Open sourceFail closedGitHub ready

Every merge
needs evidence.

AI proposes changes. MergeGrounds decides what has earned the right to merge—with design contracts, mutation testing, security gates, and exact-revision evidence.

Read the quickstartView sourceImmutable demo v1.0.1

No “100% safe” claims. Missing, stale, malformed, or self-graded evidence is denied by the documented controls instead of silently passing.

01 / THE TRUST GAP

The problem isn’t AI. It’s unearned trust.

Linters, scanners, and tests produce signals. MergeGrounds turns those signals into one fail-closed admission decision—bound to the source, policy, reviewer, and artifact that actually ran.

02 / 90-SECOND MODEL

See one admit and five denials.

Run the deterministic v1.0.1 demo before touching a repository. It is an Educational model, not a repository audit or a production assurance claim.

Why real activation begins red
mergegrounds-demo / v1.0.1EDUCATIONAL MODEL
$ python3 demo.py

ADMITTED admitted complete exact evidence
DENIED stale-evidence stale:evidence
DENIED wrong-commit revision_mismatch:unit-tests
DENIED incomplete-scope scope_mismatch:static-analysis
DENIED survived-mutant survived_mutant
DENIED missing-producer missing_producer:static-analysis

DEMO PASSED: 1 admitted control; 5 negative controls denied
Educational demo only — this is not a production assurance claim.

03 / THE DECISION LAYER

Not another scanner.

Keep the tools you trust. MergeGrounds reconciles their evidence under one deterministic policy and refuses ambiguous results.

  1. INPUTSTests · scanners · build · reviewers
  2. MERGEGROUNDSPolicy + evidence reconciliation
  3. OUTPUTAdmit · deny · explicitly waive

04 / CONTROL SURFACE

Hard gates across the lifecycle.

Stack-aware where tools differ. Stack-independent where trust cannot.

01

Design before code

Risk-ranked changes must carry reviewed acceptance oracles and failure behavior before implementation starts.

02

Strict quality gates

Formatting, lint, types, unit tests, coverage, and build are required signals—not optional suggestions.

03

Mutation that discriminates

Tests must kill meaningful mutants. Survived, uncovered, timed-out, or invalid mutants fail the gate.

04

Security by default

Secret, SAST, dependency, workflow, and supply-chain controls reject missing or inconclusive output.

05

Exact subject binding

Evidence is bound to the candidate commit, base, policy, tree, and artifact digest—never only a branch name.

06

Build once, promote

Release controls promote the reviewed artifact by digest instead of rebuilding something merely similar.

07

AI-product assurance

When AI ships in the product, evaluation extends to retrieval, context, tools, providers, drift, and rollback.

08

Humans stay accountable

Independent review and explain-back remain explicit. A second model is defense in depth, not approval.

05 / HONEST ASSURANCE

Two levels. One clear boundary.

The repository delivers strong portable controls. Maximum assurance additionally needs a separately administered verifier and protected GitHub settings.

AVAILABLE IN THE REPOSITORY

Portable controls

  • Strict policy runner and evidence schemas
  • Quality, mutation, security, and supply-chain gates
  • Risk-ranked design and change contracts
  • Dry-run bootstrap preview before reviewed apply
Start here

REQUIRES EXTERNAL TRUST

Maximum assurance

  • Independent GitHub App or protected verifier
  • Isolated execution against read-only source
  • Authenticated reviewers and signed attestations
  • Rulesets with no administrator bypass
Read the trust model

06 / ADAPTER DEFINITIONS

Included adapter definitions

Included adapter definitions, not blanket support. End-to-end validation and public green fixtures are pending for each ecosystem. Review and pin every project-owned command and toolchain before relying on an adapter.

Node / TypeScriptPythonGoRustMavenGradle.NETPHPStrict custom adapter

07 / FIRST VALUE

Preview first. Change nothing.

After the educational demo, bootstrap preview reports only CREATE and CONFLICT planning results, prints Dry run only, and exits 0 on success. Counts depend on the target. Review every entry before apply.

Full quickstart Security model
01
Clonegit clone --branch v1.0.0 --depth 1 https://github.com/ExCoder/mergegrounds.git
02
Previewpython3 -I mergegrounds/scripts/bootstrap.py --target .
03
Review, then applypython3 -I mergegrounds/scripts/bootstrap.py --target . --apply
04
Verify; expect redpython3 -I scripts/mergegrounds.py verify-repo --strictUnderstand the expected-red result

OPEN SOURCE · APACHE-2.0

No grounds. No merge.

If the demo or bootstrap preview delivered value, starring is optional. An issue or reproducible feedback is equally useful. Bring back false positives, awkward adapters, and failure modes; that is how the standard gets stronger.

View GitHub sourceJoin the community