v3.2.5 — FactorForge

Fixed

• The published Docker image (ghcr.io/eijex/factorforge-cds) crashed on every
  POST /api/optimize request with AttributeError: 'FactorForgeHandler' object has no attribute 'validate_host', returning an empty response. Root cause:
  scripts/serve.py routed requests to the API handler via an unbound-method
  call (handler.do_POST(self)) instead of real inheritance, which broke once
  do_POST started calling self.validate_host / self.send_error_response
  for host-parameter validation. The hosted web app (factorforge.eijex.com) and
  the PyPI CLI were unaffected — only the Docker image and local
  python scripts/serve.py dev server were broken. FactorForgeHandler now
  properly inherits from the optimize API handler.
• CITATION.cff's doi field was left pointing at the v3.2.3 exact-release
  DOI (10.5281/zenodo.20758131) after the v3.2.4 release commit bumped
  version/date-released but not doi. Updated to the v3.2.4 exact-release
  DOI (10.5281/zenodo.20826659), confirmed live on Zenodo. Found while
  updating the FactorForge Paper 1 manuscript's Availability section to
  cite the same DOI.

---

v3.2.4 — FactorForge

Added

• "View Predicted Structure" (AlphaFold DB / ESM Atlas) buttons now show an explicit consent modal naming the actual external operator (EMBL-EBI/Google DeepMind for AlphaFold DB, Meta Platforms, Inc. for ESM Atlas) before the sequence is sent via URL, replacing the previous passive footnote-only disclosure.
• Runtime validation registry (factorforge.validation_registry) listing all 17 validation checks with per-execution-path enforcement metadata, and a canonical validation report builder (factorforge.validation_report) that scans the final returned CDS and reports results keyed by check_id. The web "Sequence Checks" panel and the /api/optimize response now expose all 9 advisory scanners plus restriction-site and MoClo-overhang results (previously only 3 of 9+ checks were visible). validation.moclo/polya/gc legacy fields are unchanged.
• reproducibility/benchmark_v0.5.1/scripts/figures/make_cai_vs_multiconstraint_figure.py regenerates the "CAI-only baseline versus constraint-aware design" scatter figure (mean CAI vs. corrected multi-constraint pass rate per method) from benchmark_summary.frozen.json. No generation script for this figure existed anywhere in the repository or its Git history; this is a reconstruction verified against the figure's existing per-method values, not a recovery of the original code.

Docs

• Clarified that BY-2 (N. tabacum) experimental host support applies to
  the balanced/gc_target/assembly_friendly rule-based profiles only;
  high_cai and feasibility_best remain N. benthamiana-only by design.
• docs/how-it-works.md's "Design Objectives" table listed gc_target/
  high_cai as if they were working DP --objective values; passing either
  to --objective always raised (the DP engine only implements
  feasibility_best). Split the table into the DP engine's one real
  objective and the profile engine's --profile values, with an explicit
  note that the latter are not DP objectives.
• docs/profiles.md's "Stable Profiles" section claimed all four profiles
  are "fully supported ... via CLI, Python API, web app, and MCP" without
  disclosing that high_cai is N. benthamiana-only and rejected for other
  hosts at the CLI/REST/web layers. Added that qualifier.

Fixed

• Web UI "AA Preserved" badge in the Optimization Results panel displayed
  validator_status (assembly/restriction-site review outcome) instead of
  actual amino acid identity — a 33aa test sequence with a flagged
  restriction site showed "⚠️ Review" even though its translated protein
  was 100% identical to the input. The API already exposed the correct
  value as constraint_report.aa_identity; getPrimaryResult() in
  web/js/app.js now reads that field instead. No API contract change.
• Success/info toast notifications (e.g. "History item loaded") were nearly
  invisible in dark mode — the generic .dark .bg-emerald-50/.bg-blue-50
  card-dimming rules turned the toast background into a low-alpha tint while
  the dark text color was left unchanged. Scoped an opaque, high-contrast
  override to #toastContainer only; light mode and other card UIs (e.g.
  the privacy banner) are unaffected.
• /api/optimize/compare and /api/optimize/batch accepted a host or
  host_profile field in the request body but never read it, silently
  ignoring the caller's host intent and returning HTTP 200 with
  default-host output. Both endpoints now reject any request containing
  either field with HTTP 400 (HOST_NOT_SUPPORTED_ON_ENDPOINT) instead of
  silently dropping it.
• CLI --objective declared gc_target and high_cai as valid DP
  objective values via click.Choice, but the DP engine only ever
  implements feasibility_best — passing either always raised
  DP engine currently supports --objective feasibility_best.. Removed
  both from the click.Choice so the CLI rejects them at the argument-parsing
  stage with a clear "invalid choice" error instead of failing inside the
  engine.
• A direct library call (RuleBasedOptimizer().optimize(profile="high_cai", host=<non-default>)) silently substituted N. benthamiana golden-set
  output with no indication that the requested host was ignored (the
  existing host/strategy compatibility guard only covers the REST/CLI/web
  surfaces). Added a logger.warning at this call site; the host-invariant
  output itself is unchanged (this is the documented design boundary, not a
  bug) and the existing benchmark
  codon_table_path injection path is unaffected.
• Web UI "View Predicted Structure → ESM Atlas" link used a URL scheme
  (esmatlas.com/explore?tab=fold&sequence=) that ESM Atlas's router never
  branches on, so it always landed on the generic explore screen instead of
  a fold view regardless of sequence. Switched to the route ESM Atlas's own
  navigation actually uses (esmatlas.com/resources?action=fold&sequence=),
  confirmed against ESM Atlas's production JS bundle and verified live —
  it now lands on the correct "Fold Sequence" tool (ESM Atlas's UI does not
  support pre-filling the sequence box from a URL parameter, so users still
  paste it manually after arriving).
• Removed internal task-tracking references ("Job NNN") from the public changelog (CHANGELOG.md, docs/changelog.md, web/index.html) and from public docs (docs/strategy/eijex-tool-layer-classification.md, docs/validation/RELEASE_GATE.md, reproducibility/benchmark_v0.5.1/README.md). scripts/audit_public_surface.py's internal_reference pattern now also matches Job \d+ so future releases catch this automatically.
• Removed the same internal task-tracking references ("Job NNN", "analysis NNN") from source-level comments, docstrings, and registry provenance fields (src/, tests/), and deleted benchmarks/scripts/resume_job130_rerun.sh, a one-off maintainer rerun script that hardcoded a local developer machine path. Empirical citations (CAI/GC benchmark numbers) were kept; only the internal ID labels were removed.
• /api/optimize now rejects explicit objective=feasibility_best or
  profile=high_cai requests combined with a non-default host (HTTP
  400) instead of silently returning N. benthamiana-table output. The
  implicit case (host-only request, no explicit strategy) now discloses
  requested_strategy/resolved_strategy/resolution_reason and
  resolves to balanced (previously silently resolved toward high_cai,
  which is also N. benthamiana-only). CLI and web UI now block high_cai
  for non-default hosts (--compare-profiles included), matching the
  existing feasibility_best guard; the web UI's auto-selected fallback
  for BY-2 changes from high_cai to gc_target.
• Release provenance hashing now computed from the committed git blob (git show HEAD:<path>) instead of local working-tree bytes, fixing CRLF/LF drift on Windows that could silently produce incorrect SHA-256 values in reproducibility/benchmark_v0.5.1/MANIFEST.json and tests/test_docs_consistency.py.
• Public-surface DOI references (README.md, docs/index.md, AGENTS.md) switched from version-pinned Zenodo DOIs to the concept DOI, which always resolves to the latest release, so future releases no longer require manually updating these files.
• AGENTS.md no longer states a hardcoded "16 version-bearing files" count, which had gone stale; points at scripts/release.py's build_targets() as the source of truth instead.
• docs/rule-engine-roadmap.md, docs/validation.md, docs/how-it-works.md, and docs/factorforge-architecture.md enumerated only 5-8 of the 9 default advisory RuleEngine scanners and omitted the MoClo overhang assembly-review check; all four now list the complete set. rule-engine-roadmap.md additionally mis-stated "Repeat patterns" as "Planned / Not yet implemented" (it is implemented and runs by default) and described an unused legacy GC-window calculation instead of the active scan_gc_extremes thresholds; both corrected. No runtime code changed.
• Web app "Sequence Checks" badge labeled "MoClo Overhang Check" actually reported a Type IIS restriction-site scan result, not MoClo overhang validity (the real MoClo overhang check lives in the opt-in construct-builder path and was never called here); label and changelog text corrected to "Restriction Site Check (Type IIS)". validation.moclo JSON field name kept unchanged for frontend compatibility.

