Roadmap

This page lists user-visible planned work and known gaps. It does not cover internal training or evaluation pipelines.

Current release

Nassila 1.2.1 · Masdar UX — audit in-progress panel; DOI↔title conflict manual-only; Copy evidence / Jump to Bibliography; icon polish. See Changelog.

Release train (planned)

App versions use worker-themed codenames. These are planned — not promises with dates.

Version	Codename	User-visible focus
1.2.2	Throughput	Faster audits with sensible limits for registry vs local Sanad LLM
1.2.3	Quote chip	Inline quote-verification chips; cleaner header chrome
1.2.4	Masdar attach	Attach your own PDF per cited reference and re-ground it
1.2.5	Raqim Repair	Stronger bibliography verify/repair (PMCID, arXiv, publisher URLs, parser edge cases)
1.2.6	Raqim Resolve	Suggested matches + manual lookup; gray-lit hosts (HF Hub, Kaggle)
1.2.7–1.2.9	(TBD)	Further polish — not yet scoped
1.3.0	Sharh-lite	Clearer deterministic explanations; Help menu → documentation deep links

Parallel (not a numbered release): scanned-PDF OCR hardening (Maktab O2 fixtures/smoke); Sanad S15+ training when corpus is ready.

Shipped milestones (recent)

Version	Codename	Highlight
1.2.1	Masdar UX	In-progress audit panel; DOI↔title manual resolve; shortcuts; icon I2
1.2.0	Masdar-lite	OA PDF text in grounding; live N/M audit progress
1.1.3	Polish	Notifications; lighter Sanad setup
1.1.2	Raqim Bridge	Send refs to Bibliography; audit from library
1.1.1	Bibliography-first	Curate references before audit
1.1.0	Sanad	Manuscript loop + L3 passage grounding

Worker maturity

Worker	Today	Target
Raqim	Bibliography import, L1/L2, export, bibliography bridge	Continued citeproc and registry improvements
Sanad	L3 in manuscript loop; S12/S14 tiers	Deeper Masdar integration for full-text excerpts
Tasnif / Sharh	Inline deterministic copy	Richer explanations as Sharh matures
Maktab	pdf.js + OCR (O1) in loop	Broader scan coverage / fixtures
Masdar	OA PDF text in grounding (1.2.0); UX polish (1.2.1)	User PDF attach (1.2.4)
Shahid	Disabled	Table/figure evidence (Tier 3+)

Manuscript loop (Tier 3 direction)

Today: Upload/paste → parse References inline → L1/L2 → OA/abstract fetch → Sanad when enabled.

Target: One upload; Maktab segments; Masdar ingests cited PDFs automatically; Sanad runs per cite without manual paste between modules; Shahid for tables/figures.

Webpage and grey-web citations

Planned improvements for messy webpage references (Raqim / Tasnif / Sharh modules):

Baseline and reliability

HTTP status, content-type, redirect logging
Dead-link visibility; paywall / archive hints
Host-specific parsers where stable (prefer code over model)
Feature flags for AI suggestions

OpenAlex consistency when type stays webpage
Locale-aware accessed-date copy
Batch re-fetch metadata suggestions

Model-assisted (future)

When trained facets ship, they may help with:

Scenario	Notes
Suggest CSL fields from page signals	User confirms before apply
Platform / grey-literature typing	YouTube, social, Substack, etc.
Explain fetch failures	PDF URL, Cloudflare, dead link

Bibliography bridge (shipped)

v1.1.2 sends manuscript References to Bibliography and supports audit from Bibliography library with manuscript-ref:N cite keys. See Bibliography bridge.

What we will not promise here

Full reference manager replacement
Cloud LLM as default (local Sanad remains the primary path)
Automatic paywall bypass via Unpaywall email
Training schedules, eval scores, or internal checkpoint gates