Pre-meeting steelman pair

Committees arrive at decision meetings holding a scored matrix and a recommendation. What they usually lack is the best argument against the thing they are about to approve. This job produces that argument in writing, before the room meets.

Each finalist gets two short documents. One makes the strongest honest case for funding it. The other makes the strongest honest case against. Both are written in earnest, with citations into the round's evidence.

The word steelman is doing real work here. A weak case against is worse than none, because it inoculates the committee. Members read a thin objection, dismiss it, and feel they have considered the downside. If the case against cannot be made strongly, that itself is worth reporting.

This is the cheapest useful thing at the decide stage. The evidence already exists; the pair reorganizes it into two adversarial readings. There is no new research, no new data collection, and no scoring change. An afternoon of drafting and an hour of review covers a full shortlist.

Load the finalist's assessment, the criteria, and the scoring detail. The pair must argue from what this round actually found. A case against written from general knowledge about that sector is a different document. It sounds plausible and it is not about this candidate.

Sub-criterion detail is more useful than a total. Where a candidate was scored separately on each dimension, as in scoring each sub-criterion independently before recombining, the pair can point at the specific weak dimension. A single composite score gives the argument nothing to hold onto.

Load the disagreement too, if there was any. A candidate where two assessors split is a candidate where both documents write themselves. A flag on high-variance candidates with the divergence described is the natural input here. The split already contains the two cases; the pair just writes them out properly.

Set a length limit and enforce it. Half a page each is enough, and it forces selection. Committee members who receive twelve pages will read the first two. Distribute at least a day ahead, and never in the meeting itself.

If the evidence file for a finalist is thin, say so instead of generating. A steelman pair built on two paragraphs of assessment is fiction with structure. That finalist needs more work, not better rhetoric.

Generate the case for first, in one pass, from the assessment. It is the easier document and it establishes the shared facts. Then generate the case against in a separate pass, with the case for as context. Sequencing matters because a single prompt asking for both produces a balanced essay.

A balanced essay is the failure, not the goal. Each document should read as though its author wants that outcome. Hedging inside the case for makes the case against redundant. Let them contradict each other on the page.

The case against needs one structural requirement. It must state what would have to be true for the concern to bite. "The delivery timeline is optimistic" is a mood. "This depends on a permit that took eighteen months on the last comparable project" is an argument.

Add a third short block if the committee will use it. For each finalist, the one fact that would change the reader's mind fastest. That block is what people actually remember from a pre-read.

Attribution rules keep the documents honest. Every substantive claim in either document cites the assessment section it came from. An uncited assertion should be deleted rather than softened. Softening produces a document that sounds balanced and cannot be checked.

This example is invented, including names and figures. A philanthropic foundation is choosing among three finalists in community air quality monitoring. Ruth Alderman is the program director and will make the recommendation. The committee meets Thursday; the pairs go out Tuesday morning.

The first finalist is a sensor network operator with four years of deployments. Its case for cites completed installations and a maintenance record from two cities. Its case against is narrow and specific. Every deployment so far used the same municipal partner. The assessment never tested whether the model transfers.

The second finalist is a university group with a stronger measurement method. The case for rests on published validation of the instrument. The case against names an assumption nobody had written down. The proposal needs data to reach residents, and the group has never run a public-facing service.

The third finalist is a community organization with the weakest technical assessment. Its case against is easy to write and mostly about capacity. Its case for is the one that surprises the committee. It is the only finalist with existing relationships in the neighborhoods the program targets.

On Thursday the discussion opens on transfer risk rather than on scores. Ruth recommends the sensor operator, conditioned on a second municipal partner in year one. The condition comes directly from the case against, which is the point.

Forced symmetry is the failure mode built into the format. Asked for a case against, a model will produce one for every finalist. Sometimes the honest answer is that the case against is weak. Allow the document to say that, and to be shorter than its partner.

Watch for objections that apply to every candidate equally. Execution risk, team depth, and funding uncertainty appear in all three cases against. Those are true and they do not discriminate. Mark generic objections separately from candidate-specific ones, or drop them.

The pair can also flatten a real difference. Two documents of equal length imply two comparable options. If one finalist is clearly stronger, the format hides it. Keep the scored assessment attached so nobody reads the pair as the whole picture.

There is a rhetorical risk worth naming. A well-written case against a strong candidate can talk a committee out of a good decision. That is why every objection has to name its condition. A condition can be checked; a well-turned sentence cannot.

The pair is not adversarial review. It works from the assessment as written and argues about the choice. A structured attack from multiple configured reviewers tries to break the analysis itself, earlier and at much higher cost. Do not use this cheap step as a substitute for that one.

The recommendation belongs to a named person, and the pair changes nothing about that. Ruth signs the recommendation and answers for it at the board. No document generated on Tuesday carries authority over Thursday's decision. The pair exists so that the person deciding has seen the other side in its strongest form.

Never let the format imply a verdict. Ordering finalists by how strong their case against looks is a ranking in disguise. Publish the pairs in a fixed order, alphabetical or by submission date. Small presentation choices move committees more than anyone admits.

Committee members are the intended audience, not the model. They should be able to disagree with a case against and say why in the room. A document nobody argues with was probably not strong enough.

Capture what the pair actually changed. When a case against becomes a funding condition, that path should be traceable afterward. Extracting the assumption each side rested on during the meeting preserves it. Otherwise the condition looks arbitrary a year later.

Confidence deserves recording alongside the choice. A committee that approved a finalist while finding the case against nearly persuasive should say so. A decision written as a distribution rather than a binary holds that nuance. A yes with a strong dissent is a different decision from an easy yes.