A goal-holding interpreter fails soft, and its workarounds tax a bounded budget

Type: kb/types/note.md · Tags: learning-theory, llm-interpretation-errors

A procedure is a goal compiled away: the artifact holds one path, the end is no longer in the representation, and a blocked step has nothing to re-plan from — fail-stop is structural, not a design choice. An interpreting agent holds the end and derives means at runtime, so a blocked path triggers re-derivation, navigation of an admissible space rather than a line. The two execution styles therefore have inverted failure economics. In the exact layer, failure cost is a step function and detection is free — the failure announces itself. In the interpreted layer, failure cost is amortized into workarounds and detection is destroyed by the very mechanism that makes the system robust: the workaround consumes the failure event, often leaving a successful outcome on top, which is why semantic recovery breaks traditional debugging intuitions.

One property, two faces

The robustness is not a separate mechanism from the interpreted layer's characteristic error. Underspecification — the interpreter's freedom to choose among readings — is the deviation source the error-correction asymmetry charges against semantic work, and it is the same property that supplies path multiplicity when a route is blocked. Interpretive freedom has two faces: deviation from intent, and resilience under partial breakage. Codification's trade restates itself accordingly: removing the freedom buys exactness and loud failure at the price of brittleness; keeping it buys survival at the price of silence.

The tax and the ceiling

Workarounds spend context, attention, and reasoning steps from a bounded pool, so an unfixed defect in the interpreted layer behaves like debt taxed per encounter: every traversal that meets it pays again. Accumulation does not produce a crash; it produces load — and because context degrades silently under volume, complexity, and interference, the approach to saturation is as quiet as the failures themselves. The interpreted layer's ceiling is workaround load crowding out task work, reached gradually and invisibly, where the exact layer's ceiling is one wrong transition, reached instantly and loudly.

A system like this KB exhibits the pattern directly: its linking contracts accumulated contradictions over a year of normal operation, surfaced only when a review was commissioned, and an index's completeness promise degraded quietly until a validator made the breakage loud. Operation continued throughout — which is the point, in both directions.

Design consequences

  • Detection must be engineered where failure is absorbed. The exact layer gets its tripwires free; the interpreted layer needs deliberately scheduled semantic checks — sampled audits over subsystems — because nothing else fires. This is the failure-economics rationale for periodic review of machinery that "works."
  • Maintenance is budget reclamation, not crash prevention. Fixing an interpreted-layer artifact retires a recurring tax rather than averting a catastrophe — and it harvests the failure event before a workaround consumes it, since a worked-around failure is a donated oracle never collected and a triggering observation never recorded.
  • The codification boundary gains a failure-economics criterion. Codify where failures should be loud — converting a silent tax into a tripwire, which is what the tag-readme validator did — and keep interpretation where survival under partial breakage matters more than exactness; relaxing trades back the other way. This is a lens the codify-versus-LLM heuristics do not currently list.

Scope

  • The soft-failure claim covers detected-blockage failures — the agent notices the obstruction and re-routes. Failures the agent does not notice — a false rule believed, a poisoned instruction followed — are not worked around; they propagate, and they ride the authority boundary rather than this one. An interpreted system can still fail catastrophically; it fails softly at blockage, not softly in general.
  • The budget claim is qualitative: per-encounter cost and soft saturation are asserted from mechanism and worked instances, not measured. Which defects tax heavily versus negligibly is unmodeled.

Open Questions

  • Is the workaround tax measurable — workaround events per session as a debt gauge — and could that same signal serve as the trigger for sampled audits, closing the detection loop the first consequence calls for?
  • Does saturation arrive as gradual degradation or does the interpreted layer have its own phase transition once re-routing dominates the context budget?
  • Can workaround events be cheaply captured at occurrence — turning the lossy repair into a recorded signal — without taxing the very budget they spend?

Relevant Notes: