Channels

A channel is a way news of the product reaches a persona from outside the study's own survey. Jahan has three, and a scenario runs any combination of them, including none:

Channel	Modelled on	Information spreads by	A persona can
social_feed	X (Twitter)	a ranked feed of posts	post, comment, like, repost, quote, follow
forum	Reddit	ranked threads anyone can reply to	post, reply, upvote, downvote
wom	word of mouth	one persona telling a close tie	ask a peer, comment, complain, reject, buy

With no channels, nothing spreads: each persona only ever sees the concept in the survey, which makes it the cleanest baseline for "what does the concept alone do?". The survey itself is not a channel. It is the internal survey room a survey wave uses, with one action: answer.

A persona may attempt anything; the channel decides what lands. Upvoting on the feed does nothing, because the feed has no votes. Ignoring always lands.

What a persona sees in one turn

flowchart LR
  subgraph IMP["Impression: everything shown on one channel this tick"]
    E1["Exposure<br/>launch post"]
    E2["Exposure<br/>a friend's reply"]
    E3["Exposure<br/>a stranger's post"]
  end
  subgraph VIEW["View: the public context"]
    V1["likes · reposts · replies<br/>up/down votes · thread<br/>tie strength to each author"]
  end
  IMP --> T(("Turn"))
  VIEW --> T
  M["the persona's own<br/>beliefs and memories"] --> T
  T --> R["Reaction<br/>action · words · belief change<br/>· which stimulus it is about"]

• An exposure is one thing shown to one persona, with why it got through and how much attention it drew. A persona is shown at most exposure_budget (default 3) per channel per tick; anything cut is logged with a reason.
• An impression is everything one persona is shown on one channel in one tick. Personas react to the impression as a whole, not to each item separately, because seeing two things side by side is not the same as seeing each alone (ADR 0003).
• The view is the public context: engagement counts, the thread an item belongs to, and the persona's tie strength to each author. It never shows other personas' private beliefs, and never how the study is turning out. A persona that could see running adoption would react to the very result being measured.
• Social proof appears one tick late. Engagement becomes visible from the tick after it happens, never within the same tick, so the order in which personas take their turns cannot matter (ADR 0010).

The reaction records both the whole impression it saw and the one stimulus it is about.

The feed (X-like)

The feed follows X's open-sourced home timeline as ported by OASIS¹:

1. In-network first. Posts by the persona's ties and follows come first, most-liked first.
2. Then everything else, by interest × recency. For a persona with profile vector \(\mathbf{u}\) and a post with vector \(\mathbf{v}\), published \(\Delta t\) ticks ago:

\[ \text{score} = \cos(\mathbf{u}, \mathbf{v}) \times \ln\!\left(\frac{271.8 - \Delta t}{100}\right) \]

The recency factor is OASIS's, kept verbatim: it is about 1.0 for a new post, about 0.91 after 24 ticks, falls below zero after 171.8 ticks, and a post older than 271.8 ticks can no longer be recommended.

Profile and post vectors come from a separately pinned ranking embedding model (--recsys-embed-model), in the spirit of TwHIN-BERT², X's own social text encoder. The vectors are embedded once and refreshed only when a persona posts. They never enter purchase-intent scoring, which has its own embedding model.

When no ranking model is pinned, the feed falls back to random order: a seeded shuffle that ignores engagement entirely. Random is also the control arm: comparing a ranked feed with a random one separates effects caused by the ranking from effects that spread organically.

The forum (Reddit-like)

Threads are ranked by Reddit's hot score, copied verbatim from Reddit's original source as ported by OASIS. For a thread with \(u\) upvotes and \(d\) downvotes, \(s = u - d\), published at time \(t\) in seconds:

\[ \text{hot} = \operatorname{sign}(s) \cdot \log_{10}\big(\max(\lvert s \rvert, 1)\big) + \frac{t - 1134028003}{45000} \]

The time is when the post was published, not its age: later posts score higher. In Jahan \(t\) is the publication tick times the tick unit's seconds, since simulated time is declared in ticks.

Worked example: freshness beats votes

Daily ticks. Thread A was published on day 0 and has 10 net upvotes; thread B was published on day 1 and has none.

• A gains \(\log_{10} 10 = 1\) from votes.
• B gains \(86{,}400 / 45{,}000 = 1.92\) from being one day newer.

B ranks above A. To keep its place, A would need about \(10^{1.92} \approx 83\) net upvotes. That is what makes Reddit's front page turn over, and what makes herding fast: early votes on a new thread compound.

The forum also has a community-scoped preset in the engine: threads visible only within a persona's community, ranked by agreement minus age, no hot score. It produces slow hardening rather than fast herding. It exists in the world module and its tests; studies currently run the global Reddit preset.

Word of mouth

Word of mouth is not a platform. It is the social network carrying news from one persona to another. After a reaction, a persona tells a peer when both hold:

• it felt strongly, with sentiment strength at least 0.6;
• the two are close, with tie strength at least 0.3.

Sentiment strength is the largest move in the reaction: the biggest absolute belief change across belief dimensions and claims, or, when the reaction carries a scored intent with expected rating \(\mathbb{E}[x]\) on the 1–5 scale, its extremeness

\[ \frac{\lvert \mathbb{E}[x] - 3 \rvert}{2} \]

which is 0 for a neutral 3 and 1 for a certain 1 or 5. A persona tells at most 2 qualifying peers per tick, chosen by a seeded shuffle, so one strong reaction cannot cross a dense community in a single tick. The message arrives as an exposure on the next tick, never the same one.

Worked example

A persona's intent distribution is \((0.02, 0.03, 0.10, 0.35, 0.50)\), so \(\mathbb{E}[x] = 1(0.02) + 2(0.03) + 3(0.10) + 4(0.35) + 5(0.50) = 4.28\) and sentiment strength \(= \lvert 4.28 - 3 \rvert / 2 = 0.64 \ge 0.6\). It has ties of strength 0.58, 0.41, 0.22 and 0.35. Three qualify (≥ 0.3); two of them, drawn by the seeded shuffle, hear about the product next tick.

Launch reach. When word of mouth is the only channel, nobody has anything to pass on at first, so a share of personas (launch_reach, default 10%, chosen at random) hear of the product first-hand at launch. Feeds and forums need none: their launch posts reach whoever is active.

---

1. Z. Yang et al. (2024). "OASIS: Open Agent Social Interaction Simulations with One Million Agents." arXiv:2411.11581. Code: https://github.com/camel-ai/oasis (Apache-2.0). Reddit's hot ranking originates in Reddit's open-sourced code (_sorts.pyx, 2010). ↩

2. X. Zhang et al. (2023). "TwHIN-BERT: A Socially-Enriched Pre-trained Language Model for Multilingual Tweet Representations at Twitter." KDD 2023. arXiv:2209.07562. ↩