Quality & Audit Practice

• Continuous regression suite

  33-test automated suite anchored to canonical hadith (Bukhārī #1, #2, #8, #15, #7280; Aḥmad #22007) plus a 7-test cross-field consistency check. Runs on every release candidate; ship is gated on full pass.

• NAQD adversarial framework

  NAQD-1 / NAQD-2 / NAQD-3 adversarial agents stress-test grading decisions against fabrication patterns, theological subtleties, and cross-document contradictions. Findings are tracked publicly in the changelog.

• Provenance discipline

  Every record carries a `_provenanceDisclosure` block. Every grading decision carries a `verifiedBy` provenance stamp (deterministic vs model-assisted). External claims must cite the disclosure.

• External review cycle

  IPSC engages independent examiners on a recurring cadence to pressure-test framing, methodology disclosure, and headline claims. Reviewer feedback is integrated into the next release window.

Every release window includes recalibration items where the methodology was sharpened, a number was honestly de-inflated, or a framing element was tightened to match what the pipeline actually does. These are visible in the changelog as first-class release entries, not buried.

• v3.8

  Cluster integrity rebuild

  Rebuilt matn cluster taqwiyah counts to retire stale supportingChain inflation. Bukhārī sahih+hasan rate moved from 97.3% to 95.8% — an honest de-inflation, documented in the changelog.

• v3.11

  Framing recalibration

  Tightened external-facing framing to match what the pipeline actually does: applied AI grounded in classical rijāl methodology, structurally validated against documented teacher-student relationships, scholar-in-the-loop on residue queues.

• v3.11

  Ikhtilāṭ onset backfill

  11,380 onset years filled from documented sources so deterioration-period checks can run on a wider portion of the narrator graph.

• v3.11

  NRS undefined-tier explicit

  1,338 NRS narrator entries explicitly flagged as undefined-tier rather than silently inferred — improves downstream caller decisions.

• v3.12

  NAQD-1 V1 closure

  112 → 0 findings closed against the V1 NAQD-1 ruleset.

The full release-by-release timeline lives in thechangelog. The authoritative provenance disclosure that ships with every record lives at /provenance.

The 2026-05-02 review cycle is documented across three release entries on the changelog. Findings, scope, and remediations below; reviewer identities are held under agreement and not published.

• v3.11 — framing recalibration

  External-examiner cycle drove a framing recalibration on one stylistic document. Added _provenanceDisclosure tocorpus-v3/manifest.json. Established standing provenance-discipline rules. Closed NAQD-1 V1 (112 → 0 findings).See v3.11 release entry →

• v3.12 — NAQD-3 fresh-embedding re-run

  After verifying OpenITI 2025-1-9 source files, refreshed all448,237 matn embeddings (previous embeddings ran on broken-matn data). NAQD-3 adversarial framework re-run on the corrected embeddings surfaced the findings catalogued below.See v3.12 release entry →

• v3.13–v3.25 — citation cascade + 6 ingestion runs

  Generalized citation grade promotion beyond Sahihayn-only. OpenITI → IPSC ingestion pipeline (token-Jaccard + containment narrator matcher). Six new primary collections ingested (+8,241 records). 5 pipeline robustness fixes.See v3.13–v3.25 release entry →

NAQD-3 findings breakdown (v3.12)

1,851 total findings surfaced when the adversarial framework re-ran against fresh embeddings. Categorized:

Related deterministic adversarial results

• 6,822 records from fabrication-focused reference collections graded by the same engine — a deterministic check against collections whose contents classical scholars had already assessed. graded by the same engine — the system correctly identified what classical scholars identified centuries ago, providing a deterministic ground-truth check.
• 2,168 records capped via _naqd3Override source-collection caps in fabrication anthologies: Mawḍūʿāt (424), Tanzīh (1,238), Fawāʾid (506).
• NAQD-1 V1: 112 → 0 findings closed after v3.11. NAQD-2 ongoing on theological-subtlety subset.

IPSC is a computational tool, and computational tools improve fastest when they are paired with working scholars. The collaboration program runs continuously; no review queue is a backlog, it is the size of the program.

• Open scholar collaboration

  IPSC operates a scholar-collaboration program for qualified muḥaddithīn and hadith-science institutions. Collaborators inspect curated review queues, file corrections, and receive attribution in subsequent corpus versions.

• Manuscript-collation roadmap

  IPSC v3.x ships against OpenITI digital editions. Critical-edition collation (al-muqābalah) against original witnesses is a separate program of work on the long-horizon roadmap.

• Mudallis registry expansion

  Tracks 107 of ~152 narrators in Ibn Ḥajar’s Ṭabaqāt al-Mudallisīn. The remaining set requires source-text acquisition beyond the current OpenITI release window.

Qualified muḥaddithīn, hadith-science researchers, and institutional review teams: IPSC welcomes inspection, correction filings, and collaboration on the scholar-review program. Substantiated corrections are integrated with attribution in subsequent corpus versions.

Apply for scholar collaboration