Methodology: How the Executive Board Funding Dialogue Readout Was Produced

A briefing built from one UN meeting's transcript. It records what each delegation put on the record, groups them by the position they took, and shows the evidence behind every grouping. It is not a UN document and carries no official standing.

The meeting ran three agenda items in one afternoon with a different cast at each: the UNDP funding dialogue with the UNCDF financial review, the UNFPA funding dialogue, and an update on how the Executive Board runs its own governance and oversight. Nineteen people spoke — six officials presenting reports, two co-chairs presenting a working-group report, and eleven delegations.

Sources for every number and quote on the readout:

• Transcript — transcripts/un-funding/2026-08-24_k17fex6nlu_attributed.md (local, regenerable from the portal)
• Analysis — readouts/un-funding-structured-dialogue-2026-08-source.md
• Scoring sheet — research/un-funding-structured-dialogue-position-scores-2026-08.md

The transcript came from the UN transcripts portal, which publishes ASR transcripts with speaker attribution attached. Collection is scripted:

python pipeline/fetch_un_transcripts.py fetch asset/k17/k17fex6nlu --silo un-funding
python pipeline/fetch_un_transcripts.py convert "transcripts/un-funding/2026-08-24_k17fex6nlu.json"

The portal supplies each statement's speaker as a structured object — country code, full name, function and any group the delegate speaks for. Attribution therefore comes from the source rather than from reading cues in the text, which is what lets the attribution gate below check the readout mechanically.

The JSON also carries the re-cut warning that shapes this whole page:

timestamps_flagged:          true
timestamps_note:             "UN Web TV re-edited this video after the transcript
                              was created; timestamps may no longer match the
                              current video."
original_duration_seconds:   12184
current_duration_seconds:    7250

Every meeting is classified before it is analysed, because the classification decides the question the readout asks. A formal meeting is about resolving contested positions, so divergence is the finding. An informal briefing is about floating ideas, so convergence is.

This meeting is recorded formal, on four grounds:

• The Executive Board is a governing body sitting in formal session.
• The President runs a speakers' list against a clock and asks delegations to keep interventions concise so more can be accommodated.
• Delegations read prepared national and group statements.
• Each of the three items closes with the President announcing that a draft decision is under preparation for the Board's consideration — the positions stated are the ones a decision will rest on.

Because the meeting is formal, the readout uses the two-rubric scoring described below. The classification is recorded in transcripts/un-funding/INDEX.md, not inferred while drafting.

The analysis codes positions only from what a delegation actually said. Nothing is inferred from bloc membership, from what a delegation said at other meetings, or from what it might be expected to argue. Where a delegation did not address a question, the readout records that rather than filling the gap.

The defining divide was derived from the transcript rather than assumed. Every speaker accepts the same diagnosis — aid down 23.1%, UNDP core at 9% of contributions, UNFPA core at 24% against a 30% target, UNCDF core at about $4 million. Nobody contests a figure. The disagreement is over the remedy, which is what the scoring measures.

Three anti-repetition rules govern where things go, so the page can be scanned rather than read end to end:

• A delegation is named at most once in the Fault Lines. Everything further about it lives in its reference card.
• Each fact or figure has exactly one prose home.
• Talking Points is the single verbatim home of the key quotes; body prose paraphrases and points there.

The horizontal position of every dot comes from a scored reading of what each delegation said. The method is in pipeline/position-scoring.md; the worked sheet with every score and quote is research/un-funding-structured-dialogue-position-scores-2026-08.md.

What is scored

Because this is a formal meeting, two mirrored rubrics of seven indicators are built, each summing to 100. A delegation scores on whichever side it argued, and a delegation that argued both scores on both.

• A–G — the case for restoring core funding. Donors should put in more predictable, unearmarked money, measured against the funding compact's 30% target.
• A′–G′ — the case for closing the gap from inside. Cut cost, reallocate what already exists, and raise revenue outside traditional aid.

