Method and measures
This page states, in exact terms, how the daily networks are built, what each chart across the site measures, how stance is classified and published, how stages and outcomes are labelled and what the data cannot say. The English section of the site has five pages — the entry, the phenomenon, the model, this one and the cases; the rest of the site is in Portuguese, and a glossary of the chart labels used on the Portuguese pages is at the end of this page.
O Mecanismo follows eight Brazilian online pile-ons (2014–2025) as daily mention networks on X/Twitter and reads them with René Girard’s account of the scapegoat mechanism. Nobody is named, the target included. The empirical question is whether the mechanism leaves a measurable signature while it happens; the answer the data suggest is that it does — concentration of attention, centralisation on one account, and a structural scar that persists after the peak. An agent-based model, built before the data were collected, produces the same shape from a few rules. The project is run by Carlos Paes, an independent researcher and data scientist in Rio de Janeiro (postgraduate in Data Science and Analytics, PUC-Rio). Contact (in Portuguese) →
Data
Collection. Public posts on X/Twitter, in Portuguese, gathered through the platform’s search. In six cases the query followed the person — their name and their account; in the two cases in which the target had no account, it followed the episode — its keywords and hashtags. Windows of 8 to 29 days around the trigger (in one case the whole year, with three months analysed), and between 2,500 and 23,000 posts per case. In four cases the collection was a daily sample with a ceiling of about 1,000 posts per day; days at the ceiling are marked ▲ on the curves. The search does not return plain reposts (retweets). The exact queries are not published, because they contain the names of the people involved. The context before each episode — the dispute already under way — comes from press coverage; when the collection reaches the days before the trigger, it records the conversation around the person or the episode, not that dispute.
Processing. Deduplication, language filtering and normalisation of identifiers. Each day becomes a directed network of mentions and replies: every account is a node, and every mention or reply is a tie from the author to the account mentioned. Days are UTC days.
Minimum network. Days with fewer than 30 accounts in the network are not given structural measures. In the case videos they are drawn, and no total below five is shown.
Measures of the network
All of them describe the structure of the day’s network, never its content, and none is published for an individual account.
- Volume. Posts, accounts and mentions per day. The peak is the day with the most posts from the trigger on (in two cases the days before the trigger carry more posts than any day of the episode: the collection followed a person already at the centre of another controversy).
- Concentration of received mentions. Gini coefficient of the in-degree distribution: 0 when every account receives the same number of mentions, approaching 1 when a single account receives nearly all of them.
- Centralisation. Freeman’s degree centralisation, on in-degree: 0 when all accounts have the same number of ties, 1 for a perfect star. This is the measure behind “how close the day’s network is to a star”.
- Attention on the target. The share of the day’s mentions directed to the target’s account. Undefined when the target has no account on the platform (two cases); the shape of the network is then the main evidence.
- Fragmentation. The number of islands (connected components) and the share of the day’s accounts in the main island.
- Modularity. Louvain communities detected on the main island, for days with 30 accounts or more.
- Roles. Institutions (companies, broadcasters, parties, public bodies), press outlets and secondary targets (accounts that come to be treated as targets during the episode) were assigned by reading each case, account by account, and are published only as aggregates: the share of the day’s mentions directed to institutions, to press outlets and to secondary targets; and, in the cases with continuous collection, the time between the first mention of an institution after the trigger (the institution is called) and its first own post in the conversation (the institution speaks).
- Entry. The share of all accounts of the episode that appear for the first time on a given day, writing or being mentioned.
The one measure of content: stance
A language classifier trained in this research assigns each post one of three positions: it accuses, it contests the accusation, or neither. The position belongs to the post, never to the account, and it leaves the database only in aggregate: as the share of each position per day and per case; in the case videos, as the colour of each interaction (the colour of the post that produced it); and in the derived measures below. The classification is automatic, without post-by-post review; on the test set it is right about three times in four, and it errs more on contestation than on accusation, so the shape of the curves — what rises and what falls — is more reliable than the level of any one of them. In the sampled cases the shares are shares of a sample of the day.
- Opposition figure. For each day, an account that, up to that day, contested more than it accused and is among the ten most mentioned accounts of the day, the target excluded. Only the day the first one appears is published, marked on the case’s curve.
- Accusation turned on contesters. The share of the mentions made in accusing posts that is directed to accounts that had already contested.
