In the last week of February 2026 and the first week of March, an escalation involving Iran moved through Telegram in every major language at once. It did not move as one story. Read side by side, the Persian, Arabic, English, Turkish, French and Portuguese layers describe events so different in emphasis that an analyst reading only one of them would reach conclusions the others do not support. This page is a surface reading of that fortnight. It takes no position on the events themselves. It measures only how they were published.
Compared against the five preceding weeks as a baseline. Markers are lexical and overlapping. No channel, organisation or individual is named on this page, by design.
No client, no commissioned collection, no hypothesis set in advance. Six languages were read across eight consecutive weeks and the differences were allowed to appear on their own. Everything below is a direction of travel rather than a finished analysis. Point the same instrument at a defined question and it stops being a direction and becomes an answer.
Within the posts that mention Iran at all, the balance of framing differs so sharply between layers that they are barely describing the same period. In English, military framing dominates and diplomacy runs alongside it. In Portuguese the event is almost purely military and almost entirely without politics. In Turkish, military framing is practically absent even though Iran itself is discussed constantly.
One layer down: which channels inside each layer carried which frame, which of them changed frame mid-event, and which frames travelled between languages rather than arising independently.
Negotiation, talks, envoys and settlement language occupy a substantial share of the English layer and almost none of the Arabic or Turkish ones. The audiences physically closest to the events discussed them least in the vocabulary of resolution. Where a diplomatic frame exists strongly in one language and is absent in the neighbouring ones, that gap is itself a finding, and it is measurable week by week rather than asserted.
One layer down: whether the diplomatic frame ever crossed into regional languages, in which channels, and with what delay.
Persian-language publishing peaked in the week of 23 February. Arabic peaked a week later, and was still climbing at the moment Persian volume had already fallen by more than a third. External coverage reached its maximum after the internal layer had begun to withdraw. Anyone reading the loudest layer at any given moment was, by construction, reading the wrong week.
One layer down: the full lead-and-lag structure between layers, which is the difference between watching an event and anticipating one.
Across the two weeks following the peak, the number of Persian-language channels publishing at all fell by roughly a quarter, while circumvention vocabulary - proxies, mirrors, access instructions - rose within what continued to be published. In most environments the number of active publishers is a constant and can be ignored. Here it moved, and it moved in one language only. Connectivity ceased to be background and became part of the measurement.
One layer down: the disruption window mapped precisely, by region and by hour, with the channels that stayed online identified.
Attention to Iran rose sharply in every external layer during the escalation: roughly fivefold in Turkish, more than threefold in French and Portuguese, close to threefold in Arabic and English. In Persian it barely moved, because it was already the subject and had been all along. The escalation interrupted everyone except the people inside it. That asymmetry is the single most useful thing to know before reading any of these layers as evidence of sentiment.
The Turkish layer increased its attention to Iran more than any other, and almost none of that attention was framed as war. It ran instead through neighbourhood, borders, security services, energy and relations with third states. A layer can be highly engaged with an event and describe it in an entirely different register from everyone else, which is exactly the kind of divergence that a single-language monitoring programme is structurally unable to see.
One layer down: the full register map per language, and which actors are addressed rather than described.
The distinctive terms of the escalation weeks separate the layers more cleanly than any metric. Summarised without naming sources:
Arabic mourns, Persian locates, English negotiates, Turkish calculates, French and Portuguese narrate. That is one fortnight, six languages, one instrument.
The same weeks were read again through the response layer: reactions on each post, the comment traffic each post generated, the reactions attached to those comments, and the composition of the reactions themselves. Publication and reception do not line up, and where they diverge is more informative than either on its own.
Connectivity, circumvention and access to the network account for about three percent of the material carrying any marker in this cut, and they draw about thirty-three reactions per post - double the rate of military material and the highest measured anywhere in this corpus. Diplomacy and negotiation follow at about twenty-six, also well above the corpus norm, on a similarly thin volume. The two thinnest subjects in the publication layer are the two the audience responded to hardest. Publication volume is, in this window, a poor guide to what the audience was looking for.
An earlier version of this page reported the opposite for connectivity, on a pass covering only the fraction of the selection carried in the live index. Read across the full archive, that subject inverts from the weakest response to the strongest. The correction is recorded in the cut reference.
