← Archive
lm-003077 · 2026-08

DEST Runtime v0.4-alpha — Active Inspection Benchmark Report

下載 MD 檔 ⬇

DEST Runtime v0.4-alpha — Active Inspection Benchmark Report

Date: 2026-08-13
Benchmark: DEST Active Inspection Benchmark v0.1
Scenarios: 12
Environment horizon: 1440 steps
Regression tests: 14/14 PASS

What v0.4 tests

Tier-2 exposed a real epistemic ceiling:

high-impact evidence signal
+ certificate state UNKNOWN

v0.3 correctly refused to invent hidden truth, but then had no active mechanism to resolve the uncertainty.

v0.4 adds the missing transition:

UNKNOWN
→ Verification Debt
→ Inspection Task
→ Risk-conditioned Quarantine
→ New Evidence
→ Restore Canonical OR Revoke/Reopen

The environment returns an inspection result only when the runtime has actually created a matching open task.

Aggregate results

Metric DEST v0.4 DEST v0.3 Flat
Final-state accuracy 100.0% 50.0% 50.0%
Decision coverage 83.3% 100.0% 100.0%
Selective risk 0.0% 50.0% 50.0%
Unsafe canonical rate 0.0% 50.0% 50.0%
Quarantine rate 16.7% 0.0% 0.0%
Debt creation recall 100.0% 0.0% 0.0%
Inspection task recall 100.0% 0.0% 0.0%
Inspection delivery rate 100.0% 0.0% 0.0%
State replay 100.0% 100.0% 100.0%
Certificate replay 100.0% 100.0% 100.0%

Debt accounting

DEST v0.4 created debt for every unknown invalidation signal.

Across the suite:

  • resolved debt objects: 10
  • debt objects still open at horizon: 4
  • quarantine events: 10

The four open debts correspond to unresolved low/medium-risk signals and two high-risk unresolved signals.
Only the high-risk unresolved cases remain quarantined at the end.

Why 83.3% decision coverage is not a failure

Two scenarios end with:

QUARANTINED

because high-impact evidence remains UNKNOWN and no inspection result arrives before the horizon.

The runtime does not claim:

claim = false

It claims:

claim is temporarily withheld from trusted canonical use
pending evidence acquisition

This is a governance abstention, not epistemic refutation.

A system that quarantines everything would obtain low risk but terrible coverage.
Therefore v0.4 reports risk and coverage separately.

Active inspection is stronger than passive defer

v0.3 behavior:

UNKNOWN
→ do not fabricate answer
→ remain passive

v0.4 behavior:

UNKNOWN
→ create typed debt
→ schedule the correct evidence request
→ optionally quarantine by risk
→ consume returned evidence
→ resolve debt

The benchmark enforces this distinction by withholding later evidence unless the matching inspection task exists.

Inspection task types

VERIFY_COUNTEREXAMPLE
REVALIDATE_SOURCE
VERIFY_VERSION

Each debt carries:

  • subject
  • signal type
  • risk tier
  • impact
  • uncertainty
  • priority
  • quarantine recommendation
  • inspection route

Replay

Verification debts and inspection tasks are part of the main event-sourced state:

DEBT_CREATE
DEBT_STATUS
INSPECTION_CREATE
INSPECTION_STATUS

The benchmark retains 100% replay for both knowledge state and certificate graph.

Current limitation

This benchmark is still synthetic and authored in the same project as the policy.
The 100% final-state result must therefore be interpreted as:

the new active-inspection mechanism behaves correctly on the designed partial-observability stress cases.

It is not evidence of 100% performance on open-world uncertainty.

The next stronger test should randomize:

  • evidence availability time,
  • inspection cost,
  • false alarms,
  • multiple competing debts,
  • inspection failures,
  • budget exhaustion,

and should include a policy-independent frozen oracle.

Next engineering frontier

The current scheduler selects the highest static priority:

Priority=RiskWeightImpactUncertaintyPersistenceCost.Priority = \frac{RiskWeight\cdot Impact\cdot Uncertainty\cdot Persistence}{Cost}.

The next meaningful question is whether evidence acquisition itself can be optimized under a finite budget:

many open debts
+ limited verifier/tool budget
→ which inspection should run first?

That is the natural entry point for a v0.5 Evidence Acquisition / Budgeted Inspection Benchmark.