feat(0.28.0): NLI pair judge — merges decided by meaning, not word overlap #12

Merged

navigator merged 2 commits from feat/nli-judge into main 2026-07-10 01:28:13 -03:00

Conversation 0 Commits 2 Files changed 11 +610 -16

navigator commented 2026-07-10 01:08:39 -03:00

Owner

Copy link

Closes the failure class reported by the steward: sentences with near-identical words and different meanings (obra do mestre Picasso ≠ mestre de obras — measured cosine 0.068, UNDER the merge threshold). The embedding now only RETRIEVES candidates; an NLI cross-encoder reading the pair jointly DECIDES, with a confidence-anchored policy calibrated on a pt-BR matrix (8/8). Also hardens the stance lexicon with regression-tested collision fixes. Addresses the V2 direction of #10 (model-based contradiction check) — the remaining #10 item is the larger embedding model.

Closes the failure class reported by the steward: sentences with near-identical words and different meanings (obra do mestre Picasso ≠ mestre de obras — measured cosine 0.068, UNDER the merge threshold). The embedding now only RETRIEVES candidates; an NLI cross-encoder reading the pair jointly DECIDES, with a confidence-anchored policy calibrated on a pt-BR matrix (8/8). Also hardens the stance lexicon with regression-tested collision fixes. Addresses the V2 direction of #10 (model-based contradiction check) — the remaining #10 item is the larger embedding model.

navigator added 1 commit 2026-07-10 01:08:39 -03:00

feat(0.28.0-nli-judge): the merge gate reads pairs jointly — meaning, not word overlap ... 811010e5f4

User-reported failure class, verified by measurement: "A grande obra do
mestre Picasso" vs "A pica de aço do mestre de obras" measure cosine
0.068 — UNDER the merge threshold — and the two senses of "banco" 0.078,
while a legitimate paraphrase can measure 0.116. Per-text embeddings
compress each sentence independently; meaning-in-context requires
reading the pair JOINTLY.

- nli_judge.rs: multilingual NLI cross-encoder (mDeBERTa-v3-xnli via
  candle, local CPU) → entailment/neutral/contradiction per direction.
  Policy anchored to the measured pt-BR matrix with a 0.5 confidence
  bar: confident contradiction vetoes (antagonists measure 0.84-1.00);
  confident entailment in either direction accepts (creche paraphrase:
  0.98); neutral/weak pluralities do NOT merge — kills homonyms,
  same-place-different-work (cures the 0.27.0 documented limitation)
  and different-scope asks. 8/8 on the calibration matrix.
- PairJudge trait (domain) + optional judge on ClusterService;
  CONSENSUS_NLI_DIR env, OnceLock load, fail-open with audit log.
- Migration 0518: consensus_embedding.text_sample (~1200 chars) so the
  judge can read cluster members; gate checks up to 3 members.
- stance.rs precision fixes exposed by the same critique: "contra"
  became exact-match (it captured "contraTAR" as a negator), "barrar"
  removed (collided with "barragem"), prefix-collision exclusions
  (elevador, cortesia, cortina, vendedor, vendaval, fechadura,
  acabamento) — each with a regression test.
- Deploy: NLI artifacts via hostPath /opt/dsoc/model-nli; pod memory
  1.5Gi → 3Gi.

navigator added 1 commit 2026-07-10 01:18:33 -03:00

ci: retrigger after runner disk exhaustion (ENOSPC during workspace tests) 8ba573d1e0

navigator merged commit 8ba573d1e0 into main 2026-07-10 01:28:13 -03:00

navigator referenced this pull request from a commit 2026-07-23 04:03:26 -03:00

feat(0.35.0-base-contatos): a base de audiência nasce

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels   

2fa

  

agora-i18n

  

campaign

  

consent

  

data-extraction

  

epic

  

intercoms

  

l10n

  

lgpd

  

roles

  

sms

No labels

2fa

agora-i18n

campaign

consent

data-extraction

epic

intercoms

l10n

lgpd

roles

sms

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

brasil/democracia-social!12

No description provided.