Assay-Declared Prediction

Every result is tied to HeLa, firefly luciferase mRNA, the released 35:16:46.5:2.5 formulation and 24-hour mTP readout; in-vivo, tropism and safety claims remain out of scope.

Measured Neighborhood

The result shows its position in 1,100 measured lipids, count-based Morgan local support and five nearest measured analogs instead of collapsing evidence into a generic composite score.

Immutable Result History

Canonical structure, checkpoint version and assay context form a Program-scoped persistence key; stored analysis reopens without losing its model, benchmark or evidence provenance.

Reference Distribution Lanes

Each rank-driving dimension now shows the candidate against every protected reference observation, the reference range, interquartile band and median.

Auditable Score Construction

Per-dimension weights and weighted contributions reconstruct the current rank score instead of asking the reader to infer meaning from radar area.

Evidence and Scenario Semantics

Evidence type, authority eligibility and context-only dimensions are explicit; assumption scenarios are labeled as stress tests rather than calibrated uncertainty.

Reference-Calibrated Radar

The selected candidate now overlays the median of the protected reference panel across all seven historical dimensions, with ranking axes visually distinguished from context-only defaults.

Scientific Metric Rail

The top summary was rebuilt as one restrained instrument rail using cobalt, mineral teal and warm-neutral state cues instead of five heavy gray cards.

Evaluate Once, Reuse Deterministically

Canonical structure, evaluation profile and scorer version form the cache key. Existing artifacts load immediately and repeat submissions reuse the original report.

Persistent Ranked Inventory

Historical cohorts are deduplicated into one inspectable board with protected literature anchors, stored dimension analysis, provenance and current-version ranking.

Recoverable Curation

Names, notes, tags and ranking inclusion are editable. Archive is a soft delete, restore is explicit, reference anchors are protected and computed scores are immutable.

Program-Level Entry

Moved Preflight & Cohorts out of the collapsed capability drawer and into the always-visible Program workflow without adding another module.

Cohort-First Assessment

The assessment page now opens on cohort comparison by default, while explicit links preserve single-structure preflight as a separate task.

Historical Inventory

Candidates now shows selectable cohort, retained structure, reference-anchor and completed-run counts above the Program registry.

Direct Cohort Selection

A compact sample tray selects an entire candidate set or individual structures, searches within the set and runs a bounded comparison for up to 32 candidates.

Historical Run Recovery

Retained score-stage structures from prior molecular-design runs are exposed as read-only cohorts instead of remaining hidden in filesystem artifacts. Active Program CandidateSystems appear as their own cohort when present.

Governed Comparison Boundary

The comparison preserves legacy dimensions, composite and Grade for inspection, but the contract is always DEMO_ONLY/BLOCKED and cannot write a rank, task, CandidateSystem or synthesis decision.

Persistent Program Context

The application shell now owns the active Program. Context is normalized, remembered locally, propagated through canonical links and consumed across evidence, candidates, experiments, decisions, governance, model evaluation and prospective validation.

Decision Control Room

Dashboard no longer acts as a module directory. It reads Program records, identifies the current stage, states blockers and routes the user to the next valid action without converting unavailable data into a passed gate.

Progressive Governance

The work queue reveals review, revocation, scoped application and rollback controls only when server state makes each operation valid. Digest materialization remains available as a collapsed advanced control.

Canonical Product Surface

Review, Legacy Curation, separate Validation Registry and Agent are retired as product destinations. Stable compatibility routes redirect into Work Queue, Evidence, Prospective Validation and Dashboard.

One Product Name

Navigation, Dashboard, page chrome and current documentation now describe one core capability: Candidate Preflight. candidate-preflight.v2 remains the scientific contract behind that product surface.

Explicit Non-Decision Boundary

The Dashboard labels the capability NON_DECISION instead of presenting Validator as a second module. Preflight checks readiness for review; it does not rank, approve or recommend synthesis.

Route Compatibility

The /validator browser route and /api/v1/validator/preflight endpoint remain unchanged, preserving bookmarks and integrations without exposing the legacy product vocabulary.

Five-Gate Readiness Rail

The default Validator experience now presents molecular identity, computation readiness, developability review, CandidateSystem context, and evidence lineage as separate inspectable gates. A failed heuristic remains a review signal rather than silently becoming a scientific decision.

