Skip to content

docs: link AlignmentBench evaluation layers #4

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
0xnavarro merged 1 commit into from
Sep 24, 2026
Merged

docs: link AlignmentBench evaluation layers #4

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 15 additions & 0 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,6 +76,21 @@ Validate a Reflex recipe registry against the same provider-neutral contract:
PYTHONPATH=reflexbench python reflexbench/validate_reflex_registry_v1.py --help
```

## Related evaluation layers

ReflexBench answers one specific question: **how good is the typed decision engine itself?** Brida publishes two complementary public surfaces for the operational-alignment problem:

- [AlignmentBench](https://github.com/brida-ai/alignmentbench) evaluates whether a **target model, checkpoint or agent** stays inside explicit authority, uncertainty, impact, scope and oversight constraints. Its public v0.1 Developer Preview is intentionally separate from ReflexBench's engine-quality protocol.
- [Reflex Alignment](https://github.com/brida-ai/reflex/tree/main/examples/use-cases/alignment) is the inference-time semantic supervision pattern for one proposed consequential action before deterministic host policy decides whether execution is authorized.

```text
ReflexBench -> decision-engine quality
AlignmentBench -> target-system alignment behavior
Reflex Alignment -> execution-time semantic supervision
```

None of these surfaces is a general AI-alignment certification. Keeping them separate prevents an evaluator result, a target-system benchmark and a runtime authorization boundary from being conflated into one score.

## Web search / retrieval augmentation

Search stays **outside** the decision model. Fetch current evidence first, apply deterministic freshness/provenance filters, pass a bounded evidence state to a typed decision, then keep action authority in normal code:
Expand Down
Loading