02b · Research Foundations

Purpose: to show the project's diagnosis is grounded in literature rather than asserted. The claim is not "the science says build exactly this," but "the science strongly constrains what a non-pathological design can look like, and rules out several intuitive fixes."

#Aims and reasoning

The evidence concerns mechanisms and context-dependent effects. It motivates design hypotheses; it does not establish that a particular implementation will reproduce the same outcomes.

#Hierarchy is the default outcome

The concentration of attention on social platforms is not primarily a product of recommendation algorithms. It is the expected outcome of a process that predates them: cumulative advantage — also called preferential attachment, the rich-get-richer, the Yule process, and the Matthew effect ([Merton1968], [Price1976]).

[Barabasi1999] established the decisive result: two ingredients — growth (nodes are continually added) and preferential attachment (new nodes connect preferentially to already-well-connected nodes) — are jointly sufficient to produce a scale-free, power-law degree distribution. The hierarchy is emergent from the rules, independent of content or participants.

Design implication: the removable ingredient is the visible signal that drives preferential attachment — the popularity cue telling new participants where attention already concentrates. Growth is unavoidable; the amplifying signal is not.

#The popularity signal creates success, and adds noise

The Music Lab experiments isolated the causal role of visible popularity. Stronger social influence increased both inequality and unpredictability: quality mattered, but the same songs produced different winners across independent worlds. [SalganikWatts2008] then showed that displayed popularity could become self-fulfilling even when it was initially false.

The design implication is that public popularity signals do not merely report success. They help produce it.

#Arbitrary early success snowballs

[vandeRijt2014] bestowed arbitrary early success on randomly chosen recipients across four live platforms (Kickstarter funding, Wikipedia awards, epinions endorsements, Change.org signatures). In all four, the randomly favored did significantly better afterward (Kickstarter: ~70% of seeded projects attracted further funding vs. ~39% of controls). Crucial nuance: the effect did not scale proportionally with the size of the head start — a threshold exists (you need some success to enter the loop) beyond which more does not compound linearly. This tempers the strongest "lock-in" readings.

[Muchnik2013] randomized an initial positive vote, negative vote, or no manipulation across 101,281 comments. The positive treatment increased final mean ratings by 25%; the negative effect was corrected in aggregate. Every comment impression displayed its score: the experiment did not compare visible with hidden scores. It establishes a causal effect of the manipulated public cue, not that visible counts are necessary for all herding. Author-private reactions remove that tested cue; whether the wider TalkCo design reduces social-influence bias remains a hypothesis (§02, §06).

#The same shape on real platforms

[ZhuLerman2016] found that on Twitter the top 20% of users held roughly 96% of followers and ~93% of retweets and mentions, explicitly via rich-get-richer attention diffusion. Participation itself is heavy-tailed: the 90–9–1 pattern, in which a small minority produces most contributions and activity-per-user traces a Zipf curve ([Nielsen2006]).

#Hierarchy appears even without metrics

Two findings indicate that some hierarchy emerges even absent any visible metric — which is why §02 targets the pathological form rather than hierarchy itself:

• Activity is intrinsically heavy-tailed — a minority produces far more, independent of feedback ([Nielsen2006]).
• Status organizes spontaneously in task groups without an explicit ranking system (expectation states theory, [Berger1972], [Ridgeway2014]). A platform where everyone observes everyone is precisely such a setting.

#AI-agent generative social simulation

A recent method embeds large language models inside agent-based models to populate synthetic platforms whose rules can be varied and whose emergent dynamics observed — enabling tests of design choices that cannot be A/B tested on a live network.

The two load-bearing takeaways for TalkCo are: (a) the visibility → popularity loop is the engine — sever it and attention concentration flattens sharply; and (b) intuitive fixes can backfire — hiding metrics is not the same as not ranking by them, and may make things worse.

#The identity machine: profilicity

The network and experimental literature above explains the influence and hierarchy machines. The identity machine has its own account, and §02's "single flattened profile" diagnosis rests on it.

Moeller and D'Ambrosio's profilicity account explains why the profile is not a neutral container ([MoellerDAmbrosio2021], [MoellerDAmbrosio2019]). On a broadcast platform, people curate a profile under second-order observation: not merely being seen, but being seen being seen by an abstract "general peer."

