Decision capture + disagreement extraction

Decision meetings produce two things worth keeping. The decision itself, with who made it and on what basis. And the structure of the disagreement that preceded it. Most committees keep the first and lose the second within a week.

The second one is where the value sits. A committee rarely disagrees about the evidence in front of it. It disagrees about an assumption nobody stated out loud. Extracting that assumption converts an argument into a question somebody can answer.

So the output has three parts. A decision statement with a named owner and a date. A list of positions, each attributed and quoted from the transcript. And per disagreement, the assumption each side rests on plus the evidence that would resolve it.

Nothing in this pipeline recommends anything. It reads what was said and organizes it. A system that summarizes a meeting into a recommendation has invented a participant. The people in the room were the participants.

Write the decision statement before the meeting, as a sentence with an owner attached. "Whether to fund program A or program B this cycle, decided by the chair." A meeting with no stated decision produces a transcript and no record. Extraction cannot recover a decision that was never framed.

Recording consent is not a formality in this setting. People argue differently when they know a model will parse what they said. Say who reads the transcript and how long it is kept. Say whether attribution stays in the final record. A committee that discovers the terms afterward will stop speaking freely.

Load the pre-read material as context. The extraction needs to distinguish a position grounded in the brief from one grounded in experience. Both are legitimate; they resolve differently. One is settled by rereading page four, the other by asking a different question entirely.

Divergence known in advance should be on the agenda, not discovered live. A brief on high-variance candidates and where scorers split tells the chair which items need real discussion time. The rest can be confirmed in two minutes.

If consent is absent, run nothing. Notes written by a person are a worse record and a legitimate one. A covertly parsed meeting is not a governance improvement.

Run extraction in two passes and keep them separate. The first pass is mechanical. Who said what, quoted, with a timestamp and a speaker label. No interpretation at this stage.

The second pass is the hard one. For each position, ask what has to be true for that position to make sense. The answer is usually a belief about the future, a comparison, or a constraint. Write it as a claim the speaker would recognize and accept as theirs.

Then check whether two positions actually conflict. Often they do not. Two people optimizing for different time horizons are not in conflict about evidence. They are in conflict about scope, and that is the chair's problem, not the analyst's.

For each real conflict, name the resolving evidence. This field should be concrete: a measurement, a document, a person to ask. "More analysis" is not resolving evidence. If no evidence could resolve it, say so, because that is a values disagreement.

Values disagreements are common and should be labeled clearly. Whether the fund prioritizes near-term deployment or long-horizon capability is not an empirical question. Sending it back for more research wastes a cycle and annoys everyone. Escalate it to whoever sets the mandate.

This scenario is invented, including the names and numbers. A corporate research council is choosing between two battery chemistry programs for the next funding cycle. Hannah Voss chairs the council and owns the decision. The meeting runs ninety minutes with six participants.

Extraction returns fourteen positions across the two programs. Nine of them are agreement and take up half the transcript. Five carry actual disagreement, and they collapse into two underlying assumptions.

The first assumption concerns manufacturability. Three participants believe the sodium-based program can reach pilot scale on existing lines. Two believe it needs new equipment, based on a supplier conversation one of them had in March. The resolving evidence is specific: a line assessment from the manufacturing team. Nobody had asked for one.

The second assumption is about timing, and it does not resolve the same way. One side reads the cycle as needing a result within three years. The other reads it as building capability for the decade. No measurement settles that, because it is a question about what the council is for.

The extraction labels it a mandate question. It routes to Hannah rather than to the research team. She commissions the line assessment and defers the decision by four weeks. Both assumptions go into the register under her name.

Models asked to find disagreement will find disagreement. Two people phrasing the same point differently become a split. A hedge becomes an objection. Require a direct quote for every extracted position and let a participant reject the attribution.

The opposite failure is quieter and worse. Meeting summaries smooth conflict because summaries are written to be readable. The one dissenting voice becomes "some concerns were raised" and then disappears. Preserve dissent verbatim, attributed, and near the decision it dissented from.

Transcription failures distort attribution in a specific way. In a fast exchange, speaker labels drift, and a position gets assigned to the wrong person. That is more damaging than a garbled word. Where diarization confidence is low, mark the speaker unknown instead of guessing.

Silence is not agreement, and the extraction cannot tell the difference. Two participants who said nothing did not endorse anything. Record who spoke on each item, so a reader can see the coverage. A decision supported by three of six voices, with three silent, is worth knowing about.

The last risk is chilling the room. A committee that expects its arguments to be parsed and stored will pre-negotiate in the hallway. The transcript then records a clean meeting and none of the reasoning. Keep the record narrow, keep access limited, and never use it for performance assessment.

The decision belongs to a person with a name, and the record says so. In this workflow that is the chair, or whoever the governance document names. The extraction pipeline has no standing to decide, recommend, or break a tie. It reports what was said and what would settle it.

That boundary is easy to erode without noticing. A tool that ranks positions by strength of argument is making a judgment. A tool that presents one assumption as the crux is making a smaller one. Present the assumptions as a list and let the chair say which is load-bearing.

Dissent stays in the record with the dissenter's name, if they agree to that. A decision recorded as unanimous when it was not will be corrected later, badly. Participants should be able to read and amend their attributed positions before the record closes.

Send the assumptions somewhere they can be tested. An assumption about a future outcome fits a registered forecast scored when it resolves. One about how confident the committee actually was fits a decision recorded as a distribution.

Store the record where a future round will find it. The most useful thing a committee can hand its successor is why the last decision was close. That only survives in a searchable store of decision artifacts, not in a chair's notebook. Steelman documents from the case for and against each finalist belong in the same folder.