Methodology: How the UNDP Funding 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 is the UNDP segment of the Board's second regular session, continued from the previous day: an interactive dialogue with the Administrator. Alexander de Croo made one opening statement, 36 delegations responded, and he replied once, halfway through the speakers' list.

Sources for every number and quote on the readout:

• Transcript: transcripts/un-funding/2026-08-26_k18wfgb8jy.json
• Scoring sheet: research/un-funding-undp-position-scores-2026-08.md
• Audio verification: daily-runs/2026-08-27-unfunding-undp/STAGE4-REPORT.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/k18/k18wfgb8jy --silo un-funding
python pipeline/fetch_un_transcripts.py convert "transcripts/un-funding/2026-08-26_k18wfgb8jy.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 transcript holds 79 statements from 39 distinct speakers, including the President in the chair, across 3 h 02 m.

Unlike the two previous meetings in this track, this entry carries no re-cut warning. The 24 and 25 August transcripts both set timestamps_flagged: true, because UN Web TV had shortened the video after the transcript was made. This one does not, and the durations agree: the Kaltura entry, the downloaded audio and the last statement in the transcript are all consistent with a 10,919-second recording. Word timestamps therefore point at the published audio directly.

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 and takes delegations in order before inviting the Administrator to respond.
• Delegations read prepared national and group statements.
• One delegation states a position against the proposal on the record. The United States argues that the constraint is UNDP's own structure rather than the level of contributions, and asks the organisation to restructure.

The word "dialogue" in the agenda item does not settle it. An interactive dialogue could reasonably read as informal, and if it had produced only questions and design concerns it would have been classified that way. It is recorded as formal because a stated position against the proposal is present, which is the test the method sets. Because the meeting is formal, the two-rubric scoring described below applies. The classification is recorded in transcripts/un-funding/INDEX.md, not inferred while drafting.

Every position in the readout is taken verbatim from the transcript. Nothing is inferred from a delegation's bloc, its known policy, or what it said at another meeting.

The defining divide is identified first, then everything else is organised around it. Here it is who has to move to lift the funding constraint — the contributors, or UNDP itself.

That framing was chosen over a simpler "do you support the new development model" axis because the transcript does not support the simpler one. Nobody argued against the model. The United States endorses it explicitly, quoting the Administrator's own words back at him, and still refuses the funding conclusion he draws from it. What separates delegations is where they place the requirement, not whether they like the proposal.

Two rules from the project's house style bind the prose:

• No ranking of delegations. The readout reports what each delegation said; it never grades whose argument was strongest, hardest or most demanding.
• Plain language. Domain terms are explained where they first appear. Quotes are exempt from any length or vocabulary constraint — trimming a delegate's words to fit a target would falsify the record.

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-undp-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 binding constraint is core funding, and contributors must move. UNDP's model can only be fixed by restoring adequate, predictable, flexible core resources. Efficiency helps but cannot substitute.
• A′–G′ — the binding constraint is UNDP's own structure. The organisation must restructure, cut overhead and narrow its programme before, or instead of, asking for restored core.

The seven indicators

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. This is why the United States sits in the right-hand column at 91 rather than at the left edge: it argued a full case, and the axis counts argument regardless of direction. Direction lives on the Y axis and in the camp colour instead.

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 and spaced evenly by rank. Nothing is placed randomly.

The split is 10 delegations in the low band, 19 in the middle and 8 at the top. A middle band holding more than half the room is a finding about this meeting rather than a reason to move the cut points: most delegations made the core-funding argument in a similar register and at a similar length, and the spread sits at the two ends.

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.

The clearest illustration is on this readout. The three delegations at the top of the scale — the Dominican Republic at 86, Rwanda at 79 and Uruguay at 78 — are programme countries or programme-country groups. The two delegations that actually committed core money, the Netherlands and Germany, score below them at 70 and 68. This is not an error in the scoring and it was not adjusted. A state that has already contributed rarely also demands that others contribute, so it scores the material indicator and stops; a state making the ask scores the ask indicators instead. The axis measures how much was argued, not who paid.

A question is never counted as opposition

A delegation asking how something will work has not taken a position against it. Rwanda, Côte d'Ivoire, Norway, Brazil, the Netherlands, Nauru and Papua New Guinea all put operational questions to the Administrator, and none scores on the counter-rubric for them.

A demand for discipline is not opposition either, when it accompanies the core case. The Netherlands, Norway and the United Kingdom all press UNDP to prioritise and make hard choices. Each does so while arguing that core must be restored — the Netherlands announces a €40 million pledge in the same statement. Only a statement that places the constraint on UNDP instead of on contributors scores the counter-rubric. In this meeting that is the United States alone.

