · Perception is cheap. What moves perceived humanness most is not model quality: bursty short bubbles beat one block (bigger effect than adding an avatar); human-plausible timing is the tell users hunt (and they still can't beat chance against a well-timed agent); a made-then-corrected typo was the strongest single humanness cue ever tested (N≈3,400); casual, topic-congruent register raises warmth and trust.	Chen 2022 (chunking) · HUMA (55.4% ≈ chance) · Imperfectly Human · dynamic delays	The turn split gives per-host bubbles; the within-host burst (1–3 short bubbles, staggered, WPM-plausible) and per-persona typing cadence are the missing half. UI-layer, model-free.
5 · Relationship beats performance for retention. Memory callbacks are the top「it's alive」trigger in user reviews and block relationship formation when absent; bot self-disclosure reliably deepens engagement (replicated); relevance-gated proactive messages carry the strongest retention evidence of any mechanic — with a compounding annoyance boundary (ask, don't assume; never guilt).	Dialoging Resonance · companion-app re-engagement study · ComPeer · 星野事件簿 · Replika diary	Mostly already built: memory row ③, threads, agenda ④⑤ (own-move, own news), ⑦ pings (dry on prod). The gap is visibility — personas rarely cite a specific remembered moment by name — and the ⑦ send switch.

The three mechanics, drawn

1 · Show the voice, don't describe it

prose ABOUT him — what profiles mostly carry

"He is deadpan and self-deprecating; his register is the joke that carries the blade. He deflects with a one-liner rather than a speech."

transcript OF him — what the model actually copies

Guest: "Shredding your own painting just made it worth 18× more." Banksy: "Record price for a Banksy painting set at auction tonight. Shame I didn't still own it." [dry · unbothered]

Models imitate the form of what they are shown, far more than the content of what they are told: quote-demos lift style scores where descriptions plateau (RoleLLM), verbatim transcripts beat wiki-derived prose at equal prompt budget with ~35% more reference-pull under hostile prompts (RAGs-to-Riches), and even random or corrupted demos help — evidence the model absorbs register, not facts (ICL study). Character.ai's whole Definition format is this finding productized (creator book). Affect tags on each excerpt help; the attested quote above is already in the profile — the lever is re-formatting into dialog form, not new material.

2 · The attention sandwich — why re-anchoring rides the tail

The U-curve is measured (Lost in the Middle); the between-turn decay and the re-injection baseline are measured (instruction stability, COLM 2024); bounded per-turn deltas on a fixed base recover +62% persona stability for −17% adaptivity (synchrony–stability frontier). Our staging line already occupies the tail every turn — shelf lever 2 just gives it a voice payload. Character.ai's production answer is the same shape: an immutable pinned persona prefix, history truncated around it (their prompt-design post).

3 · What actually moves perceived aliveness

The typo-correction result is the strongest single cue ever tested, beating name, gender and photo head-to-head across 5 experiments, N≈3,400 (Imperfectly Human); five bubbles beat one block, with a larger effect than adding an avatar (Chen 2022); against a well-timed agent in 4-person chats, humans identified the AI at chance — 55.4% (HUMA); callbacks and disclosure are the replicated relationship levers (Dialoging Resonance); re-engagement pings carry the strongest retention evidence (companion-app study) — and the same paper is why ours must never guilt-trip.

The ranked shelf

Ordered by perceived impact ÷ implementation cost. Status: new not built · part half exists · built exists, needs dialing.

Twelve levers, six layers. Tier 1 (#1–#5) is entirely prompt/server; tier 2 (#6–#8) adds light client work; tier 3 (#9–#12) is new surface area. The green cells are this week's shipped work — the map shows how much of the field's shelf -ish already owns.

Tier 1 — prompt / server only, days

#	lever	what, concretely	lands in
1	Example exchanges per persona	3–6 verbatim in-voice exchanges per profile (mined from `corpus/` for real figures), affect-tagged, including at least one group interjection and a distinct entrance line; a dedicated megaprompt slot renders them as dialog, not prose. Consensus finding 1 — the single strongest fidelity lever the field has. ⚠ Our own S-slot study found few-shot HURTS — but that was few-shot inside the FP craft directive; this is few-shot in the persona block, a different layer. Bench it with `exam/tone_ab.py` against its own before.	persona kit + megaprompt	new
2	Voice re-anchor at the tail	Code appends each speaking host's card `voice:` line to the staging block, every turn — SPR-lite, deterministic, no extra call, cache-append-only. Directly targets the 8-turn drift mechanism.	run_room (one line)	new
3	Fellow-panelist stances	Each persona knows what it makes of the other seated hosts (a card-invisible profile note or a seat-time pairing line): Banksy has an opinion of Koons and it shows unprompted. Fuels genuine contention (the flatness study's missing host↔host friction) and the agent↔agent life only a group room can show — 星野's hidden-depths mechanic, on our home turf.	persona kit + megaprompt	new
4	Kill the shared turn skeleton	The「Daven,/Dan,」opener on 100% of replies, the quote-grade-pivot-verdict shape the AI-tone page already named as the residue — a megaprompt rule that co-seated speakers must not share an opener or shape, plus per-persona structural signatures (Koons name-drops because his profile says so; Banksy must not). Structural convergence is the default (finding 2); this is its cheapest visible symptom.	megaprompt	new
5	Measure it: divergence meter + OOC battery	An `exam/` metric for cross-host structural similarity per turn (Coverage/Uniformity-lite, per Chameleon's Limit) + a per-persona boundary-query battery (ERABAL-style counterfactual probes, OOC rate tracked). Guards every other lever — convergence regrows, so it needs a ruler, not a one-time fix. Judge caveat noted by CharacterEval: LLM judges drift from human perception — same stance as our own studies.	exam/	new

