Methodology — How the Regional Dialogue Readout Was Produced

This page documents how the readout of the 3 August 2026 Regional Dialogue was produced, so a reader who knows the field can check the work rather than take it on faith.

The session was convened in Montevideo by Uruguay in its capacity as CELAC Pro Tempore President, together with GQUAL and CEPEI, and moderated by the journalist Carmen Aristegui. It is the second readout in the secretary-general track, following the UN Town Hall of 23 July 2026.

The meeting is classified as informal. It was convened to hear candidates set out a vision and to let the region put questions to them; nothing was being negotiated and no text was in play. Under this project's house format that classification decides the analytical frame: convergence is the finding rather than divergence, the fault-lines section becomes Open Questions, and the cluster map becomes an Engagement Plot. Negotiation vocabulary is kept off the page, because at a first presentation there is nothing to concede.

This project's pipeline is calibrated for chaired, sequential-statement formats. This session is not one, and it differs from the July town hall in a way that governs the whole analysis.

The consequence for the coding is direct: a prepared statement and a live answer under a 90-second clock are not comparable objects, so placing all five on one comparative axis would manufacture a comparison the record does not support. The readout therefore:

• reports and quotes both non-live candidates in full, and gives each a reference card;
• marks them with a "no questions" tag wherever they appear in a comparative table;
• renders them as ringed, paler dots on the plot, placed in the Nothing named row and not ranked on the horizontal axis;
• states explicitly that this is a gap in the record rather than a judgement about them.

A second structural difference from July: fourteen CELAC states intervened through foreign ministers, vice-ministers and permanent secretaries, seven of them putting direct questions to the candidates. They speak as states, not as candidates, so they are coded on a separate track, carry their own colour, and are filterable separately in the reference grid. The readout does not mix a national position with a personal programme.

The meeting was supplied directly as a portal URL and confirmed against the portal's own discovery index, which lists it under the General Assembly at 2026-08-03 with a transcript available.

python pipeline/fetch_un_transcripts.py fetch asset/k1z/k1zlaorvys --silo secretary-general
python pipeline/fetch_un_transcripts.py convert "transcripts/secretary-general/2026-08-03_k1zlaorvys.json"

The fetch produced 2026-08-03_k1zlaorvys.json and .txt; the convert step produced the attributed markdown the analysis consumes — 133 statements across 24 unique speakers, about 23,400 words.

The portal's ASR garbles proper nouns heavily in this session, and the same person appears under several spellings. Every name was normalised before analysis, because coding against an unnormalised transcript silently splits one speaker into several.

Venezuela's speaker is labelled in the transcript only by office, with no personal name given, so the readout attributes the intervention to Venezuela, Minister of Foreign Affairs and International Trade rather than guessing at a name. Guyana's intervention was delivered on behalf of Minister Hugh Todd by an official who is not named in the record; it is attributed to the state with that fact stated.

The analysis consumed the attributed markdown in full. Positions are taken only from what a speaker said on this record; nothing is inferred from bloc membership, from a candidate's institutional role, or from what they said at the July town hall.

The two coded dimensions

Both were derived from the transcript rather than imposed. The candidates converge on the goal — prevention and mediation must return to the centre of the office — so the analytically useful question is what each would actually do about it:

• What they would build. Build something new (Grinspan's good-offices office; Espinosa's early-warning and early-action hub) · Reorganise existing (Rodrigues-Birkett, who declines new initiatives on principle) · Nothing named (the two who took no questions).
• Use of regional institutions. How far each answer went on CELAC, CARICOM, Mercosur, the OAS and ECLAC, which Colombia and Paraguay both asked about directly.

The horizontal axis

The horizontal axis carries how concrete the answers were — whether a candidate named instruments, figures and mechanisms, or stated a direction of travel. It is an ordinal reading of this session only. It is explicitly not a score of merit, of support for any pole, or of the quality of a candidacy, and the readout says so in the plot's own guide box.

