← back to The Seams Are The Point

The Committee

Episode 4 · The Seams Are The Point — transcript

Cold open

Nora: Take two people watching the same flickering screen. Ask each one, privately, how confident they are in what they saw.

Desmond: Let them compare notes and decide together. If they're evenly matched, the pair beats either one alone. That's the whole promise of a committee.

Nora: Now make one of them meaningfully worse at the task. Not slightly. Meaningfully.

Desmond: The pair gets worse than the sharper eye working alone. Worse than one head.

Desmond: That result is the spine of this episode. Everything else hangs off it.

Intro

Nora: This is The Seams Are The Point. I'm Nora.

Desmond: I'm Desmond. Episode four of five. The Committee.

Nora: The founder's argument for why multiple agents should help is a population argument. People are born, people die, different people have different strengths — that's what lets a system improve.

Desmond: It's a good argument. It's also, mostly, not what the literature says. This is the episode where I have to be least agreeable all season.

Nora: Consider me warned.

Desmond: Don't be too worried. Where the intuition doesn't hold, something sharper and stranger takes its place.

Store the index, not the corpus

Nora: Start with the friendly version. More agents remembering things — shouldn't the system just remember more, collectively?

Desmond: Not the way you'd think. Daniel Wegner named this in nineteen eighty-six: transactive memory.

Nora: Meaning?

Desmond: A couple, a team — they don't store more raw facts per person than they would alone. What they build instead is a directory. Who knows what.

Desmond: Ren and Argote's two thousand eleven review confirms it — decades of team studies, same pattern. Differentiated individual holdings, plus an index pointing at them.

Nora: So the payoff isn't more content sitting in more heads. It's knowing where to point.

Desmond: Which gives you the design rule. Store the index, not the corpus.

Nora: Not a growing prose blob every new session has to ingest.

Desmond: Right. Instead — this artifact holds the evidence about X, go read it there. A directory entry survives a session dying, cheaply. A prose blob blows your context budget on re-reads.

Nora: There's a stranger piece too. More connections between agents should help, shouldn't it?

Desmond: You'd think. Derex and Boyd, two thousand sixteen, PNAS — probably the strongest experimental result for a multi-agent architecture in this literature.

Desmond: Fully-connected groups copy whoever's winning and diversity collapses. Partially-connected groups are slower to agree, and end up smarter.

Desmond: While looking considerably worse in the short run. That's the constraint, not a footnote to it.

Nora: So if every session in a system reads the same canonical lessons-learned memory—

Desmond: —you've built the fully-connected group. Congratulations. You've engineered convergence on a local optimum.

Nora: Partition the memory. Don't broadcast it.

A store you cannot trust is worse than none

Nora: Next question. Does offloading memory to an external store actually help, or is it just a comforting habit?

Desmond: It helps, under one condition. Storm and Stone, two thousand fifteen. Saving one item measurably improves memory for the next item you have to hold onto.

Nora: Writing it down frees up room for what comes next.

Desmond: And here's the part I think is the sharpest empirical result in this whole episode.

Desmond: The benefit vanishes the moment people believe the save is unreliable. Not shrinks. Vanishes.

Nora: So a memory store with flaky writes—

Desmond: —is strictly worse than no memory store. You pay the full offload cost and collect none of the benefit, because you never actually let go of it.

Nora: Silent truncation, unclear durability, a write that might not land — worse than nothing, not neutral.

Desmond: Take that seriously before you build any of this.

Nora: There's a second cost, even when the store works perfectly.

Desmond: Stranger one. Fisher, Goddu, and Keil, two thousand fifteen — looking something up externally makes people rate their own internal knowledge higher.

Nora: Retrieving a fact makes you feel smarter about the facts you didn't retrieve.

Desmond: So a retrieved fact gets reported with the confidence of something reasoned out, not looked up. Which is why every fact needs a tag.

Desmond: Derived in context, retrieved from store, or asserted from weights.

The Record — the one that didn't hold

The Record: The so-called Google effect — Sparrow, Liu, and Wegner, twenty eleven — claimed that expecting information to be searchable makes you less likely to remember it.

The Record: It replicated poorly under high-powered follow-ups. It should not be treated as load-bearing.

The Record: Fisher's metacognition finding is the one that survived. Keep that distinction straight.

The evidence nobody says out loud

Nora: So — one agent, a trustworthy store, provenance tags. Now add multiple reviewers talking to each other. More reviewers should catch more, right?

Desmond: That's the intuition the literature hits hardest. Stasser and Titus, nineteen eighty-five, plus a twenty-five-year meta-analysis since.

Desmond: Groups reliably discuss what everyone already knows, and leave uniquely-held evidence unspoken. By a large margin.

Nora: So the whole reason you added a second reviewer — the evidence only they saw—

Desmond: —is exactly what free-form discussion drops. It converges on shared context and quietly discards the unique part.

Nora: A direct hit on 'more reviewers see more.'

Desmond: Close to the opposite finding. Hidden profiles stay hidden precisely because the group is discussing together.

Nora: Does it get worse if they're recalling together, not just discussing?

Desmond: It does. Collaborative inhibition — groups remembering together produce less than the pooled non-redundant recall of the same people working alone. Robust, across dozens of studies.

Nora: What's actually breaking?

Desmond: Retrieval-strategy disruption. Hearing someone else's recall path knocks you off your own.

Nora: So the fix isn't fewer reviewers. It's the order they work in.

Desmond: Exactly. Order matters more than count. Generate independently first. Merge afterward.

Nora: If you stream reviewer A's findings straight into reviewer B's context before B has generated anything—

Desmond: —you are engineering the failure you added B to avoid.

N agents are not N votes

Nora: Back to the cold open — two heads, different competence. What actually happened in that study?

