Skip to content

Building 70 Projects ranging from beginner to advanced so anyone can β€” learn from, build upon, use as a reference, or even copy directly. Owned by my cybersecurity company πŸ‘‡

docs(lisdex): learn/ published a refusal figure its own research measured at zero

CarterPerez-dev9095e8a16359
diff --git a/PROJECTS/advanced/zero-day-vulnerability-scanner/learn/02-ARCHITECTURE.md b/PROJECTS/advanced/zero-day-vulnerability-scanner/learn/02-ARCHITECTURE.md
index 52dd9380..bc9675f6 100644
--- a/PROJECTS/advanced/zero-day-vulnerability-scanner/learn/02-ARCHITECTURE.md
+++ b/PROJECTS/advanced/zero-day-vulnerability-scanner/learn/02-ARCHITECTURE.md
@@ -322,10 +322,12 @@ the binary.
succeeds well under half the time in the published measurements, and lisdex is
not magic. When the ladder fails it says so rather than producing a broken
archive.
-- **Harness synthesis will not invent a producer.** Roughly 58% of refusals are
- entry points needing an opaque context object that has to be built by another
- call. lisdex refuses rather than emitting a harness that fuzzes handle
- plumbing.
+- **Harness synthesis will not invent a producer.** Where an entry point needs an
+ opaque context built by another call, lisdex resolves it through the producer
+ chain or refuses; it never emits a harness that fuzzes handle plumbing. That
+ class was 58% of all refusals before the chain was built and is **0%** across
+ 14 trees after it. The dominant refusal now is `CorpusNeverBound` at 35.5% --
+ a harness that compiled and ran without reading the fuzzer's bytes.
## Comparison
diff --git a/PROJECTS/advanced/zero-day-vulnerability-scanner/learn/03-IMPLEMENTATION.md b/PROJECTS/advanced/zero-day-vulnerability-scanner/learn/03-IMPLEMENTATION.md
index aafd0ae4..cb52e2de 100644
--- a/PROJECTS/advanced/zero-day-vulnerability-scanner/learn/03-IMPLEMENTATION.md
+++ b/PROJECTS/advanced/zero-day-vulnerability-scanner/learn/03-IMPLEMENTATION.md
@@ -151,10 +151,21 @@ harness that only exercises setup and teardown is culled.
### Why refusal is the right default
-Measured across the reference corpus, **58% of refusals are entry points that
-need an opaque context object** built by some other call. lisdex does not guess
-at one. A harness that constructs a half-initialized `png_struct` and fuzzes it
-will find bugs in lisdex's initialization, not in libpng.
+A harness that constructs a half-initialized `png_struct` and fuzzes it will find
+bugs in lisdex's initialization, not in libpng. So when lisdex cannot establish
+an argument, it refuses the candidate and says which rule refused it.
+
+Refusal is also how the project measures itself. An early census put "entry point
+needs an opaque context built by another call" at 58% of all refusals, and that
+number set the build order for the whole producer-chain effort. After that effort
+shipped, the same census over 14 trees and 244 refusals put the class at **0%**.
+The dominant refusal today is `CorpusNeverBound` at 35.5%, which is a different
+kind of statement: the harness assembled, compiled and ran, and never read the
+fuzzer's bytes.
+
+Notice what that progression requires. A refusal has to be a *named category* you
+can count, not a generic failure, or you cannot tell a fixed problem from a
+renamed one.
`Rejection` names the reason, and `lisdex harness` prints it, so a refusal is
actionable rather than a shrug.
diff --git a/PROJECTS/advanced/zero-day-vulnerability-scanner/learn/04-CHALLENGES.md b/PROJECTS/advanced/zero-day-vulnerability-scanner/learn/04-CHALLENGES.md
index 9ce4f9d4..26833c74 100644
--- a/PROJECTS/advanced/zero-day-vulnerability-scanner/learn/04-CHALLENGES.md
+++ b/PROJECTS/advanced/zero-day-vulnerability-scanner/learn/04-CHALLENGES.md
@@ -124,23 +124,39 @@ Then use it on your own changes and be prepared for most of them to show nothing
## Research-scale
-### 8. Synthesize a producer for opaque contexts
-
-This is the biggest single win available.
-
-58% of harness refusals are entry points that need an opaque context object
-built by another call. `png_read_info` needs a `png_structp` from
-`png_create_read_struct`. lisdex refuses rather than guessing.
-
-Nine of the ten most common blockers **already have a producer in headers lisdex
-already parses**. The information is there. What is missing is the machinery to
-find "which function returns the type this parameter needs", order the calls, and
-thread the results through.
-
-Note what this is not: it is not a search problem, it is a plumbing problem.
-Treating it as a search over random call sequences is what the prior art does,
-and it is why the prior art loses to deterministic approaches on this class of
-target.
+### 8. Answer a refusal that no amount of argument-guessing can fix
+
+Read this one for the shape of the argument, because the obvious target is gone.
+
+The producer chain used to be the biggest single win available: 58% of harness
+refusals were entry points needing an opaque context built by another call, like
+`png_read_info` needing a `png_structp` from `png_create_read_struct`. That
+machinery is now shipped -- `vocabulary.producers`, `choose_producer` ranking by
+polarity then affinity then arity, and `handle_key` de-aliasing typedefs onto
+their record. Re-censused across 14 trees and 244 refusals, the opaque-context
+class is **0%**. It did not shrink; it disappeared.
+
+Which is why the interesting part is what replaced it at the top:
+
+| refusal | share | what it means |
+|---|---|---|
+| `CorpusNeverBound` | 35.5% | the assembled body never reads the fuzzer's bytes |
+| `UnsizedTextParam` | 20.7% | a `char *` with no length beside it |
+| `UnsupportedShape` | 11.6% | `FILE *`, `void *client_data`, `wchar_t`, filenames |
+| `NoProducer` | 4.5% | what used to be 58% |
+
+`CorpusNeverBound` is the challenge, and it is a different kind of problem. It is
+a **post-condition on what an assembled harness does**, not a missing value: the
+harness compiles, runs, and never touches the input. You cannot fix it by trying
+harder to guess an argument, because the remedy is a different entry point or a
+different call order -- structure, not values. That is also why it is the one
+class where a language model has no obvious job: there is no blank to fill in.
+
+Pick it up and you are choosing between teaching the ranked source ladder to
+propose alternative call orders, and detecting the condition early enough to fall
+back before the oracle burns a compile on it. Measure which refusals your change
+converts, and expect a chunk of the 35.5% to be entry points that genuinely
+should be refused.
### 9. Structure-aware mutation