The failure mode

The outcome of a slice is speed × direction. The agent supplies the speed; the spec sets the direction. On a well-specified slice direction = +1, and high speed is a large step toward done. On an under-specified slice the agent invents the missing decisions, and any one wrong invention sets direction = −1 — so the same high speed is a large step away from done.

Read that carefully: speed is not the hazard, an unset direction is. A slower agent on an under-specified slice still walks the wrong way, only slower; it does not walk toward done. The remedy is to specify the slice, not to restrain the implementer. The fix lives upstream, in the brief — the same place the two earlier dives located the value.

Hold the direction at −1 — the under-specified slice — and vary the one thing the agent controls: its speed. The outcome is speed × −1, a signed magnitude. The faster the agent, the larger the step; the sign never changes. This is why slowing the agent down does not fix the failure mode: it shrinks the magnitude of a step that still points the wrong way. Only changing the sign — by specifying the slice — turns the speed back into progress.

Drag the speed below, from low to high, over the under-specified slice. The magnitude of the wrong step grows with the agent's speed; the direction stays negative until the spec changes it.

The Portal's web-bootstrap brief, f6.1.llms.md, is a slice specified down to F6.1-R1–R8, the T1–T7 task DAG, and per-story Acceptance gates. Its risk register names exactly the two hazards an under-specified or unverified slice produces — and closes each one upstream, in the brief and in the review, never by slowing the build.

RK-1 — the opened-gap. Leave the supervision tree's child list open and an agent assembling a “three-child” tree can omit the one child every render depends on. The register states it plainly: the tree omits Portal.Store — dropping the Store courses_of/1 reads, every course render is a dead-process crash, not a 404. The brief closes it by pinning the tree: Portal.Application keeps exactly [Portal.Store, Portal.EventStore.adapter(), {Portal.Engine, []}], and the Apollo gate makes a real Portal.courses_of/1 render, not merely a boot. The boundary RK-1 protects is the one the companion course draws at the facade boundary — the web touches only the Portal facade, so the Store must be present for the facade to read.

RK-3 — loss of independent verification. The register calls it the solo-topology hazard: the same context that builds also judges. When the context that builds the slice is the same one that reviews it, a fast build is graded by the same fast assumptions that produced it. The brief closes it by restoring an independent evaluator — a different context than the builder adversarially verifies and mutation-probes the work.

# f6.1.llms.md — the risk register: what an under-specified or unverified slice produces
#
# RK-1  The opened-gap: the tree omits Portal.Store.
#       Dropping the Store courses_of/1 reads → every course render is a dead-process crash, not a 404.
  fix: Portal.Application keeps exactly
       [Portal.Store, Portal.EventStore.adapter(), {Portal.Engine, []}]   # pin the tree
  gate: a real Portal.courses_of/1 render, not merely a boot
#
# RK-3  Loss of independent verification (the same context that builds also judges).
  fix: an independent evaluator — a different context than the builder verifies and mutation-probes

The thesis fails on an under-specified slice. The agent has nothing to implement, so it invents the missing decisions; because it is fast, its speed multiplies whichever direction those inventions set, and a small slip becomes a large step away from done. Neither f6.1.llms.md's RK-1 nor RK-3 is fixed by a slower agent — both are fixed upstream, by pinning the slice in the brief and by restoring an independent review. This closes A5.7: the value is real on a well-specified, gated slice, and the failure mode is the absence of that specification.

Primary sources for design by contract, for why an open-ended goal is where an agent improvises, and for small specified steps as the guard against fast error — with the hub, the prior dive, and the companion chapter whose risk register this grounds on.