Added

• scripts/audit_public_surface.py and a CHANGELOG duplicate-[Unreleased]-header check now run on every push/PR in CI, instead of only when release.py --auto --audit-script is remembered.
• scripts/regen_manifest.py regenerates reproducibility/benchmark_v0.5.1/MANIFEST.json input hashes from committed git-blob content on demand.

Changed

• Repositioned FactorForge as a claim-bounded pre-synthesis review harness across README.md, ROADMAP.md, docs/index.md, the new docs/roadmap.md, and web/index.html, separating the research-software journey from product roadmap themes without adding any new guarantees (expression, glycosylation, folding, yield, synthesis acceptance, regulatory approval).

---

v3.2.3 — FactorForge

Release v3.2.3

v3.2.2 — FactorForge

Fixed

• multi_constraint_pass definition corrected (scoring_contract v1.1): benchmarks/scoring.py
score_cds() now defines multi_constraint_pass = biological_pass AND assembly_pass AND gc_in_target_range.
  The previous definition (biological_pass AND assembly_pass) omitted GC target compliance, producing
  inflated L3/L4 ablation values (89.0%/88.6%) that were mathematically inconsistent with their GC in range
  rates (3.7%/5.8%). Corrected values: L3=3.5%, L4=5.6%.
  All benchmark artifacts (benchmark_summary.json, ablation_summary.json, benchmark_v0.5.1 data/figures)
  regenerated from full rerun (seed=320, N=49,257). Zenodo benchmark_results.csv v2 (DOI: 10.5281/zenodo.20676276) supersedes v1.
  A canonical_multi_constraint_pass() helper added for recomputing from primitive columns in historical CSVs.
• Benchmark: source-profile codon-table injection now flows into both the design and scoring paths (design_table_sha256 == score_table_sha256 verified per run), fixing a prior gap where design always used the default table regardless of an injected profile (Job 130).

Added

• Data: added three genome-annotated N. benthamiana codon-usage profiles (SGN QLD183 v103 CDS-derived; SGN NbeV1.1 all-CDS-derived; SGN NbeV1.1 high-confidence-CDS-derived) built via scripts/build_codon_profile.py under strict_nuclear_cds_v1 filtering, alongside the existing packaged reference profile (Job 130).

Changed

• Web: Host System cards in web/index.html are now rendered dynamically from GET /api/optimize (supported_hosts + new host_metadata field) instead of being hardcoded, removing a 3-way duplication between the HTML, web/js/app.js, and the API (Job 133).

Docs

• Aligned public wet-lab validation contribution language with manual-review, public-safe submission rules.
• Clarified that public GitHub Issues must not contain raw sequences, confidential construct details, internal batch IDs, patient data, private contact information, exact process parameters, or confidential partner/customer data.
• Aligned public README, docs, web, citation, packaging, roadmap, and benchmark wording with the in-silico CDS design claim boundary.
• Removed maintainer-local file paths and internal repo references from docs/release-checklist-template.md, replacing them with generic placeholders (Job 134).