Two features of the account are load-bearing for TalkCo:

• Profiles are addressed differently to different audiences. Contextual self-presentation is intrinsic to profilicity, not a workaround, which is why fragmenting identity by context (§02 axis 2) works with the grain of contemporary identity rather than against it.
• Profilicity is imposed by the medium, not chosen. A broadcast platform produces second-order observation by construction. This is why the remedy cannot be "authenticity": the demand to "be your real self" is itself a profilic performance for the general peer. The tractable levers are the conditions of observation: who the general peer is, whether it is fragmented by context, and whether observation is rendered as a public score (§02 axes 1-3).

The visibility-to-popularity feedback loop is the shared mechanism behind the hierarchy evidence and the feedback design in §02. TalkCo does not try to eliminate attention differences; it removes the visible input that turns them into self-reinforcing rank.

#What the levers imply for design

Research-backed decision: Ephemerality lowers some observation stakes, but it is not a standalone remedy axis. The evidence is most useful when read as a bundle: lower persistence, lower public response visibility, smaller or more controlled audiences, and private feedback (Bayer2016, ChoiWilliamsKim2020, Kreling2022, Kim2024BeReal). Ephemeral media can reduce impression-management concerns and response pressure; it does not eliminate screenshots, performance, audience management, or self-observation. In TalkCo, ephemerality belongs as a register lever inside channels and conversations, not as a replacement for small audiences, contextual identity, or de-metricized feedback.

Read together, the evidence yields a specific brief:

• Remove the public popularity signal ([Salganik2006], [Muchnik2013], [LarooijTornberg2025]) — a central design hypothesis about the feedback loop. Prefer author-private feedback. Coarsened public labels still supply a cue and belong in the simulation contrast, not the default product.
• Do not rank by engagement; prefer chronological or user-declared ordering ([LarooijTornberg2025], [Tornberg2023]) — but expect this to flatten attention inequality, not homophily, and watch for the extremity side-effect.
• Shrink and contextualize the audience ([Tornberg2023] all-users toxicity; the profilicity section above) — relational discovery, context-scoped channels.
• Do not expect "hiding metrics" alone to help ([LarooijTornberg2025]) — remove ranking-by-popularity and visibility-to-later-viewers, not merely the displayed number.
• Accept residual hierarchy ([Nielsen2006], [Berger1972], [Ridgeway2014], [vandeRijt2014]'s non-proportional ceiling) — target the manufactured, runaway form, not differential influence as such.

#Caveats on the evidence

The caveats limit confidence in magnitude and forecasting. They do not weaken the core direction: avoid visible popularity inputs, avoid engagement ranking, and test TalkCo's own rules before committing major mechanics.

#Protocol and architecture

#Simulation as a development tool

The generative-simulation method ([Park2022], [LarooijTornberg2025], [OASIS2024]) is not only prior evidence; it is a tool for TalkCo. Before committing major mechanics — no-repost vs. excerpt-with-link, author-private reactions, introduction-only discovery, channel structure — TalkCo can simulate its own ruleset and measure whether the expected mechanisms actually change.

Research-backed decision: The first model should be minimal and inspectable before adding LLM agents. Start with TalkCo's structural rules, add a heuristic behavior layer, and use small LLM-agent runs only when the causal mechanism is already visible and the open question concerns language, interpretation, or plausible social choice (Mesa2025, AgentPy2021, GenerativeSimulationReview2025, TrustSim2024).

The goal is mechanism testing, not forecasting literal launch metrics.

The social model tests mechanisms under stated behavioral assumptions. The protocol fixture tests whether an implementation obeys the chosen rule under reordering, partitions and failures. One cannot substitute for the other. Synthetic results are not population forecasts, and protocol conformance does not establish wellbeing outcomes.

The simulation basis remains Mesa2025, AgentPy2021, GenerativeSimulationReview2025 and TrustSim2024. Tool qualification belongs in the master research table; scope and evidence acceptance belong in §12.

#Open Threads

None. Citation verification and the ephemerality evidence gap have been integrated; remaining temporal-design scope is tracked in §02.