Add consensus-trust learning (DynaTD) with MBFC credibility prior + e…#33
Draft
muhaiminsk wants to merge 1 commit into
Draft
Add consensus-trust learning (DynaTD) with MBFC credibility prior + e…#33muhaiminsk wants to merge 1 commit into
muhaiminsk wants to merge 1 commit into
Conversation
…valuations - Credibility-weighted consensus targets for DynaTD (fixes self-referential trust feedback) - MBFC-based credibility prior (credibility_scorer_v3) - Evals: 40-row state-media (frozen contexts + NLI-replay demo), AVeriTeC 423 prequential - Baselines: Guardian 3.3/5b, granite-4.1-8b judge, g-3.2-8b-sft-detect-v5 - switch-4.1 adapter: negative result documented (root cause confirmed via vendor io.yaml) - Unit tests for trust module (feature extraction, UTD, DynaTD, BayesianTrustFusion/CredibilityTrustFusion) - Frozen datasets, trust states, trajectory logs for deterministic reproduction Signed-off-by: Samit Muhaimin <Samit.Muhaimin@partner.ibm.com>
muhaiminsk
force-pushed
the
feature/consensus-trust
branch
from
July 17, 2026 13:26
bb5d070 to
b2c6654
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
ConTrust: Consensus Trust Learning with MBFC (Media Bias/ Fact Check) credibility prior + evaluations
Introduces ConTrust (Consensus Trust Learning) — a source-reliability layer for
FactReasoner. Contexts are weighted by an MBFC-based credibility prior blended
with a dynamically-learned reliability score (implemented as DynaTD) that updates
from evidence rather than the model's own belief, fixing a self-referential trust
pathology in the original formulation.
(fixes self-referential trust feedback)
Results (hard-fact subset, n=27): ConTrust 81.5% vs vanilla 77.8% vs Guardian-3.3 70.4%.
AVeriTeC (423 claims, gold evidence): tied at ceiling (96.7% vs 96.6%) — honest
negative result on clean evidence. Limitations (held-out dev-set non-replication,
US/English-only credibility scorer, falsified geo-independence experiment) are
documented in full in experiments/consensus_trust/RESULTS.md.