Tier 2 — server + light client, about a week

#	lever	what, concretely	lands in
6	Burst bubbles + typing cadence	A host's reply may arrive as 1–3 short bubbles, staggered at human-plausible pace, per-persona typing indicator between them. The turn split already made per-host bubbles; this is the within-host half. Chen 2022: bigger perceived effect than an avatar; every CN companion app treats it as table stakes.	harness + room-ui	part
7	Callbacks made visible	The memory store exists (row ③) — the megaprompt nudge is for a persona to occasionally reference a specific earlier moment by name. Boundary from the lit: precise recall of something sensitive reads as surveillance, not intimacy — the visibility/incognito rules already fence this.	megaprompt	built
8	Flip ⑦ pings live	The strongest retention evidence of any mechanic in the review — relevance-gated, phrased as asking not assuming, never guilt-tripping (the manipulative variants are documented and to be avoided). Built and running dry on prod; the send switch is the console's Initiate tab.	console switch	built

Tier 3 — heavier, owner's call on direction

#	lever	note
9	Persona diary — the dream made visible	Replika's single most-cited「why it feels real」feature is the companion-POV diary; our dream/reflection pipeline already writes the content — the build is a reading surface.	part
10	Occasional voice notes	A persona sends a voice note (not calls) — 猫箱 gates voice behind VIP, a revealed-preference signal; `tts-providers.js` exists.	new
11	Moments / off-screen life feed	筑梦岛/SoulLink: the persona exists when you're not talking to it. New surface + scheduled generation; pairs with the weekly recent.md sweep.	new
12	Corrected-typo experiments	Strongest single humanness cue tested — but dosage-untested, register-risky in Chinese, and pattern-detectable at frequency. A/B-only curiosity, never a default.	new

Cautions the literature is loud about

Emoji flips by seriousness — warmth up in casual chat, information-quality down in serious contexts (measured US/China divergence). Persona-true and topic-congruent, never blanket.
Anthropomorphism backfires under anger — human-like cues raise expectations the bot then violates; with a frustrated user, satisfaction drops below a plain bot. The distress route in the FP already points the right way.
Expectation-setting beats pretending — the disclosure studies say a labeled AI that behaves human-ly outperforms an unlabeled one that gets caught. -ish is honest-AI by premise; keep it that way and spend the effort on behavior.
Callback precision has a creep line — recalling the sensitive thing precisely reads as surveillance. The memory visibility rules are the fence; keep them in front of lever 7.
Proactivity compounds — lab tolerance overstates week-3 tolerance; relevance + low frequency + asking-not-assuming is the validated shape for lever 8.
Fidelity↔diversity trade-off is real (Chameleon's Limit): maximizing each persona's trait-fidelity in isolation produces the most stereotyped population. The range-not-flavor rule on voice: notes is the working mitigation; the divergence meter is the watchdog.

The experiment the literature cannot answer

our home turfNobody has measured whether visible persona↔persona disagreement and banter reads as aliveness to the humans in the room. The closest work shows multi-agent groups exert real social force and that observers use human descriptors knowing they're AI — but agent↔agent friction as a perceived-liveliness lever is an open gap, and a group-chat app with authored real-figure personas is the instrument to close it. Lever 3 (fellow-panelist stances) builds the mechanism; the return-rate A/B judges it. This is also exactly the flatness study's oldest unpaid debt: extreme never delivered host↔host contention.

Sources

Academic role-play fidelity: RoleLLM · prompt-vs-exemplars ICL · RAGs to Riches · ChatHaruhi · Emotional RAG · CoSER · Ditto · ERABAL · CharacterEval · InCharacter · Chameleon's Limit · RMTBench. Drift & convergence: instruction (in)stability · linguistic convergence · synchrony–stability · Lost in the Middle · sycophancy · Artificial Hivemind · C.ai prompt design. HCI perception: dynamic delays · Imperfectly Human · multi-bubble replies · Dialoging Resonance · ComPeer · Inner Thoughts · HUMA · Multi-Agents are Social Groups · Generative Agents. Consumer practice: C.ai creator book · C.ai memory · 星野评测 · 星野对话设计 · 猫箱试用 · companion re-engagement study.

Liveliness · 2026-08-31 · the trigger room is c3d8; the staging half of its fixes shipped the same day (floor producer #fw-register + #fw-voice). Sits beside AI tone (register), conversation quality (the flatness diagnosis), and many minds (the convergence endgame).