Who is not scored

The Administrator is unscored. He authored the proposal the axis measures reactions to, so he cannot engage with his own case. He appears in his own band below the dashed line, with no horizontal position, because the axis counts how much was argued and he argued nothing — he presented. The President is procedural and is named as excluded in the scoring sheet.

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/2026-08-26_k18wfgb8jy.json \
    --exclude "President" --md

Across the scored delegations the median is 385 words and the brevity threshold (70% of the median) is 269.5. No delegation falls below it. The shortest statement is Japan's at 304 words and the longest is Nepal's at 726, delivered for the least developed countries. No brief marker appears on this readout, and the absence is worth stating: every delegation that reached the floor used comparable time.

An axis was rebuilt after a headless check

The second Y dimension originally carried five values — the level of core funding, presence on the ground, which choices UNDP makes, making the capital model work, and UNDP's structure. The plot grid holds three rows, so a browser check found it rendering only 10 of the 38 dots and silently dropping the rest. The three smaller groups were merged into one row, "how UNDP works and what it spends on", which restored all 38 dots and produced a legible cluster: the Netherlands and the United States share its right-hand cell, both arguing at length about how the organisation should work, one while pledging €40 million and the other while naming no contribution at all.

This track carries a short list of tracked figures: specific numbers or claims from an earlier meeting that reframed what a debate was about, and that a reader would want to know the answer to going into the next session. The list lives in the track's transcript index and is read before drafting each new readout in the track.

For each open figure, the new transcript is searched for a reaction at three levels, loosest last:

1. Direct citation — the figure's own numbers, or a paraphrase close enough to be unambiguous.
2. Concept engagement without the numbers — a delegation argues the same question the figure raised, without repeating it.
3. Loose topical mention — a statement touching the same ground without engaging the point.

A result is recorded every round, including a clean miss. "Nobody engaged it this round" is itself the finding, and it only means something if it is checked and stated the same way each time.

This round produced one of each, which is what the three levels exist to separate:

• Figure 1 — answered at level two. Nobody repeated the UNDP budget numbers from 24 August. Two speakers argued the question those numbers raised: the United States asked that allocation formulas be modernised, and the Administrator opened the formula himself and gave his own discretionary residual. The figure stays open, because the Board discussion he proposed has not happened.
• Figure 2 — no reaction at any level. Three delegations pressed for the UNDP–UNOPS assessment to proceed. None asked whether the divergent numbers the agencies conceded on 25 August had been reconciled.

A search hit is not the same as a reaction. A text search for the liquidity threshold returned the Dominican Republic saying "Two months ago, in June of 2026" — a date, not the funding figure. It is recorded as a false positive rather than counted, which is the reason the levels are adjudicated by reading rather than by grep alone.

Full method: pipeline/anchor-tracking.md.

The project's standard is that every quote in a readout is checked against the meeting audio: the video is downloaded, sliced at the transcript's own word timestamps, re-transcribed with a second speech-recognition model, and compared. That check ran on 27 August 2026. All 25 quotes matched the recording. No corrections were made.

No offset was needed, and that was checked rather than assumed. Both previous meetings in this track required one, because UN Web TV had re-cut the video after the transcript was made. The pre-flight check here found no re-cut flag, and three independent figures agree on a 10,919-second recording: the Kaltura entry duration, the downloaded audio, and the last statement's timestamp at 10,886 s. Slices were therefore taken at the JSON timestamps directly.

Which audio channel. The entry carries ten flavors. The English one was selected on its declared language, never on a remembered identifier — that is the channel the portal ASR transcribes, carrying floor audio for English speakers and interpreter audio for everyone else. The raw floor channel was not used, because five of the scored delegations spoke through interpretation and their words would not match the English transcript.

The result, and the one reading that looked like a correction. Nineteen quotes matched on a first-pass slice. Six needed wider or re-anchored windows, and every one of those failures was the window rather than the wording: a quote spanning several sentences needs a slice anchored on its last fragment as well as its first, or it ends before the quoted passage does.

One quote is worth naming. On a wide window, Norway's earmarking argument appeared to read "becomes too strong, and the preference of donors risks weighing more", against the readout's "then the preference". A tight re-slice of the same passage returned "then", matching both the readout and the transcript. The wider window had smoothed the word. This is the second run in this track where a wide-window reading looked like a correction and a tighter slice confirmed the original, so a single wide-window disagreement is now treated as a prompt to re-slice rather than as a finding.