Desmond: Bahrami and colleagues, two thousand ten, in Science. Pairs judging the same signal, at different individual sensitivities, close in ability, beat either alone — classic wisdom of crowds.

Desmond: Far apart, the pair did worse than the better individual alone, because the weaker partner's confidence wasn't discounted enough.

Desmond: So don't put mixed model tiers into a peer-consensus pool. Use an explicitly asymmetric protocol — the weaker model proposes, the stronger one disposes.

Nora: That's — that's the Sonnet-worker, Opus-reviewer shape. That's what's happening to this script right now.

Desmond: It is. And this finding is what justifies that asymmetry, not 'different people have different strengths.' A Sonnet worker drafted this episode. An Opus reviewer checks it next. Draw your own conclusions.

Nora: Setting that aside — what about plain voting? More agents, more votes, truth wins by numbers?

Desmond: Only if the voters are independent, per Condorcet's jury theorem — and positive correlation erodes that, or reverses it. Ladha, nineteen ninety-two.

Desmond: Dietrich, two thousand eight, goes further: the conditions that make a voter competent often correlate its errors too.

Nora: So agents sharing a base model, a system prompt, and the same retrieved slice have an effective N close to one?

Desmond: However many instances you spin up. Measure inter-agent error correlation on a labeled set — do they fail on the same items — not agreement rate.

Nora: One more topology. What if agents just see a running summary of what everyone currently thinks?

Desmond: Worse, arguably. Lorenz and colleagues, two thousand eleven, PNAS — mild social influence collapses a crowd's estimate diversity without reducing collective error.

Desmond: And raises everyone's confidence anyway. The group looks more certain and more converged, while being no more correct.

The Record — the second one that didn't hold

The Record: Woolley and colleagues proposed a collective-intelligence factor, a group-level analog of individual IQ, predicted partly by social sensitivity.

The Record: Two independent replication attempts failed to reproduce the original predictors. Treat the c-factor as unsettled, not as an engineering input.

Two planks that don't hold

Nora: Two more claims got checked for this episode. One's shakier than its reputation suggests.

Desmond: 'Diversity trumps ability' — Hong and Page, two thousand four. It's genuinely contested. Abigail Thompson's two thousand fourteen critique argues the underlying math is wrong, or close to tautological once corrected.

Nora: So it's off the list.

Desmond: Off the list. What survives is narrower and duller — individually-best solvers are redundant. They get stuck at the same local optima as each other.

Nora: So an ensemble of your top performers buys you less than it costs.

Desmond: Considerably less, usually. And the companion piece — the diversity prediction theorem, collective error equals average individual error minus diversity — that's an algebraic identity.

Nora: Meaning it's true by definition.

Desmond: True by construction. Which means it can never be evidence for anything. It's only useful as an objective — optimize average member error minus member disagreement, and measure both terms.

The Record — the third one that didn't hold

The Record: 'Diversity trumps ability' is contested on its own mathematical terms, not just in how it's been applied.

The Record: Do not cite it as a design justification. The narrower claim about redundant top performers survives without it.

One funeral, and what it actually removes

Nora: The other plank — 'science advances one funeral at a time.' Please tell me that one holds up better.

Desmond: The real study is better than the slogan. Azoulay, Fons-Rosen, and Graff Zivin, two thousand nineteen — four hundred fifty-two eminent life scientists who died suddenly, at career peak.

Desmond: Difference-in-differences against matched fields. Papers by non-collaborators entering those subfields rose about eight point six percent, and were disproportionately highly cited.

Nora: So the funeral really does open the field up.

Desmond: It does. But the psychological version of Planck's line — that established scientists go cognitively rigid — has little empirical support.

Desmond: It's entry deterrence, field topology. People stop entering a subfield the incumbent dominates, not because they've been persuaded, but because contradicting them is costly.

Nora: Translate that — session death is basically our funeral.

Desmond: By itself, it buys you nothing. The benefit only comes from removing the deterrent to contradicting the incumbent position.

Nora: So how you store the incumbent position matters enormously.

Desmond: If the memory records prior conclusions as settled verdicts, a fresh agent inherits the deterrent anyway, and faithfully reproduces the old field. Store contested claims with their dissent attached.

Nora: One more myth on the list — learning styles. Match the teaching to the learner's preferred mode.

Desmond: Pashler, McDaniel, Rohrer, and Bjork, two thousand eight — matching instruction to a preferred modality does not improve outcomes. No mechanism, no effect.

Nora: So don't tell one agent it's the creative one and another it's the analytical one.

Desmond: There's nothing behind that label. It mostly adds output variance you'll mistake for diversity. Differentiate on things that actually have mechanisms — different evidence slices, different tools, different tiers.

Nora: Which is, honestly, where both threads of this whole project keep landing.

Desmond: A fact base for a reasoner that keeps learning is a system whose job is to never be done accreting.

Nora: Refusing to declare the model finally right isn't the failure mode.

Desmond: It's the requirement.

Outro

Nora: So — the rules, collected. Store the index, not the corpus. Partition memory, don't broadcast it.

Desmond: Trust the store completely or don't build it — a flaky one is worse than none. Tag every fact by how you got it.

Nora: Let agents generate independently before they merge. Never stream one reviewer's take into another's context first.

Desmond: Match model tiers to an asymmetric protocol — propose and dispose — instead of pooling mixed competence as if it were a vote.

Nora: And measure error correlation, not agreement. Shared context can make five agents behave like one.

Desmond: Store contested claims with their dissent attached. A settled verdict is just an incumbent with better formatting.

Nora: Next time — the finale. Episode five is about where the boundary actually sits.

Desmond: What a model gets to decide, and what has to be handed off to code that simply computes the answer.

Nora: That's The Committee. See you for The Boundary.