The seven indicators

#	For restoring core (max)	For closing it from inside (max)
A	Core funding as the answer (20)	Counter-endorsement (20)
B	Defence of the funding model (10)	Performance critique (10)
C	Maximal ask (15)	Maximal counter-ask (15)
D	Specificity of the case (15)	Specificity of the counter-case (15)
E	Obligation framing (15)	Conditionality framing (15)
F	Material commitment (10)	Material counter-commitment (10)
G	Speech-act verb register (15)	Counter-pole verb register (15)

How the axis works

Write P for the total scored on A–G and N for the total scored on A′–G′. Then:

net        = P - N     which way the delegation argued, and how far off neutral
on record  = P + N     the total it put on the record, either way  <- the X axis

The X axis is on record, never net. Adding the two poles rather than netting them keeps a delegation that argued both ways visible: the Netherlands put 70 points on the record and would otherwise plot at 26, reading as though it had barely spoken. Direction lives on the Y axis and in the camp colour instead, which is why Chad (88 for) and the United States (78 against) share the right-hand column while sitting on opposite rows.

Columns are fixed thirds of the 0–100 scale — 0–39, 40–66, 67+ — so a column means the same thing here as on every other readout in the series. Within a cell, dots are ordered by on record ascending and spaced evenly by rank. Nothing is placed randomly.

What the score does not claim

The claim is ordinal. Read a score as "argued more than", never as a measurement of how right or how influential a position is. Scoring the counter-pole on its own 100-point rubric means a delegation arguing that the fix is internal can sit in the same column as one arguing donors must pay. That is a statement about how much each put on the record, not that the two positions carry equal weight in the room. The evidence panel keeps its bars scaled within each rubric rather than across them for the same reason.

A question is never counted as opposition

A delegation asking how something will work has not taken a position against it. Detailed questions usually mean the opposite — the delegation read the papers closely enough to find the gaps. The Netherlands asked two questions about how UNDP has adapted and what it is prioritising; those score nothing on the counter-rubric. Its 22 points come from one explicit sentence making further support contingent on understanding the reform.

Who is not scored

The six officials presenting reports and the two working-group co-chairs are unscored. They present the material the axis measures reactions to, so they cannot engage with their own proposal. They appear in their own band below the dashed line, with no horizontal position, because the axis counts how much was argued and they argued nothing — they presented.

No observers spoke. Every delegation on the floor was a member state or a member state speaking for a group, so the separate observer band the method provides for is not used here.

Speaking length

Word counts are computed mechanically from the transcript, never estimated:

python pipeline/word_counts.py transcripts/un-funding/2026-08-24_k17fex6nlu.json \
    --exclude "Executive Board / President,Executive Board Secretariat,unattributed,UNFPA /,UNDP /,UNCDF /,ERI / Co-Chair,DEU / Co-Chair" --md

Across the eleven delegations the median is 452 words and the brevity threshold (70% of the median) is 316. The shortest delegation, the Netherlands, spoke 321 words — five above the cut. No delegation carries the brief marker on this readout.

A readout normally ships only after its position-critical quotes are checked against the meeting audio. The check downloads the English audio channel, slices it at each quote's word-level timestamps, re-transcribes the slice with a second speech-recognition model, and compares the two readings. On earlier readouts this has caught real errors, including one where the ASR dropped a negation and reversed a delegation's meaning.

UN Web TV re-edited this recording after the transcript was made, cutting it from 3 h 23 m to 2 h 01 m. The transcript's timestamps describe the long version, so slicing the short version at those raw offsets would land roughly 17 minutes away from the intended moment and return coherent audio from elsewhere in the meeting — the dangerous failure mode, because it reads like a mismatch and invites a "correction" that would edit a correct quote into a wrong one.

The offset was derived empirically before any quote was sliced, per the method. A plan drawn up before probing considered the two internal gaps in the transcript (each over 60 s, both podium changeovers between agenda items) and asked whether the recording might have been re-cut at those points too, which would mean each of three segments needed its own offset. Six probes were placed to test that directly: two bracketing the whole recording, one deep inside the first segment, and — the two probes that mattered most — one immediately after each internal gap, guarding the two quotes structurally most exposed to a segment break.

All six converged on the same constant, offset by no more than about four seconds from each other — ordinary sentence-boundary rounding, not a segment break:

audio_time = transcript_time − 1021.0 s

The three-segment hypothesis was tested and rejected, not assumed away. The offset must never be taken from the difference between the two reported durations: here that difference is 4,934 s, nowhere near the true 1,021 s constant. On a previous re-cut meeting the same mistake would have sent every quote about four minutes wide.