Full derivation and the per-quote table: daily-runs/2026-08-27-unfunding-undp/STAGE4-REPORT.md. Method: quote-verification.md.

Attribution

Every quote rendered in the readout is matched back to the transcript and checked against the speaker object the portal supplied. This catches a quote attributed to the wrong delegation, a quote that does not appear in the transcript at all, and an actor who never spoke.

python pipeline/verify_attribution.py \
    readouts/un-funding-undp-2026-08.html \
    transcripts/un-funding/2026-08-26_k18wfgb8jy.json

Result: clean. 25 attributed quotes checked against 39 speakers, all matched. Six warnings name countries that appear in the prose but never spoke — Fiji, Guatemala, Lebanon, Palestine, Serbia and Syria — each cited as an example of UNDP work rather than presented as a participant. All six were read and confirmed correct.

Scoring-sheet completeness

The scoring sheet is checked for invented score ranges, deferred work, and any transcript speaker the sheet never names. An actor left out must be named with a reason rather than covered by a group sentence.

python pipeline/verify_scoring_sheet.py \
    research/un-funding-undp-position-scores-2026-08.md \
    transcripts/un-funding/2026-08-26_k18wfgb8jy.json

Result: clean. Every one of the 39 speakers is scored, banded as a presenter, or named as excluded with a reason.

Arithmetic and evidence

Two checks were run over the scoring sheet and the plot data together, because a readout can be internally consistent and still disagree with its own sheet:

• Every total was re-summed from its own indicator rows. Three drafting errors were found and corrected against the rows, never the reverse: Nepal 66 → 56, Thailand 61 → 51, Pakistan 52 → 48. After correction, P, N, net and on record reconcile for all 37 scored delegations and no indicator exceeds its cap.
• Every evidence quote was checked against its own speaker's turn — 305 checks in total, each required to appear exactly once in the transcript of the delegation it is attributed to. Two quotes in the sheet turned out to be near-paraphrases and were replaced with the verbatim wording; one register fragment read "must eliminate" where the speaker said "must favor opportunity for all and eliminate", and was corrected.

Rendering checks

Run in a headless browser against the built page:

• All 38 dots render on both Y settings, with no overlapping dots, nothing escaping the plot area, and no collision with the column or row headings. The first run of this check found the second axis rendering only 10 dots, which is what prompted the axis rebuild described above.
• 172 camp-coloured names were checked against their computed colour; all matched their camp. An earlier build had 13 dots falling back to grey because a camp name had no colour defined — caught here, and fixed by aligning the plot's camp names with the ones the prose already used.
• Edit mode enters cleanly, exposes 198 editable regions and 10 section controls for 10 sections, and leaks no editable node into the plot, toolbar, legend, evidence panel, glance table or axis explainer. Removing a section and undoing it leaves all 38 plot dots live and the evidence panel still wired.
• The page opens in light theme regardless of the reader's system setting, and does not scroll horizontally at desktop width.

• The source is ASR, not an official record. The UN publishes these transcripts as a preview tool. Wording in any quotation may differ from the official record when it appears.
• The meeting was cut short, and about ten delegations never spoke. They are absent from this analysis because they did not reach the floor, not because they had nothing to say. The Administrator's second reply round never happened, so questions asked after his single reply went unanswered in the room rather than being declined.
• Papua New Guinea's statement was truncated by its own account, with the full text to be submitted afterwards. Its score reflects what was said in the room, not what was written.
• Scores are ordinal, not measurements. They say how much a delegation put on the record, not whether it was right, and not how much influence it carries in the room.
• One delegation carries the counter-rubric. A two-pole plot with 36 delegations on one side and one on the other is an accurate picture of this meeting, and must not be read as a balanced split.
• Longer statements have more chances to score. Group statements carry group mandates and cover more ground, which lifts the specificity and obligation indicators relative to a national statement of the same conviction. This is flagged rather than corrected.
• Five delegations spoke through interpretation. The register indicators — obligation framing and verb choice — are the ones most exposed to an interpreter's word choice. The content indicators are not.
• One meeting only. Positions reflect what was said on 26 August 2026. A delegation silent on a point here may hold a clear position on it elsewhere; silence is never read as a position. A score does not carry across meetings either: France scored 82 on the 25 August merger question and 22 here, against a different divide.
• Figures are as stated by speakers. The funding numbers on the readout are what delegations and the Administrator said in the room, not figures checked against the underlying reports. Where two speakers give different numbers for the same thing, both are shown and neither is reconciled, because neither stated a basis for the difference.