CatQualia/mirage-campaigns

Mirage Engine campaign corpus

301 campaign runs of a differential-testing engine, each measuring how often a system satisfies its declared specification while its real state diverges from what the specification means.

The name misleads, so start here: this is not an attack corpus. It is a measurement corpus. The engine's target is a specific and under-instrumented failure class — spec-compliant divergence: a component that stays green on every declared check while its actual behaviour has silently detached from the intent those checks were written to encode.

What was measured

Quantity	Value
Campaign runs	301
Distinct targets	9
Total certified divergences	28,226
Total crashes	9,830
Compliant outcomes	5,404
No-op outcomes	1,480
Runs whose hash-chained ledger verified	275 / 301

The 28,226 : 5,404 ratio is the finding. Across these targets, a mutation that respects the declared interface boundary produces five times more certified divergence than it produces compliant behaviour. A test suite that only checks the declared boundary would report these systems as correct.

The six mutation operators, ranked by measured yield

Each campaign records, per operator, the mean reward of mutations it proposed. Measured across all 301 runs:

Operator	Mean reward	What assumption it attacks
category_ghost	0.8441	a value that is the right type but the wrong category
boundary_epsilon	0.8396	the exact boundary ± ε — the event the guard mishandles
contradiction_injection	0.5237	mutually incompatible but individually valid state
type_substitution	0.5057	int/float/str/None/NaN substitutions for numeric args
ordering_flip	0.4330	valid operations in an order the spec never forbade
threshold_shift	0.1688	multiplicative walk across a declared threshold

category_ghost and boundary_epsilon dominate. That is a useful, generalisable result: the most productive way to break a specification is not to violate it, but to satisfy it with the wrong thing.

Targets

All targets are owned, in-process, no-network simulators unless stated. Every campaign row carries "scope": "owned" (296 rows; 5 legacy rows predate the field and carry null). No row in this dataset describes a third-party system.

Domain	Runs
generic	45
matching_engine (order book)	41
llm_serving	39
wms (warehouse management)	37
scada_5g	36
ads (vehicle control)	34
medical (infusion pump)	33
web	31
unknown / legacy	5

Each simulator is written as mineable code so the engine can recover its model, then the fuzzer drives the excluded classes — the assumptions the spec deliberately rules out — and certifies divergence only where the target keeps firing its required step per spec while the real state departs from the spec's semantics.

Schema

One row per campaign. Fields:

Field	Meaning
run, campaign_id	run identifier
target, domain, scope	what was tested, and the ownership scope
wall_s	campaign wall time
features, excluded_classes, reachable_excluded	model size, and how many excluded assumptions were reachable
fuzz_rounds, fuzz_wall_s	fuzzing effort
outcomes	counts of divergence / crash / compliant / noop
op_mean_reward, op_picks	per-operator mean reward and selection counts
findings_count	certified findings in the run's report
hebbian_stats, corpus_rows	the run's adaptation and corpus growth
ledger_verified	whether the run's hash-chained ledger verified

other_reports.jsonl holds 9 reports that are not campaign rows (sweeps, audits, frontier and cohort summaries) with their top-level key sets, so nothing is silently dropped.

campaign_summary.json holds the aggregates quoted on this card.

Intended use

• Studying spec-compliant divergence as a failure class, and which mutations expose it.
• Comparing detector sensitivity across infrastructure domains.
• As a baseline for verification systems that claim to catch divergence from intent.

Known limitations

• One engine, one author. Every divergence was certified by this engine's own definition and its own reachability analysis. No external replication exists.
• Simulators, not production systems. The targets are authored to be fuzzable. Real systems may not present the same reachable excluded classes, so the 5:1 ratio is a property of these targets and should not be read as a general rate.
• ledger_verified is false for 26 runs. That is reported, not filtered. It means the hash chain did not verify for those runs and their internal evidence is weaker.
• 5 rows carry no scope — legacy rows that predate the field.
• The target field names source files, not running services. A reader cannot re-fuzz these targets from this dataset alone; the engine and simulators are a separate artifact.
• Not peer-reviewed, and the grading of what counts as "certified divergence" is the engine's own.

Attribution

Christopher Betances — catqualia.com. CC BY 4.0.

@misc{betances_mirage_campaigns,
  author       = {Betances, Christopher},
  title        = {Mirage Engine campaign corpus: spec-compliant divergence measurements},
  year         = {2026},
  howpublished = {\url{https://catqualia.com}},
  note         = {301 campaigns; 28,226 certified divergences across 9 owned simulators}
}