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.
| 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. |
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.
Ordered by perceived impact ÷ implementation cost. Status: new not built · part half exists · built exists, needs dialing.
| # | 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; the profile’s own Voice section carries them as dialog (single-source: the quote list reshaped, not duplicated). 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 | built 08-31 (banksy · jeff-koons pilots + the template rubric) |
| 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) | built 08-31 (the [VOICES] line rides the staging bundle; speakers parsed Name(-anchored, fail-open to all voiced hosts) |
| 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 | built 08-31 (“On other artists” sections, attested-or-[inferred]; template §8b) |
| 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 | built 08-31 (register rules 8+9, both SP variants: no shared opener, no shared skeleton; name-openers only where the profile says so) |
| 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/ | built 08-31 (exam/diverge.py — prod baseline INDEX 0.380 over 417 co-seated pairs · exam/ooc_battery.py — 9 probes, dry by default) |
| # | 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. | megaprompt + FP (server-side only, owner’s call) | built 08-31 — the firing design: the FP owns it (“two quick bubbles” in the pace vocabulary + a WHEN bullet: casual/playful/warm beats only, one host, never twice running, never analysis/serious/tender/game; the megaprompt allows a self-initiated burst on the same beats). Each block already lands as its own bubble — no client work. Typing-cadence stagger stays open (client). |
| 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. | memory block manners | built 08-31 — reconciled with the measured 「you told me」 defect: SPECIFICS without PROVENANCE (name the project/city/number; never where it came from) |
| 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 |
| # | 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 — recon 08-31: reflections already reach the per-persona memory page (reflection:true rows); the remaining delta is a DIARY-VOICE presentation pass, which must honour the 600-column + row-language laws — its own design pass, proposed separately |
| 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 — BLOCKED on an owner decision: tts-providers.js is the md-viewer’s browser-side path on the USER’s OpenRouter key; a persona voice note needs a server-side TTS provider + key on the box (+ a billing kind + an audio store). Options to pick first. |
| 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 — DEFERRED (owner 08-31: this is the Vibes track’s ground; revisit there) |
| 12 | Corrected-typo experiments | Strongest single humanness cue tested — but dosage-untested, register-risky in Chinese, and pattern-detectable at frequency. | built 08-31, owner-shaped: CODE owns the rarity (the punch slot’s seed arms ~1 turn in 13; the model still gates by beat, so the lived rate is far lower), the license names REAL slips only (adjacent key · 的/得 · 在/再 · teh→the, never inside a fact or a number), the correction arrives as its own tiny bubble, never on kit rooms or early turns. |
Nothing here is built (except where marked) — each specimen is the lever made concrete, in the app's own dress. Chinese where the room speaks Chinese; annotations in English.
3–6 exchanges, attested lines wherever the corpus has them (the first two are real), each affect-tagged, at least one group interjection — the situation our rooms actually put him in and 1:1 example formats never cover. The greeting doubles as the strongest single style sample (C.ai's own finding).
The freshest tokens in the whole context, renewed every turn — SPR made cheap. The voice: cards shipped 08-31; this block is the one-line code change that makes them fight drift directly.
Banksy reacts to Koons, unprompted — not to the guest. Persona↔persona friction on standing priors is the group-room mechanic no 1:1 app can copy, and the literature's open gap (multi-agent groups read as social groups; friction-as-aliveness unmeasured).
Koons keeps the name-drop — his profile documents it ("often drops the listener's name into the sentence"); Banksy loses it. The ban targets the shared shape, never a persona's own authentic tic — same-model outputs converge to shared templates by default (Artificial Hivemind).
Numbers illustrative — the meters are the build. Structural-similarity per Chameleon's Limit; boundary probes per ERABAL. Convergence regrows, so every lever above reports to this ruler.
Same generation, split server-side at natural seams, painted staggered with WPM-plausible delays and a per-persona typing row. Five bubbles beat one block with a larger effect than adding an avatar (Chen 2022); timing is the tell users hunt and fail to catch (HUMA). The turn split already made per-host bubbles — this is the within-host half.
A specific earlier moment, by name — the canonical「it's alive」trigger across every companion app's reviews. The store (memory row ③) exists; the megaprompt nudge is the whole build. The fence: precise recall of something sensitive reads as surveillance — the visibility/incognito rules stand in front of this.
Re-engagement messages carry the strongest retention evidence of any mechanic reviewed — and the same study frames the emotionally-loaded variants as manipulation (arXiv 2508.19258). References the last conversation, asks rather than assumes, never guilt. Built (⑦), running dry on prod; the console's Initiate tab is the switch.
The companion-POV diary is Replika's most-cited「why it feels real」feature — and our dream/reflection pipeline already writes this content nightly; the build is a reading surface. First person, something held back from the room itself: an interior.
A note, not a call — the lighter cut of what 猫箱 gates behind VIP (revealed preference). tts-providers.js already exists; rarity is the point: a voice that arrives sometimes is an event, a voice that always arrives is a setting.
The persona exists when you're not talking to it — 筑梦岛/SoulLink's mechanic, and the natural surface for the weekly recent.md sweep a living figure already gets. Heaviest lever here: a new surface plus scheduled generation.
The strongest single humanness cue ever measured — 5 experiments, N≈3,400, beating name/gender/photo head-to-head (Imperfectly Human). And the most dangerous at frequency: an agent that typo-corrects on schedule patterns out, and the un-corrected variant flips to incompetence. Run it as a rare A/B curiosity or not at all.
voice: notes is the working mitigation; the divergence meter is the watchdog.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.
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).