Deliberate practice / V8 source research
A deliberate-practice environment for learning how to orient in V8 source, form and test vulnerability hypotheses, explain root causes, and retain reusable bug patterns over time.
What it is
Built for security researchers who can read C++ and want structured practice between studying known bugs and auditing unfamiliar V8 code.
Each case is built from a known V8 security fix, but you work in the vulnerable tree without the patch or answer. The system controls what context is exposed, records the reasoning process, and delays ground truth until you have committed an explanation.
This is not a from-zero C++ course, a sequence of finished exploits, or a promise that completing cases will make you discover novel vulnerabilities. It is a controlled way to build and measure the component skills that unguided auditing otherwise teaches with very slow, ambiguous feedback.
The case loop
Real source, a committed theory, high-resolution feedback, then delayed recall.
Begin from the vulnerable pre-fix tree. Guided cases identify the bug family, training area, and a small set of source files so the first task is building a map, not searching the whole repository.
After a first source pass, record a tentative mechanism and the first trace you would follow. The model is saved before investigation and feedback begin.
Read and navigate the real source, annotate decisive lines, cycle through hypotheses and evidence, run probes, and request graduated help only when needed.
State the mechanism and candidate trigger before seeing the fix. This preserves what you independently understood instead of letting the answer rewrite the attempt.
Compare vulnerable and fixed code, identify the decisive change, and encode the case as a causal chain and a reusable detection rule.
Reconstruct prior causal chains after increasing delays. As performance improves, source cues and hints are removed until cases require blind real-versus-clean decisions.
Intended outcomes
The goal is not recall of CVE trivia. It is a durable catalog of causal mechanisms and better judgment while reading unfamiliar code.
Progressive independence
The case workflow stays familiar while the task-defining cues disappear.
Research basis
The evidence base is cognitive science and educational psychology, not a claim that this interface has been neuroscientifically validated.
The recall queue asks for active reconstruction of the causal chain after 1, 3, 7, 14, and 30 days. A missed link remains due until the chain is reconstructed.
Evidence boundary: Research supports retrieval over restudy and distributed over massed practice. The exact interval sequence is a product policy, not a validated optimum for V8 knowledge.
Rowland, 2014 - testing-effect meta-analysisCepeda et al., 2006 - distributed-practice meta-analysis
You commit a working theory and later explain the entry, trusted fact, invalidation, unsafe use, missing guard, and trigger before reviewing the known fix.
Evidence boundary: The published evidence concerns learning across many instructional domains. A V8 causal-chain form is our domain-specific implementation, not a directly validated protocol.
Bisra et al., 2018 - self-explanation meta-analysisWisniewski et al., 2020 - feedback meta-analysis
Post-mortem review places vulnerable and fixed source together and asks you to select the decisive lines before opening the sealed explanation.
Evidence boundary: Case-comparison research supports comparison as a learning activity. Transfer from instructional cases to vulnerability discovery remains an inference that the product must measure.
When the queue permits it, recall cases from confusable bug families are mixed instead of practiced in one uninterrupted block.
Evidence boundary: Interleaving effects depend on the material and task. It is used here to practice discrimination, not as a universal improvement claim.
The support ladder moves from curated files and hints to self-selected source routes, uncued classification, cold localization, and blind existence decisions.
Evidence boundary: Evidence for fading worked steps is strongest for near transfer in structured problem domains; far transfer was not reliable in the cited study. That is why novel performance is evaluated separately.
Long-term memory is an explicit product target. The system uses delayed, effortful reconstruction to strengthen access to prior causal models. The papers support the underlying retrieval and spacing mechanisms; they do not establish that this particular schedule or product will improve zero-day discovery.
Release signal
One message when the first V8 training release is ready. No newsletter and no address sharing.