Since July 2026 this project scores within-camp position intensity with the evidence rubric in pipeline/position-scoring.md, and displays dots ordered by that score. That rubric was deliberately not run here, for three reasons that are all properties of this meeting rather than choices about it:

• There is no contested pole to score toward. The rubric measures intensity of position on a hearing's defining divide. This session converged; no divide of that kind exists in the record.
• The cast is below the rubric's differentiation floor. Three scored speakers cannot support the rubric's minimum of five positive indicators in a way that would mean anything.
• The format is out of scope. The rubric states that free-flowing and rotating-answer formats need adjustment before it applies.

The rule this project works to is that a within-cell ordering is part of the analytical claim: either it is evidence-based, or it must not be displayed. Rather than manufacture a score, the plot shows a defensible ordinal reading (concreteness) and labels it as exactly that.

Brevity marker

Word counts were taken mechanically from the attributed transcript before any scores were read: Espinosa 3,468 · Rodrigues-Birkett 3,311 · Grinspan 2,909. Median 3,311, so the 70% floor is 2,318 and no live candidate carries the brevity marker. Grossi's video runs 310 words and Bachelet has no turn of her own at all, since her letter is read inside the moderator's opening turn. The marker is suppressed for both: it only carries meaning where a roughly equal turn is the convention, and a video message and a read letter are not turns. Their brevity is a fact about the format and is stated as one on the readout.

Result: 15 of 15 matched, 1 ASR garble corrected

Run on 5 August 2026. Every position-critical quote was sliced from the meeting audio at its JSON word timestamp, re-transcribed locally with whisper large-v3-turbo, and compared against the ASR text.

The one correction. The ASR rendered Grinspan’s line as “estamos aquí porque traemos una experiencia” (we bring experience). Two independent whisper runs at different window widths both returned “tenemos” (we have). The reading was stable across runs and the audio is clear floor Spanish, so the readout was corrected to the audio and the English rendering changed from “we bring experience” to “we have experience”. The sense of the quote — experience rather than entitlement — is unchanged.

Pre-check before the audio pass

Every quote was first grepped against the attributed transcript and confirmed to appear exactly once, with whitespace and curly quotes normalised. Quotes were taken from the transcript itself, never from the intermediate source-analysis document, because a paraphrase in an analysis file can read plausibly and reach the page as a fabrication with a clean surface.

Translation

Most of the session was conducted in Spanish. The readout gives an English rendering with the Spanish original beneath it, so a reader who works in Spanish can check the translation directly. These renderings were made for this readout and are not the words of the live Zoom interpreter, which the portal does not transcribe.

This readout is filed in the secretary-general track, which is thematic: it is grouped by subject — the selection and appointment of the next Secretary-General — and fans across the General Assembly, the Security Council, PGA-convened informal dialogues and regionally convened events like this one. It is not grouped by speaking body.

Filing this session moved the track's time-series intent from planned to active. The 2026 selection is a bounded, connected process: the candidates' informal dialogues in April, the July town hall, this regional dialogue, Security Council straw polls from 30 July, a recommendation anticipated in September, and appointment later in the year. When a Secretary-General is appointed the cluster closes and the track reverts to retrieval-only until the next cycle.

• ASR source, not an official record. The portal transcript is machine-generated. Proper nouns are garbled throughout and were normalised by hand (see Stage 2).
• Quotes are audio-verified; the surrounding text is not. The 15 position-critical quotes were checked against the recording on 5 August 2026. Names, figures and non-quoted prose elsewhere on the page still rest on the ASR.
• Translation risk. English renderings of Spanish statements are this readout's own. A translation is an interpretation, and readers working from the Spanish original may reasonably render a nuance differently.
• Two candidates are under-represented by the format. Grossi and Bachelet took no questions. Nothing on this page should be read as evidence about how they would have answered.
• The concreteness axis is ordinal and session-bound. It supports "more concrete than" within this meeting only. It is not a score, and it does not transfer across readouts.
• Length bias. Speakers with more time have more opportunity to name specifics. Word counts are disclosed in full rather than corrected, because correcting would mean inferring what someone would have said.
• Candidate positions are personal programmes. They are not the positions of the states that nominated them. The ministerial interventions are state positions on the selection process and are coded separately.
• One session, not a trend. Everything here reflects the 3 August 2026 dialogue alone.