Takeaways: Lessons learned during the session

The Closed Loop

     Nassim Taleb describes a turkey fed every day for a thousand days. Each feeding confirms the pattern: the farmer is reliable, the system works, everything is fine. The turkey's confidence in its model grows with every confirmation. On the thousand-and-first day, the pattern breaks in a way the turkey's model had no variable for. The fatal error was structural: it had no mechanism to detect that the farmer's interests and its own had never been the same.

     This is the entry point the session kept returning to. Not bad intent, not incompetence, rather a structural gap between the person making decisions and the person bearing their consequences. When that gap exists, the system produces confident, consistent, increasingly wrong outcomes. And the longer it runs without correction, the more legitimate it appears from the inside.

The Show and the Restaurant

     The Michelin Guide works. That is the point most critiques miss. It optimises for a specific, clearly defined experience: technically precise, consistently reproducible, calibrated to the expectations of a professional evaluator who has eaten at thousands of restaurants across decades. Guests who arrive knowing what they are buying receive exactly what the star promises. As a quality standard for a specific kind of culinary performance, it is genuinely effective.

     In September 2017, Sébastien Bras, chef at Le Suquet in Laguiole, holder of three Michelin stars for eighteen years, contacted the guide and asked to be removed. Not because the restaurant was failing. Because he had realised he had stopped cooking for the people in the room and started cooking for the inspector. The show had replaced the restaurant. Michelin honoured his request and omitted Le Suquet from the 2018 guide. Then in January 2019, without his consent and without explanation, it re-entered him with two stars. His response, published in a statement to AFP: he was no longer concerned with the stars or the strategies of the guide. The system had no mechanism to let him leave.

     The question this raised was not whether Michelin is good or bad. It is more precise than that: good for whom, and measured by what? The inspector bears no consequence for what the rating costs the chef who holds it. The feedback loop ran in one direction only and the person bearing the full weight of that direction had no voice in it.

The expert system doesn't fail because experts are wrong. It fails because expertise without consequence creates a closed loop: self-validating, internally coherent, and increasingly disconnected from the human experience it was built to serve.

The Mass Review Trap

     The natural response to the opacity of expert evaluation is to democratise it. If the anonymous inspector cannot be trusted, surely the aggregate of thousands of genuine diners can be. Google Reviews appears to solve the Michelin problem: anyone who ate there can judge, the score reflects a broader reality, the expert is replaced by the crowd. The anticipation is understandable. The result follows exactly the pattern Elinor Ostrom predicted.

     Her research on collective governance identified the conditions under which shared judgment produces reliable signal: bounded membership, continuous participation, and consequence that flows back to the person making the judgment. An anonymous reviewer who leaves a one-star rating because the waiter was slow bears zero consequence for what that does to the restaurant's livelihood. The signal looks democratic. It is noise dressed as data. Campbell's Law completes the picture: the moment the metric becomes the target, rational actors optimise for the metric rather than the thing it was measuring.

     The same failure appears across industries. Hospital ratings aggregated from patient surveys produce institutions that optimise for survey scores rather than clinical outcomes. It discharges timing, bedside manner, car park accessibility ranking alongside treatment quality in the final number. University league tables built from graduate salary data produce departments that optimise for graduate employment rather than intellectual development. The mass review, across every domain, suffers from the same structural flaw: the reviewer disappears after clicking submit and bears no consequence for being wrong.

The Signal That Was There All Along

     The honest evaluation signal is neither the anonymous inspector nor the aggregate score. It is the composition of the dining room on a quiet Tuesday. A restaurant that fills with returning locals has already passed the only evaluation that carries genuine consequence — people chose to come back, with their own money, on an ordinary night, when nobody was watching. That signal is almost never represented in formal reviews, which skew toward occasion dining and tourist visits. The Michelin inspector eats once. The regular has eaten forty times and will eat forty more. Their judgment costs them something if they recommend poorly. That is what makes it worth trusting.

     The same principle operates in every domain where evaluation matters. The software tool that the development team adopts voluntarily and recommends to colleagues. The consultant whose previous clients take their calls without being asked. The hospital that attracts patients who have already been treated elsewhere and chose to return. These signals are harder to aggregate, impossible to game, and almost entirely ignored by formal evaluation systems — precisely because they resist the kind of legibility that makes metrics publishable.

     Identifying the honest signal is the easier half of the problem. The harder half is designing systems that produce it structurally, where the people making consequential decisions are kept close enough to the outcome that the signal reaches them naturally, rather than having to be extracted through evaluation frameworks that are themselves gameable.

Designing for Consequence

     In ancient China, physicians were paid a retainer to keep their patients well. When a patient fell ill, payment stopped until health was restored. The incentive structure aligned the doctor's interest with the patient's outcome rather than with the volume of treatment delivered. Prevention became more valuable than intervention. The doctor who would feel the consequence of their patient's decline paid a different quality of attention than the doctor paid regardless of outcome. This is documented in historical records of the period and noted as a striking contrast to Western medical practice as early as the nineteenth century.