Result: all 11 position-critical quotes checked, 10 matched the ASR text verbatim, one corrected. Tanzania's line about UNCDF's capital instrument reads "bridged bond" on stable, repeated audio — not "sub-regional bond", which is what the ASR transcript has. The readout's Talking Points blockquote and evidence panel both carry the corrected wording; the source analysis file logs the original ASR reading, the correction, and the full probe table. Full derivation: daily-runs/2026-08-25-unfunding/STAGE4-OFFSET.md (not published — internal working file).

Two scripted checks run against the finished files. Both look at everything rather than sampling, which is the point: a reviewer confirms what they happen to read, a gate confirms the whole page.

Attribution

python pipeline/verify_attribution.py \
    readouts/un-funding-structured-dialogue-2026-08.html \
    transcripts/un-funding/2026-08-24_k17fex6nlu.json

It takes every quote rendered on the page, finds the speaker the readout credits it to, and checks that the words appear in that speaker's own statements in the transcript. It also flags any country named on the page that never spoke.

Result: 10 of 11 quotes matched, against a roster of 22 speakers. The one flagged quote is expected: Tanzania's quote was corrected against the meeting audio (see Stage 4 above), so its wording no longer string-matches the ASR transcript this gate compares against — a known, documented limitation of a mechanical text match, reconciled by hand in the source analysis file rather than a defect in the readout. Six countries are named on the page without having spoken, and each is correct in context: Austria, China, Luxembourg and Thailand appear inside a quoted list of UNCDF's six contributors; Bangladesh appears as the location of a UNFPA financing agreement; and Spain appears as the delegation whose statement the shortened recording does not carry.

Scoring-sheet completeness

python pipeline/verify_scoring_sheet.py \
    research/un-funding-structured-dialogue-position-scores-2026-08.md \
    transcripts/un-funding/2026-08-24_k17fex6nlu.json

It fails a sheet for invented score ranges, deferred work, placeholder entries standing in for unscored delegations, and any speaker in the transcript the sheet never names. A delegation left unscored has to be named with a reason.

Result: clean. Every one of the 22 speakers is either scored or named as excluded with its reason.

Rendering checks

The plot was checked in a headless browser across all three Y-axis settings: no dot escapes its cell, no two dots overlap, no dot collides with the axis labels or column heads, every delegation keeps the same column whichever dimension is selected, and each dot's column matches its own arithmetic. All 131 coloured delegation names were checked for their computed colour, because a name can render grey through a CSS conflict while the markup looks correct — that check caught two real faults during the build.

• Audio verification is quote-level, not transcript-wide. The 11 position-critical quotes reproduced verbatim in Talking Points were checked against the meeting audio and match (one word corrected). The wider evidence base behind the position scores — every fragment quoted in the evidence panel and the scoring sheet — still reproduces the ASR transcript text unheard.
• Spain's statement is missing. It took the floor and its words are not in the record. It is shown as unknown rather than silent.
• ASR errors survive in quotes. The transcript renders some words imperfectly — "UNIB" and "UNIP" for UNDP, "NDP" for UNDP, "FNUREP" for UNFPA, "UNAT" and "UNAD" for the same reform initiative. Quotes are reproduced as the transcript has them rather than silently corrected, because correcting them by guesswork is how a paraphrase becomes a fabricated quotation.
• Scores are ordinal. They support "argued more than", not a measurement.
• Length affects scores. A longer statement has more chances to hit indicators on either rubric. Nepal spoke 1,425 words across three turns. This is disclosed rather than corrected, because correcting it would mean inferring what a delegation would have said.
• Scope affects scores. Three agenda items ran with different casts. Rwanda spoke only on governance; Finland and Côte d'Ivoire only on UNFPA. A delegation that never had the funding item in front of it cannot score on it, and a low total can mean scope rather than silence. The readout says which case applies for each delegation.
• One meeting only. Positions reflect what was said on 24 August 2026. A delegation may hold a different view elsewhere.
• Interpretation. Some delegations spoke through interpreters, so the English text is the interpreter's wording rather than the delegate's own.