- Side changes. An account has a side on a day when it publishes at least two posts of the same position that day, and more of that position than of the other; changing side is having a side on one day and the opposite side on a later day. Only accounts with a side on two days or more are counted. Totals are published per case and per window of days — never the position of an account, and no total below five.
- Sanction and its day. Decided for each case without looking at stance — dismissal, suspension or removal from a role, loss of a contract, elimination, exclusion — from the share of posts that state the sanction as a fact, as distinct from the ones that demand it, confirmed by the reading of the case. Contestation is then compared before and after that day (in the cases without a sanction, the first three days against the rest), together with how organised it is: the share of the mentions made in contesting posts during the first three days that goes to other accounts that had also contested.
- Naming the act and naming the person. For each post, whether it names the act — through a lexicon of the transgression’s terms for that case — and whether it names the person, by name, nickname, account, mention or reply. Only the daily share of posts in each condition is published. The measure depends on how the corpus was collected: in the six cases collected by the person’s name, the person appears in nine out of ten posts by construction; only in the two cases collected by the event is the share that names the person a real measure.
Stages, alignment and outcomes
Cases are aligned on the trigger: day 0 is the day the triggering post or event enters the collection. In the published metrics each day carries a phase — pre-crisis, trigger, escalation, peak (the day with the most posts from the trigger on), decline and post-expulsion. The narrated stages use a canonical sequence shared with the model: baseline → trigger → convergence → peak → isolation → crystallisation (the expulsion) → residue. Each case names its stages with the terms of its own episode, mapped onto that sequence.
Each case also carries an outcome, decided by reading before any stance measure was computed: consummated (the sanction took place), consummation contested, not consummated, or under analysis. Six of the eight cases had a sanction; two did not.
What the videos show, and what they hide
Each case page shows a video of the network, frame by frame: one frame per day and, on the day with the largest network, when the timestamps of the collection are reliable, one frame every three hours (Brasília time, so the first block may show as 21:00–24:00 of the previous day). Every account that interacted in the period enters the frame’s network, including on days with a very small network — the framing follows the main network, and loose islands far from it may fall outside the screen; the video carries no identifiers: each account is only a position in the drawing, a size band and a type — target, account, institution, and, when there is one, secondary target, marked with a dashed ring. Accounts with five or more ties in the period are drawn larger; the others are small, unattributed dots. From one frame to the next each account starts from where it was, so that movement shows a change of relation rather than a new random layout; the target stays fixed at the centre. The frame file that links accounts across frames is not published — only the rendered video is. Institutions are marked only with a letter (A, B, C…), with no name and no role; the same letter is the same account throughout the video. The video itself carries no text: date, phase, totals and legend are written on the page next to it, where a browser translator can reach them. Interactions are coloured by the stance of the post that produced them; in frames with fewer than five interactions, all of them are drawn in the neutral colour.
Findings, in short
The full discussion, with the charts, is in Portuguese at Análise; an English summary of each case is at Cases, and each Portuguese case page ends with “Os números do caso”, the day-by-day figures of that episode.
- In seven of the eight cases the day with the most posts is not the day the network comes closest to a star. The crowd and the star are two events.
- Convergence takes four forms: a star around the target; a star around another pole — an institution, a TV programme, an outlet; an archipelago of islands that do not touch; and a one-day star, in which the target is the centre only on the day the exclusion is consummated.
- Entry has three rhythms: concentrated in the first days after the trigger; almost seven in ten accounts entering on a single day, the day of the exclusion; or flat, one twentieth to one tenth of the accounts per day for one or two weeks. The flat shape appears in the sampled cases, where it may be a limit of the collection rather than of the conversation.
- In the three cases where hours can be measured, the institution is called before it speaks, and speaks one to a little over four days later; in two of them, only after the day of the peak.
- In three cases the accusation extends to secondary targets — in two of them, an institution — which together come to receive about one tenth of a day’s mentions.
- Contestation is present from the day of the trigger in every case, between a sixth and a little over a third of the posts, and it does not grow: its median stays between a sixth and a quarter throughout, with no trend. Accusation is highest on the day of the trigger — a median of six posts in ten — and falls to a little under half over the next three days; what it loses goes to posts that take no side, not to contestation.
- An opposition figure appears in five of the eight cases, almost always early — on the day of the trigger or within the next five days. After it appears, the share of accusation does not change level. Contesters are rarely targeted: the share of accusing posts’ mentions directed at them stays below two per cent on almost every day.
