4f13bd6733
A real model conflated an optimized CFG edge with the bytes at its instruction origin. Make provenance explicit and add bounded, batched instruction windows from the existing Kuna CLI, preserving exact bytes, per-window failures and Evidence. Record the task-cost comparison without treating the incomplete machine-path answer as an equal-quality win. Verified 33 focused tests, targeted lint, real stock/patched MCP decoding with mixed success, and a real 4096-row capacity run under the resource observer. BREAKING CHANGE: Experimental operation.address/addresses and block.start become origin_address/origin_addresses and origin_start. Printed CFG target_address becomes target_origin; historical source captures are unchanged. Co-authored-by: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com> Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>
Atomic RE tools: an experiment
Experimental branch only: no PR, merge, production registration, or Code Mode.
Started from REA main 5b16f2431062eea6ffd245a9e05512790007cb13.
Read measured results and mechanism / prior-art notes. The first round includes an object-reference counterexample and a failed code-mapping composition; it does not claim every hypothesis succeeded. The control-flow / SSA follow-up repairs three zlib mapping paths, tests a SQLite call guard, and records where structured slices cost more.
The question is whether composable facts improve completed reverse-engineering tasks, especially on complex software. Three outcomes matter together:
- Time, native work, CPU and process-family memory, including startup and retries.
- Whole-task estimated tokens: catalog, arguments, results and necessary follow-up.
- Correct answers with traceable evidence, meaningful coverage and explicit unknowns.
Smaller replies, fewer tool names and fewer calls are intermediate measurements, not proof of a better investigation. Negative results are useful experiment results.
Working definition
An atom is a coherent fact or relationship about an identified object, with its producing evidence and uncertainty. It need not be a single field, MCP call, instruction, native command or engine pass. A string's encoding and the status of that encoding belong together; separating them can change the meaning.
Candidate reusable building blocks:
| Object / fact | Relationships and questions |
|---|---|
| Artifact and analysis context | Which bytes, architecture, address space and analysis profile produced this? |
| Location and byte span | What is mapped here? Where are the source bytes? Is the span initialized? |
| Instruction | What bytes decoded here, and with which mode and provenance? |
| Function | Which entry is identified? Which locations does the engine actually assign to its body? |
| String, symbol, type | What was observed or recovered, and what remains unknown? |
| Reference / call / control-flow edge | What are both endpoints, the reference site, kind and coverage? |
| Decompiled view | Which engine reconstruction and machine locations support this selected text? |
An inventory extent is an upper bound, not proof that an interior address belongs to the preceding function. A reference to a string's start address is not proof of a runtime string use, and misses references into its interior. Different providers' recovered functions and types are not automatically the same entities.
Identity needs artifact digest + analysis profile + address space + object kind and location/discriminator. A separate session lease controls whether a handle is still usable. Identity, existence, lookup success and readiness are distinct.
Hypotheses to try to falsify
- Fine-grained getters can lose to coherent fact batches through repeated envelopes, schemas and round trips.
- A small workflow can be built from the same facts without maintaining a second analysis implementation. Explicit selection must survive composition.
- Batching at MCP may reduce agent turns without saving any native analysis. The engine's unit of work must be measured separately from the fact/read unit.
- Richer structured relations can improve investigation completeness more than another wrapper around whole-function C text.
- On complex targets, a narrow result can omit the decisive evidence. Compare coverage and final answers, counting all extra reads and failed paths.
First prototype
facts.mjs consumes REA's existing normalized observations. It does not parse
assembly/C or guess ownership. The same methods support individual reads,
homogeneous batches, and one string-reference-function workflow. Native analysis,
target admission, caches, cancellation and full Evidence ownership stay in REA.
server.mjs adds experimental tools to the complete production MCP catalog.
It does not hide unrelated schemas to manufacture catalog savings. The additional
schemas are included in costs. Its operations are static reads; it never executes
the target or mutates an attached analysis database.
benchmark.mjs compares production calls, individual fact calls, batches and
composition using actual MCP clients and actual Kuna/Ghidra. Verification runs
are separate from timed equivalent-work runs. Each performance run repeats the
task in one session to distinguish first queries and warm native caches. Cold
Ghidra import/open cost must also be reported, not hidden in the query timer.
The prototype projects the current address-only xrefs contract. Reference kinds
stay unknown. Unsupported containing-function resolution stays unknown without
futile retries. This is intentionally a way to expose integration gaps rather
than claim that the adapter can recover information already discarded upstream.
Full Evidence is explicitly readable while its existing session owner retains
it; closing the binary expires those records. Paging covers the provider-returned
set and never establishes complete artifact discovery.
This is not a production contract proposal. Overlays, non-default address spaces, typed edges, native batching and live agent decision quality need further work. No measured claim about improved complex-program reasoning is established yet.
Native experiments
After the build described below, with REA_KUNA_COMMAND configured:
mkdir -p .cache/atomic-re-tools
gcc -O2 -fno-pie -no-pie -o .cache/atomic-re-tools/interior-reference \
experiments/atomic-re-tools/interior-reference.c
# Run once for each variant: per-address, then inventory.
python3 experiments/analysis-efficiency/resource-run.py --observe-payload-exit -- \
node experiments/atomic-re-tools/native-relations.mjs \
.cache/atomic-re-tools/interior-reference HEAD:REA per-address \
.cache/atomic-re-tools/native-interior-per-address
python3 experiments/atomic-re-tools/verify-interior.py \
.cache/atomic-re-tools/interior-reference \
.cache/atomic-re-tools/native-interior-per-address.report.json \
.cache/atomic-re-tools/native-interior-inventory.report.json
Use the installed zlib/SQLite paths and incorrect/malformed to repeat the
native plan comparison. Append --code-slice to the per-address experiment for
an exact instruction-to-line probe. mapped-code.mjs rejects unsupported address
precision and never substitutes a nearby line for a missing mapping.
The experiment adds an opt-in typed reference projection at the existing Kuna protocol owner. The existing public address-only result and its retained cache shape remain the same; ordinary queries do not materialize the richer facts. Experimental MCP tools have live-validated input schemas, but no canonical output-schema/product-catalog integration. They are not a production SDK/API.
The v2 facade labels reference scope explicitly and connects locations in its
point-reference edges. A result's target preserves the caller's selector; it
is not a separate observation proving that an object exists at that address.
All experimental MCP results carry REA's canonical untrusted-target-content
marker. The v1 measurement table is retained with its revision label; it is not
relabelled as a measurement of this later response shape.
Reproduce
Use the repository's pinned Node/npm and run npm ci, then
npm run build:cached. Supply existing installations through REA_KUNA_COMMAND,
GHIDRA_INSTALL_DIR and JAVA_HOME; this experiment installs no provider.
node --test experiments/atomic-re-tools/facts.test.mjs
# Existing bounded runner: one CPU, 4 GiB hard memory cap, early stop at 3 GiB.
# Its payload-exit counters include all processes in the owned scope.
python3 experiments/analysis-efficiency/resource-run.py --observe-payload-exit -- \
node experiments/atomic-re-tools/benchmark.mjs \
kuna batch /usr/lib/x86_64-linux-gnu/libz.so.1 incorrect \
.cache/atomic-re-tools/kuna-batch
Use baseline, fine, batch, or workflow, and ghidra or kuna.
Append --verify for additional lifecycle, paging, negative and Evidence checks;
those runs must not be used as equivalent-work performance samples. Binaries,
provider projects, captured catalogs/traffic and local measurements stay ignored
under .cache/atomic-re-tools/, never in commits. Summarize reproducible results
and limitations here as experiments finish.