CandidateSystem-Native Input

Users can start with a quick structure intake or validate the full candidate-system.v1 object, including formulation, payload, biology, assay, operating constraints, metrics, and provenance.

Non-Decision API Contract

POST /api/v1/validator/preflight returns candidate-preflight.v2 with explicit NON_DECISION authority, fail-closed blockers, evidence gaps, descriptors, checks, and governed handoff actions. It never returns Grade, rank, composite score, or synthesis decision fields.

Calmer Navigation and Account Surface

The sidebar returns to a 200-pixel workbench width with a more generous row rhythm. The bottom account surface follows the compact avatar/name/settings pattern; sign-out moves into Account settings instead of competing with navigation.

Ordered Pipeline Rail

Ten workflow pages now sit behind one collapsible Pipeline entry and open automatically around the active step. The 01–10 rail communicates the real sequence without presenting every route as a separate top-level product.

Validator as Core Capability

Validator is promoted immediately below Pipeline as a persistent first-level destination with a Core marker, reflecting its current evidence strength and day-to-day utility.

Consolidated Platform and Admin Navigation

Release Notes, Methodology and Architecture are available through one Platform Guide with Updates, Method and System sections. Admin access and activity records move into a protected Settings section while /admin remains a compatibility redirect.

Research-Workbench Density

Sidebar width, navigation rhythm, top bar, page gutters, large headings, cards, tables and form controls use a tighter shared density tuned for information-heavy scientific work.

Feishu Super-admin Boundary

The sidebar and dedicated Admin workspace appear only for the configured Feishu super administrator. The API revalidates the live platform role, demotes password-era admin rows, and returns 403 for every other identity.

Login and Task Audit Ledger

Admin now combines an expandable per-user Feishu login history, failed-login review, and cross-user pipeline and retained historical task records with status, owner, source and timestamps.

genprime-ai GPU Runtime

The TransMA default and persisted platform setting now target the verified genprime-ai Modal endpoint. Known legacy workspace values migrate automatically while unrelated custom endpoints remain unchanged.

Simplified Navigation

The sidebar uses tighter vertical rhythm and aligned horizontal padding, removes the Agent entry, and gives the administrator a dedicated audit destination.

One Session Across Every Module

The frontend establishes one host-scoped Feishu session and every protected API revalidates it against the live user and tenant boundary. There is no per-module password prompt or proxy-supplied demo identity.

Historical Work Preserved Read-Only

Runs, agent threads and plans created before trusted ownership are visible to signed-in members as legacy shared records, but cannot be resumed, cancelled, renamed, deleted or continued through normal routes.

Private Ownership for New Work

New runs, conversations and plans are bound server-side to the verified Feishu user. Owners and super admins can access them; unrelated users receive a non-disclosing not-found response.

Fail-Closed Cutover

Production refuses missing or default signing secrets, OAuth state is browser-bound and single-use, password login is retired, and deployment verifies the public edge before reporting success.

Legacy Replay, Not a Candidate Leaderboard

The results workspace now labels historical triage DEMO ONLY · NON-DECISION, hides legacy Grade, composite and rank, and removes score-derived seed/synthesis actions. Current work routes to Program definition and complete CandidateSystem onboarding.

Twenty-Capability Admission Registry

COMET/LANCE and Helix-mRNA join the registry with pinned source revisions, model/artifact roles and explicit blockers. The current boundary is 15 supporting-only, 5 rejected and zero decision-grade external capabilities.

License Truth at Time of Use

COMET is rejected because its custom agreement prohibits commercial R&D, redistribution and most modification; Helix-mRNA is rejected because the published weights are non-commercial and the task is payload representation, not LNP delivery.

Mechanism Endpoint Method

Methodology v6 and architecture v17 define independent formation, exposure, uptake, escape, damage/recovery, inflammation, function, repeat-dose and process endpoints, with hard gates, conservative bounds, Pareto and information gain instead of a universal score.

Eight-Source Activation Dossier

A strict Program-scoped request binds five JSON cards, an activation-grade dataset package, candidate and rollback artifacts, plus one completed permanent-blind result. Every role requires a distinct Evidence source and the final fingerprint covers source/artifact attestations, exact bytes, blind receipt and logical card hashes.

