Abaris

Abaris is a maintainer-controlled patch-behavior validation and evidence-preservation harness.

Given a local repository, a known-vulnerable baseline revision, a candidate revision, one supplied reproducer, and one vulnerability-presence oracle, Abaris is intended to run the same paired experiment against both revisions in isolated environments and preserve the observed evidence.

Abaris does not claim that a vulnerability is universally fixed. Its strongest defensible conclusion is:

The supplied reproducer demonstrated the defined vulnerability-presence oracle on the baseline revision and did not demonstrate it on the candidate revision under the recorded conditions.

Project status

Abaris is in a documentation and architectural-foundation phase. There is no working CLI or supported execution backend. Security controls described in this repository are requirements, not implemented guarantees.

v0 experiment

The v0 workflow accepts:

• a local repository path;
• a known-vulnerable baseline Git revision;
• a candidate Git revision;
• one curated reproducer for an already-public historical vulnerability;
• one explicitly defined vulnerability-presence oracle; and
• an explicit execution policy.

The same reproducer, oracle, and comparison policy are applied to both revisions. Each run produces one primary observation:

Observation	Meaning
PRESENT	The oracle was evaluated successfully and demonstrated the defined vulnerability-presence condition.
ABSENT	The oracle was evaluated successfully and did not demonstrate the defined vulnerability-presence condition.
INDETERMINATE	The oracle could not be evaluated reliably under the recorded conditions.

Every INDETERMINATE observation must include a bounded technical failure_reason separate from the paired conclusion.

The paired observations produce one conservative derived conclusion:

Conclusion	Required observation pair
REPRODUCER_BLOCKED	Baseline PRESENT, candidate ABSENT
REPRODUCER_STILL_SUCCEEDS	Baseline PRESENT, candidate PRESENT
BASELINE_NOT_REPRODUCED	Baseline ABSENT, candidate ABSENT
CANDIDATE_ONLY_PRESENT	Baseline ABSENT, candidate PRESENT
INCONCLUSIVE	Any pair containing INDETERMINATE

REPRODUCER_BLOCKED does not mean universally fixed. The paired observations, immutable input identities, execution conditions, and evidence remain the authoritative result. CANDIDATE_ONLY_PRESENT is a blocking comparative asymmetry, not proof that the candidate introduced a vulnerability or regression.

Safety requirements

Repository contents, revisions, setup steps, reproducers, and workload output must be treated as untrusted executable content, including when they come from a curated public corpus.

The future supported execution profile must:

• expose no host secrets;
• expose no home-directory mounts;
• expose no Docker or equivalent host-control socket;
• use disposable isolated environments;
• deny external reproduction egress;
• for service-target cases, permit only declared reproducer-to-target communication on a closed per-run network;
• apply an explicit setup-network policy;
• enforce CPU, memory, process, disk, output, and wall-clock limits; and
• fail closed when required controls or evidence are unavailable.

Containers alone must not be described as a sufficient hostile-code boundary.

v0 scope

Abaris v0 is:

• maintainer-controlled and self-hostable;
• limited to local repository paths;
• limited to curated, already-public historical vulnerability cases;
• deterministic at its trusted observation and conclusion core;
• evidence-first and conservative; and
• human-controlled for any publication or export.

Abaris v0 explicitly excludes:

• private vulnerability reports;
• untrusted community reproducers;
• vulnerability discovery or exploit generation;
• root-cause verification or variant analysis;
• affected-version inference or duplicate detection;
• automatic patch generation or disclosure;
• hosted execution;
• broad regression claims; and
• any AI dependency in the deterministic core.

v0 source materialization must treat Git as a security boundary. It reads only from the operator-supplied local repository, ignores inherited Git configuration, and disables hooks, filters, credential helpers, LFS downloads, submodule initialization, and external protocols by default.

v0 may use private, disposable draft contracts while validating representative cases. It must not publish or promise a stable schema before 3–5 representative historical cases have completed review. Every draft artifact must self-identify as private-draft with no compatibility guarantee.

Documentation

• Vision
• Roadmap
• Threat model
• Security policy
• Contributing
• Architecture
• Reproducer specification
• Decision log

Standards and interoperability

Abaris should reuse established formats where their semantics fit, without allowing advisory or scanner metadata to determine observations. Potential future integrations include OSV for public vulnerability identity and affected range context, SARIF for compatible result export, VEX for carefully scoped statements, and public OSS-Fuzz testcase metadata.

Those integrations are not part of v0 unless separately accepted. Abaris needs its own narrow evidence model because existing formats do not fully express the paired experiment and its immutable execution conditions.

Responsible use

Abaris must only be used against repositories and environments the operator is authorized to test. It must not test third-party or production systems and must never publish vulnerability information automatically.

See SECURITY.md before attempting implementation or execution.