- Almost nobody holds a side. Across the eight cases, 93 accounts have a side on two days or more, out of thousands that write; side changes add up to about thirty in the whole set, in both directions in similar proportions, and they come late.
- Early contestation did not prevent the sanction in these cases: before the sanction, contestation occupies between a sixth and a third of the posts in the six cases where it was consummated, and a tenth and a third in the two where it was not.
- Where the comparison is possible — the two cases collected by the event — the act is named far more than the person: in the 2017 case the act reaches six to eight posts in ten the day after the trigger while the person never passes a quarter; in the 2014 case the person is named in fewer than one post in a hundred. In those two cases the accusation goes to the act and to secondary targets, not to the person.
The model
The Scapegoat model is an executable hypothesis in NetLogo, built before the data were collected. Thirty to a hundred agents on a small-world ring (Watts–Strogatz, taken from NetLogo’s Small Worlds model) accumulate tension and transfer it onto a target; whoever transfers becomes an amplifying pole, and the target stays marked. Once the marked targets exceed a tenth of the living population and a pole exists, the expulsion occurs: one marked target leaves the network, the others adhere to the pole — new ties with it — and the tension of the whole network resets; the expelled agent is replaced by a new one. Three parameters: population (30–100); closure (0–19), which lowers the chance that the network creates new ties; and skepticism (50–99), the probability that a transfer without adhesion is blocked. Five scenarios, twenty runs each (seeds 1 to 20, ceiling of 1,500 steps), time aligned on the expulsion. The curves are on the model page; the logic behind the rules — the transfer that closes a triangle against a third party — is drawn, without formulas, in A Forma (in Portuguese), which rests on structural balance theory (Heider, 1946; Harary, 1953). The model and its ODD description are archived at CoMSES Net (version 1.0.1).
Three of the model’s measures have empirical counterparts: concentration (the share of accumulated transfers that fell on the most targeted agent ↔ the Gini of attention), centralisation (Freeman degree centralisation of the agents present ↔ the same measure on the day’s network) and modularity (Louvain, at each step ↔ Louvain on the main island). Local cohesion (mean clustering coefficient) has no counterpart: the model’s network is the substrate — who knows whom — while the empirical network is the episode itself — who mentioned whom that day.
Headline results: the base scenario (population 60, closure 10, skepticism 75) reaches the expulsion in all twenty runs, at a median step of 546.5, with a median of 10 adhesions to the amplifying pole; the closed network (closure 19) reaches it in only 10 of 20 runs, with the smallest adhesion (3.5); the open network (closure 0) and the skeptical network (95) bring the expulsion forward to median steps of 279.5 and 278; with 30 agents it comes at step 190. In every scenario modularity and local cohesion fall during the episode and have not recovered a hundred steps after the expulsion.
What the data cannot say
- Whether an accusation was deserved. The mechanism has the same shape when the accusation holds and when it does not; the shape decides nothing about merit.
- Who the people are: no account is identified, and stance belongs to the post, not to the person. Contesting an accusation is not the same as defending the target.
- The exact levels of stance: the classifier is automatic and right about three times in four; shapes are more reliable than levels. With side changes in the tens, a single misclassified post already produces one.
- Whether opposition changes the course of an episode: the opposition figure appears together with the escalation, the order of events does not separate cause from coincidence, and the cases without a figure follow the same path.
- Whether a call caused an institutional response: the data show the order of events, not the effect, and three cases are too few for a rule. Actions taken off the platform — a dismissal, a contract ended — do not appear in the data.
- What happened before the window: collection starts at the episode, and the prior crisis is documented from press coverage.
- In the sampled cases, a flat curve may be the ceiling of the instrument rather than the shape of the conversation; and the naming measure is a real measure in only two cases.
- Outcomes and roles are decisions of reading, taken before the stance measures were computed.
Publication rules
Nobody is named, the target included; each case is identified by the transgression, the year and the kind of target. No identifier leaves the research database. No measure is published for an account, and no account is labelled by stance or conduct. When the target is a private person, the page is kept out of search engines and omits the elements that would identify them. No post text and no screenshots: contents are described, not reproduced. Every publication or change of a case is reviewed and logged internally, with date and sources. The full data policy is in Portuguese at Política de dados. Anyone concerned may ask for correction, further anonymisation or removal at [email protected]; requests are answered within seven days.
Theory, in two pages
The phenomenon presents Girard’s mechanism and the four stereotypes of persecution that organise the reading of the cases. A Forma (in Portuguese) draws, without numbers, why a group under tension closes against one person and why reopening is harder than closing — the three balanced forms of a group, “unanimity minus one”, and the pressure on the ties of whoever stays.
Code and contact
The analysis code is available to researchers on request. The model is on CoMSES Net, with its ODD description. Questions about method, data or the model: [email protected].
Glossary of chart labels
The charts on the English pages carry English labels. On the Portuguese pages — the case pages and the analysis — the labels are in Portuguese; the table gives the English for every label that appears on their curves, videos and legends.
| Portuguese | English |
|---|---|
| contas · postagens · menções | accounts · posts · mentions |
| dias desde o estopim | days since the trigger |
| estopim | trigger (the event: day 0) |
| linha de base · gatilho · convergência · pico · isolamento · cristalização · resíduo | baseline · trigger (the narrated stage) · convergence · peak · isolation · crystallisation (the expulsion) · residue |
| escalada · pós-rito · esvaziamento · dispersão | escalation · post-rite (the phases’ name for post-expulsion) · emptying · dispersal |
| votação · expulsão · sanção · sanção anunciada | vote · expulsion · sanction · sanction announced |
| Estágio N de 5 · Dia N · conversa difusa · repercussão institucional · concentração · acusação · acusações de dentro · contágio · declínio · ausência · retorno · acordo · pico do episódio | stage N of 5 · day N · diffuse conversation · institutional repercussion · concentration · accusation · accusations from within · contagion · decline · absence · return · settlement · peak of the episode (stage captions of the narrated graphs) |
| atenção sobre o alvo | attention on the target |
| não se aplica: o alvo não aparece como conta na coleta | not applicable: the target does not appear as an account in the collection |
| centralização · concentração (Gini) | centralisation · concentration (Gini) |
| fragmentação | fragmentation (share of the day’s accounts outside the main island) |
| modularidade · coesão local | modularity · local cohesion |
| ilhas · ilha principal | islands · main island |
| rede pequena demais para as medidas | network too small for the measures |
| N contas · N interações · menos de 5 · nenhuma interação coletada neste período | N accounts · N interactions · fewer than 5 · no interaction collected in this period |
| antes · depois da expulsão · Como ler o vídeo · instituição (letras A–…) · alvo secundário · não há neste caso | before · after the expulsion · How to read the video · institution (letters A–…) · secondary target · none in this case |
| volume no teto do coletor | volume at the collector’s ceiling |
| contas que entram | accounts entering (first appearance) |
| atenção sobre instituições · sobre a imprensa · sobre alvos secundários | attention on institutions · on press outlets · on secondary targets |
| cobrada · fala | called · speaks (institution) |
| primeira cobrança · primeira postagem da instituição · dia de pico de volume · +1d … +7d · N h depois · N dias depois | first call on the institution · the institution’s first own post · day of peak volume · days after the trigger · N hours later · N days later |
| acusa · contesta · nenhuma das duas | accuses · contests the accusation · neither |
| figura de oposição | opposition figure |
| mudança de lado · lados | side change · sides |
| ato · pessoa | the act · the person (naming) |
| Consumado · Consumação contestada · Não consumado · Em análise | consummated · consummation contested · not consummated · under analysis |
| alvo · alvo secundário · conta · instituição A · veículo | target · secondary target · account · institution A · press outlet |
| mediana · um caso · todos os casos | median (thick line) · one case (thin line) · all cases |
| ▲ | day at the ceiling of a sampled collection |
| tensão média da rede (0–3) | mean tension of the network (0–3) |
| medidas de rede (0–1), run principal | network measures (0–1), main run |
| concentração de atenção | concentration of attention |
| transferências de tensão por passo · transferência com adesão | tension transfers per step · transfer with adhesion |
| tentativas bloqueadas pelo ceticismo | attempts blocked by skepticism |
| focos marcados (fração da rede) | marked targets (share of the network) |
| limiar da expulsão (10% da rede) | expulsion threshold (10% of the network) |
| polo amplificador · adesões · expulsão: o foco sai | amplifying pole · adhesions · expulsion: the target leaves |
| semente · linha cheia · faixa | seed · solid line (the fixed-seed run) · band (middle half of twenty runs) |
| passos | steps |
| população · fechamento · ceticismo | population · closure · skepticism |
| cenário base · rede fechada · rede aberta · rede cética · população pequena | base scenario · closed network · open network · skeptical network · small population |