Repository navigation
test(pal256): 64-hop numerical containment vs epistemic hallucination falsifiers - #1715
Conversation
…llucination falsifiers
…llucination falsifiers
|
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 |
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01G22yT6htkcdyXsihxxXdrg
Moves crates/bgz-tensor/tests/palette_multihop_error_probe.rs here from #1721. 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. #1721 is closed as superseded once this lands.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. |
…rror-probe-20261022 # Conflicts: # .claude/board/entries/README.md
…nerate entries index
…nerate entries index
…nerate entries index
The big hypothesis, separated into what we can test
Immutable Palette256 needle addresses and a frozen pairwise LUT avoid repeated re-quantization of reference vectors. Under independently proved contractive composition, errors can be bounded per hop. That does not alone prevent unsupported inference.
Includes
64-hop numerical test with known Lipschitz envelope; immutable lookup test; explicit adversarial negative controls (saturated false belief, repeated common ancestry).
Next falsifiers
Real 256x256 LUT compositions; independent f64 oracle and held-out semantic labels; source overlap; pairwise rho vs trajectory rho; abstention/false-positive calibration. This PR asserts no general hallucination prevention.
Run
cargo test --manifest-path crates/bgz-tensor/Cargo.toml --test palette256_multihop_error_probeDraft research probe only; CI not yet observed.