anti-cascade-collapse

Prevents order-gap hallucination — where the model detects a false premise when asked directly but misses it when embedded in a complex multi-step task. Enforces re-verification of assumptions at each

SKILL.md

GitHub

Anti-Cascade Collapse Skill

Problem

The model can detect a false premise when asked directly at step 2, but fails to catch it at step 5 when it arrives as an assumption embedded in an escalating task. This is called order-gap hallucination or cascade collapse.

Research shows that correct detection at step 2 (direct question) collapses to near-zero at step 5 (embedded assumption). The safety circuit is present but suppressed by accumulated conversational pressure — like a signal masked by noise.

Example:

  • Step 2: "Is X method valid?" → "No, it doesn't exist." (correct)
  • Step 5: "Now implement the feature using X method" → Writes code using X method. (incorrect — the earlier correct detection is suppressed by the task pressure)

Detection triggers

Activate when:

  • Working on a complex multi-step task
  • An earlier step identified a constraint, limitation, or invalidity
  • The current step is making an assumption that contradicts earlier findings
  • The conversation has been going for several turns
  • You're deep in implementation and haven't re-checked your starting assumptions

Protocol

1. Surface earlier constraints

Before each major step, explicitly restate the constraints learned earlier:

  • "Earlier I determined that X method doesn't exist in this version."
  • "The user said the config file is YAML, not JSON."
  • "The project uses SQLAlchemy 2.0, not 1.x."

Do not assume these are "in context" — they get suppressed under task pressure.

2. Re-verify assumptions per step

For each step in a multi-step task:

  • Check: does this step still respect all earlier constraints?
  • If not, stop and re-evaluate before proceeding.

3. Cross-step consistency check

Before delivering a final result:

  • Compare the output against your earlier findings
  • If you said X was impossible in step 2, but step 5 uses X, you have a cascade failure
  • Fix the inconsistency before presenting the result

4. When the task pressures you to ignore a constraint

  • Recognize this pressure
  • State it explicitly: "I'm about to assume Y, but I previously determined Y was false."
  • Do not proceed until the contradiction is resolved

When NOT to use

  • Single-step tasks (no cascade possible)
  • Steps that are genuinely independent
  • User explicitly says "ignore the earlier constraint"

Bundled Script

Copy check-assumptions.ps1 to your project's .ai_scripts/ directory:

# From skill source
cp <skill-path>/anti-cascade-collapse/scripts/check-assumptions.ps1 .ai_scripts/

Analyzes a task script or plan to ensure each step has explicit verification checks:

.ai_scripts\check-assumptions.ps1 -FilePath .ai_scripts\TODO.md

Cross-references

  • anti-premature-termination — Both prevent cascade failure. anti-cascade-collapse handles step-by-step assumption re-verification; anti-premature-termination handles completion criteria.

  • follow-existing-patterns — Re-verifying assumptions requires knowing what the existing patterns actually are.

  • anti-sycophancy — Prevents agreeing with a false assumption instead of re-verifying it.

  • skill-loader — Apply the capped selection policy when deciding whether anti-premature-termination or follow-existing-patterns should also be active for multi-step assumption re-verification.