Isolated Physical Gate Replay

The server ignores caller filesystem locations, safely materializes current custody in a private environment and invokes the existing activation-v2 gate. Duplicate keys, NaN, unknown card fields, unsafe ZIPs, artifact drift and a hand-authored blind slice fail closed.

Large Artifact Streaming Custody

An admin-only route streams candidate and rollback artifacts up to a configurable 2 GiB ceiling, limits bytes during transfer, fsyncs staging, computes server SHA-256 and atomically adopts content. Dossier verification rehashes large files with bounded reads instead of loading model weights into memory.

Append-Only Handoff and Replay

Only a fully passing physical preview can be registered as READY_FOR_INDEPENDENT_ACTIVATION. Every list operation rebuilds the blind result and activation gate from current custody; drift becomes BLOCKED without rewriting history. READY remains NON_DECISION and NOT ACTIVATED.

Promotion Control Plane

The /models workspace now separates external capability admission from model promotion. A custody rail exposes Evidence → Blind → Artifacts → Policy → Handoff, supports real source assembly and large artifact staging, and never seeds a fake dossier or metric.

Pre-Reveal Truth Commitment

An independent custodian freezes the exact truth-file SHA-256 before candidate training. Plan creation binds holdout, endpoint, unit, candidate code/config/artifact, all eight registered baselines and policy thresholds while keeping truth bytes outside platform custody.

Exact Nine-Model Prediction Freeze

The server requires the candidate plus eight registered baselines to predict the identical blind sample set with point and interval estimates. Unknown fields, duplicate keys, missing pairs, identity drift and early truth reveal fail closed.

Three-Party Reveal and Evaluation

Truth can be attached only after predictions and only by the preregistered custodian. Finalization requires a third identity that is neither custodian, trainer nor prediction submitter, and no client-provided metric is accepted.

Server Metrics and Blind-Slice Adapter

The server computes nine-model RMSE, MAE, Spearman, tie-conservative top-k enrichment, interval coverage and calibration with deterministic 95% structure-group bootstrap intervals. The candidate is compared with the strongest frozen traditional baseline.

Custody Rail Workbench

The /evaluations workspace now separates development and permanent-blind evidence. Its commit → freeze → reveal → recompute rail performs real operations, exposes identity boundaries and shows no metric until an actual immutable result exists.

Complete Archive Verification

A deterministic ZIP must contain the exact 8 × 5 run file set. The server rejects traversal, symbolic links, encrypted or duplicate members, unsafe compression and extra files before recomputing nested hashes, folds, predictions, metrics, intervals, code and runtime identity.

Append-Only Evaluation Registration

A zero-write Program preview verifies current Evidence custody. Import replays the archive and records source, archive, run, dataset and all eight model identities in an immutable SQLite record fixed to UNVALIDATED, NON_DECISION and activation_allowed=false.

Time-of-Use Integrity

Every registry read reopens the current custody artifact and repeats full bundle verification. Missing bytes, digest drift, code/runtime mismatch or registration-binding drift becomes a visible BLOCKED state without rewriting history.

Same-Split Evaluation Workbench

The new /evaluations workspace places all eight baselines on one frozen dataset rail, exposes selected metrics and immutable bindings, and keeps activation blockers visible. Empty Programs show no synthetic metrics.

Durable Evidence Custody

Production now host-mounts runtime evidence artifacts and protects the directory from deployment deletion. The API image includes the shared read-only frozen-package reader while Modal training code remains excluded.

Visible Capability Admission

The new /models workspace exposes the reviewed 18-capability registry, current tier counts, exact licenses, deployment/input policies, permitted roles and declared blockers. It reports the current zero decision-grade external model boundary without turning the registry into a model leaderboard.

Intended-Use Preflight

A deterministic API assesses capability, intended role, input classification and endpoint/unit. Remote non-public data, disallowed roles, rejected licenses and non-authoritative candidate-ranking attempts fail closed.

Prediction-Run Lineage

Candidate batch preflight now replays prediction-run.v1 artifacts and verifies run/model/activation/capability identity, endpoint, unit, per-candidate value and uncertainty, canonical activation hash, policy and independent issuer binding.

Proposal Admission Boundary

Production ExperimentProposal creation derives admission from append-only batch receipts and re-verifies current prediction/activation bytes at time of use. Manual drafts and candidates whose activation artifact later disappears are excluded.

Deterministic Batch Preflight

The server replays the custody artifact and reports every schema, context, formulation, evidence, uncertainty, duplicate-ID and duplicate-semantics blocker by row. A preview never writes CandidateSystems.

Evidence-Bound Candidate Metrics

Every Pilot-eligible measured or predicted metric must carry uncertainty and point to a physically verified, reviewed source with a matching evidence class. Predictions must also bind a registered model or run identity.

Atomic Append-Only Import

An administrator submits the preview digest; the server replays the source and writes all CandidateSystems, members and the batch receipt in one transaction. Any conflict rolls the full import back.

Round 1 Execution Funnel

The new /onboarding workspace exposes live custody-batch, READY-system, frozen-proposal and blinded-pack state, plus a neutral JSON template and row-level correction guidance.

Deterministic Five-Arm Pilot Pack

A complete ExperimentProposal can now be replayed into platform, random, diversity-only, uncertainty-only and preregistered human arms. Every policy receives a canonical digest and must fill the same N within the same cost ceiling.

Role-Separated Blinding

Authorized lab custody sees real sample mappings; analysts receive wells, sample type, replicates and blind IDs only. Controls remain blinded in the observation contract.

Server-Derived Results

The independent evaluator uploads a complete blinded JSON artifact. The server verifies physical custody, coverage, controls, QC, units and actual cost before unblinding and generating the immutable result manifest.

Visible Pilot Operations

The new /pilot workspace covers replay, preregistration, pack freeze, role-specific downloads, observation custody and result recording. /validation is now a read-only registry with no placeholder hashes or manual aggregate entry.

Complete Five-Arm Preregistration

Every plan freezes the platform policy, random, diversity-only, uncertainty-only and human baselines against the same candidate pool, candidate count, endpoint and cost ceiling.

Independent Two-Round Evaluation

The planner cannot evaluate their own plan. Both immutable rounds require a server-verified machine-readable result manifest; material deviations, missing custody, digest drift or aggregate mismatch block evaluation.

Predeclared Increment Gate

A plan passes only when the platform arm exceeds every baseline by the preregistered hit-rate delta in each of two evaluable rounds. The summary remains NON_DECISION.

Atomic Governed Target Adapter

Approved application receipts can now create immutable JSON snapshots and atomically advance a real current head. Rollback restores only the exact recorded prior snapshot and preserves append-only history.

Visible Validation & Transition Surfaces

The new /validation workspace manages preregistration, rounds and summaries; /governance now exposes server-side payload hashing, bootstrap, materialization and transition history.

Independent DecisionReview

A DecisionPacket can be approved or rejected only by an authenticated identity distinct from its preparer. The immutable review binds the packet digest, recommendation, rationale and conditions.

Revocation & Rollback Authority

Approved decisions can be revoked without deleting history. Revocation records the reason, affected targets and mandatory rollback actions, and makes the previous approval non-effective.

Scoped LearningReview

LearningUpdates enter a separate review queue. Approval grants only explicit evidence, dataset, model-evaluation or experiment-backlog scopes; unapproved target classes fail closed.

Version-Bound Application Receipts

Learning application requires a third identity, target ID, prior/applied version IDs and SHA-256 digests, plus a rollback reference. Rollback can restore only the exact recorded prior version and digest.

Program Design

A neutral ProgramBrief defines the decision, context, competing hypotheses, constraints, evidence needs and required outputs before candidate selection begins.

Evidence Backbone & Artifact Custody

Program-scoped sources, claims, contexts, evidence edges and citation packs are persisted with digest checks and server-verified physical artifacts.

CandidateSystem & Frozen ExperimentProposal

Candidates are evaluated as complete context-bound systems. Fixed-budget selection, plate layout, controls, replicates, metrics and success criteria are frozen into a replayable proposal.

Execution, Decision & Learning

Immutable execution records feed NON_DECISION evaluations. DecisionPackets require human review, while LearningUpdates remain review-only and cannot automatically mutate evidence, datasets or models.