---

v3.2.1 — FactorForge

Added

• Protein risk annotation layer for CDS sequences
  • Transmembrane helix prediction (Kyte-Doolittle, window=19, threshold=1.6)
  • Signal peptide heuristic (N-terminal 30 aa scan)
  • Risk classification: HIGH / MEDIUM / LOW / UNKNOWN

Fixed

• Correct CAI provenance annotation in benchmark output (Job 110)
• Correct Type IIS restriction site warning status (Job 110)
• Pin JSON files to LF line endings for Windows reproducibility
• Add pandas to dev test dependencies
• Manifest SHA-256 reproducibility drift on Windows (JSON/EOL normalization)
• wet-lab result GitHub template: add protein_class options (Reporter / Antigen / Cytokine) and validation consent checkbox

Changed

• Add Google Form as wet-lab submission channel alongside GitHub Issue and email
• Standardize wet-lab submission link labels across README, docs, and web app

Docs

• Add public-safe FactorForge agent guidance
• Expand release checklist with public surface audit steps

Chore

• Bump actions/configure-pages, codecov/codecov-action, actions/deploy-pages, actions/setup-python, softprops/action-gh-release (Dependabot)

---

v3.2.0 — FactorForge

Added

• MFE metadata fields — Design Package and API response now include mfe_used (bool), mfe_status (computed / not_computed), and mfe_warning (string when ViennaRNA unavailable). score_components added to expose per-term weights used in composite score calculation.
• Design Package schema v1.0.0 — Formal IUPAC/FASTA I/O contracts and MFE null invariant established (090).
• Registry production constants export — DEFAULT_CAI_TARGET, DEFAULT_GC_LOW, DEFAULT_GC_HIGH importable as public production constants (091).
• Benchmark seed injection — --seed flag for deterministic reruns; most_frequent_codon tie-breaking deduplication (099).
• Codon table provenance disclosure — codon_table_manifest.json with sha256 pin, build_path_status: incomplete, and known limitations for nbenthamiana_codons.json (097).

Fixed

• Domestication Silence Fail — pipeline.py now raises ValueError when restriction-site domestication fails (previously returned the undomesticated sequence silently as success).
• Pipeline Output Validator — validate_cds_output() is now called in pipeline.py before final sequence return, catching AA identity violations and internal stops at the pipeline level.
• MFE not-computed value — mfe_kcal_mol is now null (not 0.0) when ViennaRNA is unavailable. Composite score is unchanged; this corrects misleading metadata only.
• Input validator — IUPAC ambiguous DNA/AA sequence misclassification corrected (098).

Documentation

• Stale constant corrections — 5 doc/comment locations corrected to match live code.
• Claim wording alignment — Public-facing API and CLI output wording unified; no expression-level or yield improvement claims (092).
• Formal benchmark — N. benthamiana SGN CDS (N=49,257, seed=320). All metrics are in-silico; no wet-lab validation claimed.

---

v3.1.9 — FactorForge

Documentation

• Internal housekeeping — project tracking references updated. No engine changes.

---

v3.1.8 — FactorForge

Breaking Changes

• gc_target profile default changed — calling gc_target without an explicit target_gc now produces sequences targeting ~60% GC (host midpoint) instead of the previous 42.5%. If you relied on the 42.5% default, pass target_gc=42.5 explicitly to preserve the old behavior.

Changed

• gc_target profile default — now targets the host-profile GC midpoint (60% for N. benthamiana) when target_gc is not supplied, instead of the legacy hardcoded 42.5%. To target a lower GC, pass target_gc explicitly. Output sequences from gc_target without an explicit target will differ.
• GC scoring — calculate_composite_score now scores GC using a band function (gc_band_score): full score inside [gc_min, gc_max], linear decay outside over gc_decay_width (default 20 pp). Replaces the previous 1 - |GC - GC_opt|/50 proximity formula, which under-discriminated GC quality.
• assembly_friendly scoring weights — changed from balanced-identical (0.5, 0.3, 0.2) to (0.3, 0.4, 0.3) (lower CAI pressure, higher GC/MFE weight) to align scoring with its Type IIS site-avoidance translation strategy.
• feasibility.py defaults — target_cai 0.92 → 0.82 (achievable; aligns with industry >0.8 practice); target_gc 41–44% → 55–65%; fallback GC ranges realigned to the 55–65% output distribution.

