Fission Benchmark

Fission ↔ Ghidra layered parity

Fission and Ghidra share a p-code-class intermediate representation. This page measures structural agreement — assembly bytes, p-code ops (strict/loose dual), CFG topology, and function inventories — with Ghidra as reference and Fission as candidate.

Match rates here are not multi-decompiler quality rankings. Oracle semantic pass rate lives on the Multi-decompiler page.

Ranking surface

Multi-decompiler quality →

Semantic oracle, coverage, same-function matrix, and full tool grid.

Reference contract

Ghidra inventory & IR

Unified runner: python -m runner.run_parity with decompilers fission,ghidra.

Dual metrics at a glance

Prefer dual rates for triage. Strict set / full equality is intentionally hard (CRT, space IDs, inventory heuristics).

Assembly match

100.0%

40/40 comparable · coverage 100.0%

P-code opcode seq

100.0%

Dual · sequence agreement (not full IR eq)

P-code loose full

100.0%

Stub space-selector abstracted

P-code strict full

100.0%

Primary status match 100.0%

CFG block starts

99.5%

Set match 97.5% · edge jaccard 98.8%

Function discovery

100.0%

Set match 100.0% · manifest 100.0%

Layer stack · full telemetry

Primary publishable stages: assembly · p-code · CFG · function discovery. Decode / extension tracks are listed but not headline quality.

Layered parity with Ghidra as reference and candidates (typically Fission). Headline stages = assembly + p-code + CFG + function inventory. Match rate is among comparable rows; coverage shows infra health. This is not multi-decompiler semantic ranking. Mode: strict (conservative / no leniency).

Headline rollup: match_rate=99.2% · coverage=100.0% · rows=129 · Headline quality = assembly + pcode + cfg + function_discovery (Ghidra inventory). Decode/IR/golden are diagnostic or meta.

P-code dual rates (do not read strict 0% alone): opcode=100.0% · loose_full=100.0% · strict_full=100.0%

Function inventory dual: set match_rate=100.0% · mean presence_recall=100.0% · mean manifest_recall=100.0% (exact set equality is strict; Fission often finds fewer CRT/helpers)

Reliability notes: function_discovery near-100% inventory match — verify Ghidra/candidate are not under-discovering the same way. · ir_invariants high match rate reflects weak structural checks only. · pcode literal_full << strict_full: remaining gap is space-table id encoding (documented selector_abstract_under_strict policy).

Stage	Total	Match	Mismatch	Match rate	Coverage	Mismatch rate	Top mismatch kinds
Assembly assembly_parity	40	40	0	100.0%	100.0%	0.0%	—
CFG cfg_parity	40	39	1	97.5%	100.0%	2.5%	block_count:1
Functions function_discovery	9	9	0	100.0%	100.0%	0.0%	—
P-code pcode_parity	40	40	0	100.0%	100.0%	0.0%	—

Extension tracks — diagnostic or incubating; not scored for primary ranking. Promote a stage by removing it from NON_PUBLISHABLE_STAGES in benchmark/telemetry/aggregate.py.

Stage	Total	Match	Mismatch	Skipped	Match rate	Coverage	Top mismatch kinds
Golden repros golden_repros	12	0	0	—	—	0.0%	—
IR invariants ir_invariants	40	40	0	—	100.0%	100.0%	—

Rows: 181 · pairs: ghidra→fission, invariants→fission · fetch_error_rate=0.0%

Core pair · semantic side-by-side

From the multi-decomp envelope (dev), filtered to Fission + Ghidra only. Useful context — not the ranking surface for other tools (see Multi-decompiler).

Fission · candidate

Semantic mean56.4%

Perfect / tested99/210

Adapter clean / attempted210/216

Boundary invalid5

Mean time2537ms

Ghidra · reference

Semantic mean77.3%

Perfect / tested164/215

Adapter clean / attempted215/216

Boundary invalid1

Mean time12606ms

Core · same-function rates

Boundary contract on the core pair (expected ~1.0 after address anchors). Core cohort rate: 98.5%.

Request contract (binary, addr) — did the adapter return that function unit? Primary rate = direct / (direct + boundary_*). Not a semantic ranking axis. Address anchor rate: 0.5%.

Cohorts — core: 98.5% (loose 98.6%) · multi: — · all: 98.5%

Decompiler	Cohort	same_fn_rate	loose_rate	Status counts
fission	core	97.7%	97.7%	direct_function:211 · whole_program_output:5
ghidra	core	99.4%	99.5%	direct_function:179 · needs_normalization:36 · whole_program_output:1

Reproduce

export FISSION_HOST_PORT=8007
python -m runner.run_parity --corpus dev --limit 20 --decompilers fission,ghidra
python -m benchmark.telemetry.aggregate
# → public/parity-telemetry.json

python -m benchmark.function_discovery.run --limit 10
python scripts/check_reliability.py public/parity-telemetry.json