PDA Failure Modes

Hovered up from pda-lab, Progressive De-Arbitrarization now appears to be especially suited to a region of tasks where raw-task-to-action mapping is not trustworthy without semantic repair.

Within that region, two dominant model failure modes appear repeatedly.

1. Specificity Erosion / Invention

The model reaches the right general abstraction level, but:

• invents concrete policy that was never earned
• smooths over settled distinctions
• replaces already-earned closure with plausible defaults

This is the failure mode most visible in specificity-sensitive tasks.

Example shape:

• the model understands that a task is about interaction semantics
• but fills in thresholds, proportions, or scope rules that were not actually grounded

Hovered Up

Specificity erosion includes invented details, and its deeper form is loss of coherent closure under temporally distributed decision-making.

When work unfolds thread-wise as:

• intent
• decisions
• implementation

then later decisions are made inside a field of already-made decisions. If those prior decisions are not actively preserved, two things tend to happen:

1. previously settled decisions are implicitly reopened
2. new local decisions become globally inconsistent with earlier closure

So for entangled decisions, coherence usually requires one of two structures:

• persistent global decision-state
• temporary joint closure of the entangled set

PDA tends to force the second. It squeezes a temporally dispersed set of interacting decisions into an explicit shared decision surface before implementation proceeds.

In that stronger sense, specificity erosion can be described as:

the re-fragmentation of previously closed decision surfaces under sequential reasoning

2. Authority / Locus Smearing

The model fails to preserve differentiated loci such as:

• who observes
• who decides
• who stores durable truth
• who initiates transitions
• what counts as transient failure
• what counts as true state transition or terminal state

Instead, it collapses these into one generic control flow.

This is the failure mode most visible in multi-authority tasks.

Example shape:

• the model sees a coherent operational problem
• but blurs room authority, budget authority, failure semantics, and grace-state transitions into one generic pipeline

Hovered Up

This is mainly a problem of structural differentiation.

The task requires distinct loci to remain explicit, such as:

• who observes
• who decides
• who stores durable truth
• who initiates transitions
• what counts as transient failure
• what counts as actual terminal state

The failure occurs when the model collapses that differentiated operational topology into one generic causal narrative.

In that stronger sense, authority / locus smearing can be described as:

the collapse of a differentiated operational topology into a single generic flow

The Unit Case

closure-sufficient can be understood as the unit case:

• PDA is still useful
• but these stronger failure modes do not activate beyond what a good closed formulation already repairs

So the current conceptual picture is:

• closure-sufficient is the baseline case
• specificity erosion / invention is one dominant non-unit failure mode
• authority / locus smearing is the other

Common Pullback

The two failure modes differ, but they also share a deeper form.

• specificity erosion / invention collapse of distinction through time

• authority / locus smearing collapse of distinction through structure

Their common pullback is:

failure to preserve articulated difference under generative pressure

Or more operationally:

collapse of distinctions that must remain explicit for coherent action

This also sharpens what PDA is doing at a deeper level.

PDA is an anti-collapse discipline for decision-relevant distinctions.

Organizational Transfer

These failure modes also transfer naturally into organizational coordination.

See:

• PDA Organizational Transfer

Why This Matters

This sharpens what PDA is doing.

PDA clarifies tasks in the abstract and counters the dominant ways a competent model reintroduces arbitrariness when grounding is insufficient.

That is why PDA remains useful even when the model is otherwise coherent, articulate, and operationally capable.