Across every marker the dominant reaction is affirmative rather than hostile, and hostile reactions stay inside a narrow band between nine and thirteen percent - a composition markedly unlike other regions measured on the same instrument, where anger leads and separates the subjects. Military, energy and nuclear material sit around sixty percent affirmative. Connectivity is the exception at about thirty-eight, and the reason is visible in the distribution: roughly a quarter of the reactions on that subject are channel-specific custom emoji rather than the shared vocabulary everyone else uses. On this subject the audience answers in a private register, inside closed communities, which is itself consistent with what the subject is about.
One layer down: the same composition per language, and the point in the window at which affirmation gives way to doubt.
Material on civilians, evacuation and humanitarian consequence carries the highest hostile share in the corpus at about thirteen percent, but the composition is grief rather than rage: affection is the leading reaction at roughly three in ten, followed by sorrow and heartbreak. It also carries the lowest comment rate of any subject, about one reply for every nine posts. The audience marks this material and does not debate it. Reading its hostile share as anger, without the composition underneath, would invert the meaning.
The highest-response replies were read back against the findings above. In the full archive they are overwhelmingly Persian-language and directly concerned with the events and with domestic authority: comparisons between the current casualties and those of 1979, demands that the state be condemned before it is defended, arguments about referendum and legitimacy, and first-person accounts of families divided across the two sides of the argument. Arabic replies carry a different register, of endurance and steadfastness rather than argument. English-language replies exist but are a minority and are largely generic geopolitical point-scoring detached from the events.
This corrects an earlier reading on this page, taken from the live-index subset, which found the reply layer to be predominantly external and non-regional. Across the full selection that is not the case: the reply layer belongs to the same space as the publication layer. The operational point stands in a weaker form - on a subject with global attention, part of the comment volume is external and separating it by language of reply remains necessary - but the bulk of the response here is local.
One further constraint holds across every subject: comment rates run between one reply for every nine posts and one for every four. This is a broadcast environment rather than a discussion environment, and reply volume should not be used as a sentiment proxy in it.
One layer down: the reply layer separated by language and by the location profile of the channel carrying it, with the point at which external commentary enters each subject.
Response figures are computed from the same cold archive as the publication figures, across all 5,993 channels in the selection and the full window. Reactions are deduplicated per post and emoji before aggregation; comment traffic and reactions on comments resolve only across the channels that have discussion enabled. Reaction composition is reported as shares within each subject, not as absolute counts. Comment findings are drawn from the highest-response replies; no channel, account or individual is named, and no comment is reproduced. The distinctive-vocabulary section above is drawn from a reading sample ranked by reach and is used to characterise register only; no share or trend on this page is computed from it.
It reads six languages out of more than a hundred and forty carried by the index, and the same event ran in Hebrew, Russian, Urdu and others that are not shown here. It reads publication only; audience reaction is a separate layer and frequently points the other way. It reads eight weeks, while the history behind it runs back years. It reads channels in isolation, while the repost graph shows who originated a claim and who merely carried it, which is usually where the answer lives. It takes no position on the events and attributes nothing to any state, organisation or person. Attribution work is done under agreement, not on a public page.
One subject goes in - a claim, an actor, a location, a policy, a date. This is the shape of what comes back, with identifiers withheld here because this is a public page.
First results in days. The same question can be asked of any of the other languages in the index, and answered backwards across years of history.
This is six languages, eight weeks, read from the surface, produced without asking the data a single hard question. The index behind it spans the great majority of the public Telegram ecosystem: well over a hundred languages, every populated region, more than a decade of continuous collection, growing daily, with a large discovery queue behind it. The constraint on depth is not the data. It is the question, and the question is yours to set.
Method, limitations and reproducible queries are provided to institutional partners under agreement. Work under NDA is the norm, not the exception. Nothing is published about a client engagement. NarraPrism holds no political affiliation and works for no side.
Most of what is worth knowing about this fortnight is one question away from the surface you just read. If a line above sits closer to your remit than to your curiosity, that is the point.
Commission a cut on your question Put a subject under watch
Not ready to scope a project? Send one subject, one location or one claim and we will come back with a short read on what the environment holds on it. No commitment, no cost.
Cut reference IRAN-2026-09. Channel selection, queries, marker dictionaries and archive locations for every figure on this page are retained unchanged and supplied to institutional partners under agreement.