Repository navigation
probe(nars-bayes): preserve sample size and predictive uncertainty beside CE64 - #1720
Conversation
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars. ⚙️ Run configuration
Warning Billing warning: we have not been able to collect payment for this subscription for more than 72 hours. Please update the payment method or pay any pending invoices in Billing to avoid service interruption. Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
…ompiles Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01G22yT6htkcdyXsihxxXdrg
Moves crates/causal-edge/tests/palette_bayes_nars_bnn_probe.rs here from #1720. Unlike the local helpers already in this series, it exercises the real implementation, so the series keeps its experiment contract and gains the concrete counter-probe. #1720 is closed as superseded once this lands. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01G22yT6htkcdyXsihxxXdrg
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
2 participants
Purpose
Experiment with Bayesian learning and BNN-style predictive uncertainty while preserving canonical NARS frequency/confidence and evidence provenance. BNN is intentionally scoped to Bayesian predictive uncertainty here; binarized neural network execution is a separate meaning and must not be silently substituted.
Existing anchor
crates/causal-edge/tests/d_rpf_p5_beta.rsalready establishes NARS expectation at horizon k=1 equals the Jeffreys-prior posterior mean, and documents u8 evidence-weight stalls. This PR does not reproduce a new NARS revision engine.New falsifiers
tanh(|z|).Verification
cargo test --manifest-path crates/causal-edge/Cargo.toml --test palette_bayes_nars_bnn_probe.Lab-only; no changes to production NARS or BNN runtime.
Next
Wire exact disjoint-source counts via DR/SoA and benchmark posterior predictive calibration against independent held-out evidence; define whether a separate binarized-NN experiment is required.