Fixed

• Homopolymer thresholds documented — expression-stability (≥6 nt) and synthesis/manufacturing (≥8 nt) scans now use named constants (HOMOPOLYMER_EXPRESSION_WARN_NT, HOMOPOLYMER_SYNTHESIS_WARN_NT) and emit context/threshold_nt metadata so the two intentionally different thresholds are no longer mistaken for a bug.
• Misleading docs removed — gc_target no longer described as "42.5% (N. benthamiana optimal)"; 42.5% was a legacy assumption inconsistent with the 55–65% codon-table output.
• CLI docs corrected — docs/cli.md --gc-min/--gc-max defaults fixed from 40/55 to the actual 55/65.

Documentation

• docs/profiles.md — added missing assembly_friendly profile; corrected gc_target description.
• docs/tutorials/gfp-nbenthamiana.md — regenerated profile-comparison metrics under the new GC scoring and gc_target default.

---

v3.1.7 — FactorForge

Added

• Web UI host selector — expression host toggle (N. benthamiana / BY-2 Experimental) in the input panel. BY-2 selection disables Feasibility Best objective and shows experimental warning. Result panel displays active host profile.
• E2E smoke tests — 5 Playwright smoke tests covering UI load, protein input, BY-2 host toggle, Feasibility Best guard, and result rendering. Runs automatically after each deployment via e2e.yml.

Documentation

• Eijex MCP access — added Eijex MCP as access option in README.md and docs/index.md
• API endpoints — added POST /api/optimize, /compare, /batch endpoints section to docs/cli.md
• MCP getting started — added Eijex MCP connection guide to docs/getting-started.md
• ml_enhanced profile — docs/profiles.md에 ml_enhanced 프로파일 문서화
• AGENTS.md — 새 API 엔드포인트 추가 시 eijex-mcp 동기화 항목 명시

---

v3.1.6 — FactorForge

Added

• SynCodonLM scoring dimension — optional 5th composite score component (w_syncodonlm, default 0.0). Integrates Boehringer-Ingelheim's BERT-based codon language model (SynCodonLM, NAR 2025; HuggingFace: jheuschkel/SynCodonLM-V2). Graceful fallback (score 0.5, WARNING) when transformers is not installed. No change to existing scoring behavior.
• ml_enhanced scoring profile — w_cai=0.35, w_gc=0.25, w_mfe=0.15, w_syncodonlm=0.25. Opt-in; existing four profiles unchanged.
• [ml] optional dependency group — pip install factorforge-cds[ml] installs transformers>=4.40 and torch>=2.0 for SynCodonLM inference.
• scoring_ml.py — SynCodonLMScorer class with lazy model loading; calculate_syncodonlm_score(sequence, organism).
• Profile comparison mode — factorforge optimize input.fasta --engine profile --compare-profiles balanced,high_cai,gc_target outputs a side-by-side CAI / GC% / score table. First profile result saved to --output when specified. POST /api/optimize/compare endpoint added with same functionality via JSON API.
• Tutorial: GFP N. benthamiana — end-to-end worked example at docs/tutorials/gfp-nbenthamiana.md. Covers CLI, Python API, profile comparison, and MoClo assembly preparation.
• Batch optimization API — POST /api/optimize/batch accepts up to 20 sequences in a single request. Returns per-sequence CAI, GC%, score, and optimized CDS. Auto-generates IDs (seq_1, seq_2, ...) when omitted. CLI multi-FASTA was already supported.
• Tobacco BY-2 host support (experimental) — --host by2 CLI flag and "host": "by2" API field optimize for N. tabacum BY-2 suspension culture cells using a Kazusa-derived codon table (1,534 CDS, species 4097). Default host remains nbenthamiana. CAI difference between hosts is < 0.05. Experimental: uses N. tabacum codon usage as proxy; not wet-lab validated for BY-2 expression performance.
• Structure prediction links — AlphaFold DB and ESM Atlas fold links appear in the result panel after optimization. No API calls — links open external services with the input sequence.

---