Data & Model Governance

Partner onboarding, blind custody, traditional baselines, artifact integrity, activation attestations and external capability admission now fail closed when lineage or evidence is incomplete.

Privacy-Safe Curation Page

New Curation entry for in vivo CAR-T delivery intelligence. The page presents target-hypothesis context, internalized prior work, competitive benchmarks, technical literature, review-harness status and upgrade decisions without exposing raw source locations.

Operational Review Workbench

Curation now includes an interactive candidate review console and model lifecycle gate. Users can submit a neutral candidate intake, run the ICL/verification/panel harness, and check whether calibration, adapters, fine-tuning or retraining are allowed.

Candidate Review Harness

New pipeline/review/ package with CandidateIntake, Claim Ledger, ICL gate wrapper, verification gates and static multi-agent review panel. APIs: /api/v1/review/intake/validate, /icl-gate, /verify, /run, /eval/run.

Verification & Eval Gates

Hard gates now catch uncited material claims, predicted-as-measured language, target-finality overclaim, in vivo CAR-T delivery overclaim, weak benchmark context and duplicate-source double counting.

Model Lifecycle Readiness

New pipeline/model_lifecycle/ package and /api/v1/model-lifecycle/* endpoints. Blocks premature fine-tuning or retraining when sample size, endpoint consistency, split strategy, formulation metadata or calibration reports are insufficient.

Architecture v6

New architecture_overview_v6.md documents the decision architecture, privacy-safe curation rules, review harness, eval suite and model lifecycle gate.

Safety & Immunogenicity Assessment

New pipeline/safety/ module with 4-dimensional scoring: cytotoxicity (reactive groups, permanent charge), hepatotoxicity (biodegradability, CYP liability), hemolytic activity (amphiphilicity, charge density), immunogenicity (TLR motifs, complement activation). API: POST /api/v1/safety/assess.

Multi-Tissue Tropism Prediction

New pipeline/tropism/ module predicting organ selectivity (liver, spleen, lung, T-cell, DC) based on SORT mechanism (Cheng/Dilliard 2020-2021), pKa-driven charge ratio, PEG shielding, particle size, and tail saturation. Returns probability distribution + selectivity index. API: POST /api/v1/tropism/predict.

AGILE Fine-Tuning Infrastructure

Data preparation pipeline extracting 2,219 ICL-specific training samples from lipid_master.db (80/10/10 split). Fine-tuning script ready for GPU execution. Activity range: -2.35 to 15.96 log RLU across HeLa + Raw264 assays.

pKa Prediction v2

Replaced linear heuristic with RDKit-based model using LogP, TPSA, MW, Gasteiger partial charges, and EWG counting. Calibrated against SM-102 (6.44 vs 6.68), MC3 (6.61 vs 6.44), KC2 (6.61 vs 6.70). MAE reduced from ~1.3 to 0.32.

SISSO Feature Extraction — Validated

sisso_repro module fully functional: 22-dimensional geometric features from XZ/YZ density maps, 6 symbolic regression models (Model 7/8/9/12/15/17). All models reproduce within <5% RMSE of Lin et al. published results.

Lin 121 Conformation Data Import

Imported 121 lipid conformation records (XZ/YZ density map file paths), 121 transfection activity values (RLU), and 121 SISSO 22-dimensional feature vectors into descriptors table. Database: 1,338 lipids, 121 conformations.

Architecture v4 + Version Switcher

New architecture_overview_v4.md with closed-loop pipeline, pKa v2, SISSO, and deployment sections. Frontend version dropdown defaults to latest, preserves all previous versions (v1–v4).

JWT Auth Removal

Removed all JWT Bearer token references from frontend (agent, formulation, conjugation, validator, settings, mRNA pages). Platform uses HTTP Basic Auth exclusively.

Closed-Loop Seed Extraction

New API endpoint GET /exports/{run_id}/seeds extracts top-scored SMILES from completed runs. Pipeline page 'Seed from Previous Run' selector lets users load candidates from any prior run as seeds for the next iteration — closing the generate→score→seed loop.

