cd /news/artificial-intelligence/mca-open-source-neuro-symbolic-wrapp… · home › topics › artificial-intelligence › article
[ARTICLE · art-170252] src=discuss.huggingface.co ↗ pub= topic=artificial-intelligence verified=true sentiment=· neutral

MCA: open-source neuro-symbolic wrapper that teaches 7B LLMs to challenge false premises, explain paradoxes and abstain (code + dataset + Space + 2×T4 benchmark)

Developer Mati83mon opened pull request #6 on the Mati83mon/MCA GitHub repository, adding an audit mode that runs every layer of the open-source neuro-symbolic wrapper on every item without changing the decision, plus a per-rule verdict table for uncensored threshold calibration. The change adds `audit_all_layers=True`, a `DecisionController.rule_verdicts()` method tested to reproduce `decide()` on all 300 stored rows, and `benchmarks/rule_table.py` for offline grid-search of thresholds tuned on 100 items and evaluated once on held-out probes. The author disclosed that the entropy threshold τ = 0.6 was a prior default and that the R8b evidence-answer thresholds were chosen on a censored sample covering only 20–26 of 100 items per model, and said an uncensored audit run across 3 models is underway on Kaggle with a 24-item held-out set (6 control facts) too small to detect small over-refusal differences.

read1 min views1 publishedNov 16, 2026

Thanks — this is exactly right, and I checked it against the stored telemetry before replying.

Where it holds: R8 is censored. The slow path (entropy sampling) is cancelled as soon as the fast path decides, so entropy was recorded for only 20–26 of the 100 items per model. The entropy threshold itself (τ = 0.6) was a prior default, not fitted to these logs — but the thresholds of an “evidence answer” rule (R8b) I added a few hours ago after a Space test were chosen by inspecting exactly that censored sample.

Where it’s partly already covered: the fast-path rules (R1–R7, incl. R5) store every premise verdict, so their independent verdicts are recoverable for all items. Standalone, R5 fires on 6/25 control facts for Llama-3 and 0/25 for Qwen2.5 and DeepSeek-R1.

What I changed (Audit mode + per-rule verdict table for uncensored threshold calibration by Mati83mon · Pull Request #6 · Mati83mon/MCA · GitHub):

audit_all_layers=True runs every layer on every item without changing the decisionDecisionController.rule_verdicts() returns each rule’s own verdict (tested to reproduce decide() on all 300 stored rows)benchmarks/rule_table.py builds the per-item rule table and grid-searches the thresholds offline — tune on the 100, evaluate the chosen config exactly once on the held-out probes An uncensored audit run (3 models, 100 items + held-out) is running on Kaggle now; I’ll post the per-rule table and the held-out result here, whichever way it comes out.

One caveat on my side: the held-out set is only 24 items (6 control facts), so it can show a calibration that clearly fails to travel, but not a small over-refusal difference. I’ll say so with the numbers, and a larger held-out set is the next thing to build.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @mati83mon 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/mca-open-source-neur…] indexed:0 read:1min 2026-11-16 · —