     Someone in the group drew the direct corporate parallel, and it was exact. A software vendor contracted to build and ship a system is paid on delivery. Their incentive is to produce something that passes acceptance criteria on time and within budget. The internal team that inherits the codebase will live inside it for years: debugging under pressure, extending it in conditions the original designers never modelled, absorbing every architectural shortcut made by someone who would not be present when the consequences arrived. The vendor optimises for the handover. The internal developer optimises for the long term. The system that produces poor outcomes is the one that gives the vendor full design authority without requiring them to maintain any relationship with what they built.

     What the session pushed toward was the two-directional version of this principle. The Chinese doctor is accountable to the patient, but the patient is also accountable to the doctor's advice. The internal developer owns the consequence of the architecture, but the vendor who built it should remain reachable when the assumptions prove wrong. The omakase chef reads the diner across twelve courses and remembers them next time — but the diner who treats the counter disrespectfully finds the next reservation unavailable. Consequence flowing in one direction produces the Michelin inspector. Consequence flowing in both directions produces a system that can actually learn.

Whoever bears the consequence must hold the authority. But whoever holds the authority must also be able to see clearly. Proximity, taken too far, is its own form of blindness.

The Golden Path

     Medical ethics prohibits surgeons from operating on their own family members because emotional investment beyond a certain threshold impairs the clinical judgment that makes the skill useful. The surgeon who loves the patient too much cannot make the dispassionate assessment the patient needs. They second-guess the difficult call. They avoid the intervention that is right because it is too hard to make. Consequence, taken to its extreme, becomes its own form of blindness. Too little consequence produces the detached inspector. Too much produces the surgeon who cannot hold the scalpel steady.

     The productive zone is professional investment with enough distance to see clearly: enough skin in the game to pay genuine attention, enough separation to maintain objectivity. In organisational terms, this means internal ownership of consequence combined with external challenge of assumptions. The internal developer who will maintain the system should have design authority. The external perspective that is not emotionally attached to any particular solution should have a formal seat at the table. Functioning as a deliberate structure that compensates for the specific blindness each position produces alone.

     The session ended where it had begun, with the question of what it takes to keep the feedback loop honest. The answer is not more evaluation, more data, or more democratic access to judgment. It is ensuring that the people making consequential decisions remain connected to the consequences of those decisions — not so remotely that the signal never reaches them, and not so personally that emotion overrides judgment. A dish made from excellent ingredients, assembled by someone who will also eat what they made, and honest about what the result actually tastes like. That is the model. Everything else is a version of the closed loop.

More Personal, Less Connected

     This session felt different in two ways: one encouraging, one worth sitting with honestly. The encouraging part: people shared personal examples more readily than in previous sessions. The opening question: when did you experience optimization destroying the core thing landed directly. Someone spoke about cooking, describing how combining individually excellent ingredients without understanding their relationships produced something worse than the sum of its parts. That observation opened the room in a way that abstract theory rarely does. The personal example arrived first, and the theory grew around it rather than the other way around. That is the order that works.

     Nonetheless, something in the group dynamic felt slightly more distant than usual. It is hard to name precisely. People were present, engaged, intellectually generous. But there was a register of warmth that was not quite there in the way it has been before. Whether that was the composition of the room, the topic's corporate framing, or something in how the session was facilitated is not yet clear. It is worth acknowledging rather than explaining away.

When the Example Arrives Before the Theory

     The cooking observation that opened the session illustrated something Amy Edmondson's research makes precise: personal, concrete examples create psychological safety in a way that abstract frameworks cannot. When someone shares a vulnerable moment, the group doesn't just listen. It begins to recognize the pattern in its own experience.

     What worked this session was that the personal example was genuinely first. Previous sessions have struggled with the register: the intellectual tone setting in before the personal one has a chance to establish itself, leaving personal disclosure feeling like a departure from the implicit contract rather than the point of the exercise. This time the example arrived before the contract was written, which meant the contract was written around it. Carl Rogers called this the difference between intellectual understanding and felt sense: theory that has been confirmed by personal recognition lands differently than theory that has merely been understood.

What the Distance Might Mean

     The slight distance in the group is worth naming without over-interpreting. Edgar Schein's work on group dynamics suggests that warmth is a signal about psychological safety, about whether people feel genuinely seen rather than merely included. A group can be intellectually engaged and personally guarded simultaneously. The two registers require different conditions, and those conditions do not always arrive together.

     One hypothesis: the topic itself carried a corporate framing that may have oriented people toward professional rather than personal territory. The sessions that have produced the most warmth have been ones where the distance between the illustrative example and the personal experience is short enough that people feel they are talking about their own lives rather than about a case study. That gap may have been slightly wider this time than it needed to be. Something to adjust at the opening next session: a question that places the person directly inside the experience rather than observing it from the outside.