Synthesis Candidate Triage

Results page now includes a 'Synthesis Candidates' section. Runs the 7-dimension validator on scored candidates and shows those graded B+ or above with composite score, predicted pKa, and synthesis action. Direct links to seed next run or inject wet-lab data.

Discovery Loop Visualization

Results page shows the full closed-loop flow: Score → Triage → Synthesize → Validate → Next Iteration. Each step links to the relevant platform action.

GCP Docker Deployment

Migrated from ECS to GCP VM (iaso-research-platform) with 3-service Docker stack (FastAPI API, Next.js frontend, nginx reverse proxy) on port 8090. System nginx proxies lnp.iaso-research.com with Let's Encrypt SSL.

FDA Reference Lipid Data

Seeded published activity data for SM-102, ALC-0315, DLin-MC3-DMA, DLin-KC2-DMA: pKa, particle size, PDI, encapsulation efficiency, zeta potential from Hassett 2019, Schoenmaker 2021, Jayaraman 2012, Semple 2010.

Science Audit Fixes

Chemical Space page: replaced mock SAR data with proper awaiting-data state. Library page: removed disabled enumerate button, show real reagent inventory. Basic auth user display in sidebar.

15-Page Smoke Test — All Pass

Complete page-by-page verification: Overview, Pipeline, Runs, Results, Formulation, mRNA, Conjugation, Iterations, Chemical Space, Library, Agent, Architecture, Settings, Login, 404. 0 TypeScript errors, 0 build warnings.

Library Endpoints

New GET /api/v1/library/stats and GET /api/v1/library/reagents endpoints. Queries lipid_master.db for real-time stats (1,216 lipids, ~4,600 activities). Previously returned 404.

CSV Upload Fix (Iterations)

Rewrote /inject-data from JSON body to multipart/form-data UploadFile. Frontend sends CSV via FormData — old endpoint returned 422 on every attempt. Critical for active learning loop.

CI/CD Deployment Pipeline

GitHub Actions workflow with IAP tunnel to GCP VM. SSH key-based deploy, docker-compose.prod.yml with localhost-only port binding, nginx reverse proxy with SSL.

Release Notes Page

This page — platform update documentation with feature descriptions, smoke test reports, and verification records for each release.

Next.js 15 Frontend

Complete migration from Streamlit to Next.js 15 + React 19 + Tailwind CSS 4. 15 pages with consistent design system, sidebar navigation, JWT authentication, and responsive layout.

TransMA Model Training

3D Transformer + Mamba architecture trained for lipid property prediction. Ensemble of 5 models for uncertainty quantification. CPU-only inference strategy for 16GB RAM.

AI Agent Chat

Claude-powered chat interface with thread persistence, suggested prompts, and markdown rendering. Integrated at /agent with animated gradient border effect.

Architecture Documentation

Comprehensive v3 architecture document served via API endpoint. TOC sidebar with intersection observer for scroll-tracking, collapsible sections, anchor navigation.

Formulation DOE Module

LNP formulation design-of-experiments with component percentage sliders, validation, prediction, and optimization suggestions.

mRNA Design Module

Protein-to-FASTA workflow with tissue selector, modification options, poly(A) configuration, quality metrics, and downloadable FASTA output.

5-Stage ML Pipeline

DAG-based pipeline: SISSO feature extraction → LSTM molecular generation → AGILE GNN screening (top 500) → RDKit ETKDG conformer generation → TransMA scoring (top 20).

External Model Integration

Integrated 4 external models: SISSO (Fortran, symbolic regression), AGILE (PyTorch GNN, 60k compound pretrained), TransMA (3D Transformer + Mamba), REINVENT4 (generative). Bootstrap script for automated setup.

FastAPI Backend

REST API with JWT authentication, run management (CRUD + progress + cancel/resume), exports (CSV/SDF), and system status monitoring.

SQLite Database

WAL-mode SQLite at lipid_db/lipid_master.db. Tables: lipids, activities, sources, conformations, iterations, iteration_results, users, settings, agent_plans, agent_threads.

Conda Environment

Reproducible icl-discovery conda env: Python 3.11, PyTorch CPU, PyTorch Geometric, RDKit, OpenBabel, gfortran, openmpi. Single environment.yml for all dependencies.