Headless

The world

The environment has a set of dimensions that drift over time. Most are ordinary: they move steadily and matter moderately. A minority are rare and high-impact: usually near zero, then occasionally hit by a large shock that fades over a few dozen ticks. Each tick the organization picks one of several options, and each option's true value depends on every dimension.

The hierarchy

Frontline agents each observe a few dimensions with noise. Every manager averages what its reports send, adds its own consistent bias, and passes up only the dimensions it judges most important. That judgment is rational: it ranks dimensions by their long-run average impact. This is exactly why the rare, high-impact dimensions get discarded, including in the moments they matter most. Each hop also adds a tick of delay, so the leader always decides on a slightly old world.

The leader

The leader estimates each option's value from what arrived and picks the best by cost and benefit. It tracks two numbers: actual completeness, the share of what currently matters that it can see, and believed completeness. The gap between them is the blind spot. The leader can't see missing data, but it can feel surprise when outcomes miss predictions. Surprise lowers belief; quiet periods slowly restore it.

Reaching down

Trigger

Surprise is the miss between the predicted and realized value of the chosen option, measured either against recent surprise (so persistent problems habituate) or against a baseline fixed in the first 150 ticks. A reach goes out when surprise crosses a threshold. The reflection knob lowers the threshold, and higher believed completeness raises it, so false confidence makes the leader reach less.

Route

Through the chain, the request travels down and back through every layer. New questions can be reinterpreted into questions about things already known, and answers are filtered by the same managers who discarded them the first time. A skip-level reach is fast and unfiltered but carries a bypass cost.

Scope

Verification audits numbers the leader already has. It can remove bias, but only if the audit doesn't pass through the same managers, and it can never reveal a missing dimension. It also reassures the leader, which raises believed completeness. New questions pick unseen dimensions at random, because the leader doesn't know what it's missing.

Return

Found dimensions are watched for a while. Raw skip-level data costs twice the attention of compressed data. When total load exceeds capacity, the leader's estimates get noisier.

Costs you can switch off

Manager bias and reporting delay can each be turned off without changing anything else. With delay off, every layer passes on what it received in the same tick, so the leader decides on the current world rather than one that is several ticks old. Reaches still take time to travel; that is set separately by the delay per hop.

The audit and the model of the business

The organization has a model of how each dimension affects each option's value. The leader decides with it, and when the audit is on, the audit uses it too. Model error sets how far that model is from the truth. Persistent error is wrong in the same way every tick; random error is freshly wrong each tick. Regime shifts change how a quarter of the dimensions affect outcomes, and the model never catches up, so shifts make it steadily more stale.

Each tick the audit takes every dimension the leader can't see but someone in the organization measures, adds it back, and re-runs the decision with the model. If including it would improve the decision by more than the flag threshold, the dimension is flagged and watched for 20 ticks. The divergence plot costs the leader one unit of attention. A quiet audit reassures the leader a little; flags lower confidence. Unmeasured dimensions are invisible to the audit and can only be found by a reach asking a new question.

Delegation and direction

With delegation on, each department owns a slice of the dimensions, and its frontline observes only those. The leader chooses a direction from a market view of each direction plus the compressed data that reaches it, and holds that direction until a review. Each department head chooses one of four implementations using its own department's data, which is fresher and less compressed than the leader's, and picks whichever best serves the direction as the department understands it. Data that stops at the department and gets used there is terminal discard; nothing is lost.

Two couplings decide when that stops being true. Upward coupling sets how much conditions in the departments' domains decide which direction is right, which the leader can only see through the compressed chain. Lateral coupling sets how much each department's choice affects the next department's domain, which the choosing department can't see at all.

The direction is reviewed on a planning timer and whenever strategy pressure crosses a threshold. Pressure builds from sustained surprise, measured against the leader's prediction of the direction's value plus the departments' reported expectations, and decays over time; audit flags add to it, since each one says another direction would be better. A review re-chooses the direction and may keep it. After a change, departments keep serving the old direction for the realignment time, which is what makes pivots costly.

Who matters

Any role can be made less skilled: the leader when choosing a direction, one department head, or all of them when choosing implementations. Skill gap is decision noise scaled to how far apart that person's options are, so the same gap means the same degree of misjudgment for every role. Strategic stakes scales how much value rides on the direction compared with the departments' own results.

Whenever a role is weakened, each sweep run also runs an unweakened twin in the identical world, so the difference is caused by that one person. Two measures come out of it: the score lost, and the share of time the organization is heading in a different direction from its twin. Read the second with care. The direction path is sensitive: a department head can shift it through surprise, pressure and forced reviews, and even a negligible skill gap can send the two paths apart. That measures how much the path depends on a person, not whether it matters; the score lost says whether it matters.

Tail measures

Averages can improve while the worst outcomes get worse, so the sweep also reports surprise frequency, the size of the worst 1% of surprises, large misses (choices at least 1.5 times the usual spread below the best), the worst 50-tick stretch of decisions, and how much wider the blind spot was in the ten ticks before the worst surprises than on average.

Turnover

At each tenure the leader is replaced and forgets its watches and corrections. Managers are replaced at random with the same average tenure; a new manager has a new bias and no sense yet of which dimensions matter, so for a while it passes up a more random selection.

What this model is not yet

The parameters are illustrative, not calibrated to real organizations, and the headline curves depend on them. Tenure does not yet shorten planning horizons, which is where the dual-class and CEO-horizon research would come in as validation data. The environment is a drifting set of dimensions rather than a full rugged landscape. Each of these is a place to extend the model, and each mechanism should be checked against its source result before being trusted in combination.