Field Notes

A workbench you can arrange, with tested science underneath

Start here

• Open the Observatory workbench. If you are not signed in, Observatory will take you to its first-party sign-in screen and return you to the workbench afterward.
• Approved collaborators can create or return to their account with their approved email address and a one-time verification code.
• This remains a private research preview. Observatory is not for diagnosis or treatment decisions.

Release ledger

• Current delivered edition: Stable 5, published to production on 2026-09-20 through PR #104.
• Application package version: 0.1.0. Field Notes editions describe product releases without implying that the application has reached 1.0.
• Included release boundary: `24fa243..b48c429`.
• Reviewed staging cutoff: merge `b48c429`. Production release: merge of PR #104.
• Contribution count: 3 production promotion PRs + 1 scientific core PR + 9 stale PRs closed = 13 PR actions since Field Notes 04.

The bridge from Field Notes 04

• Field Notes 04 ended with the first scientific surfaces, EXO intake, a unified omnibar, and staging under EXO. Field Notes 05 turns the workbench into a configurable grid of scientific surfaces you can arrange yourself, replaces the last prototype heuristic with a tested calculation, adds the first offline Python pipeline, closes four of five open roadmap issues, and cleans up all stale draft PRs.

Progress against the MVP

The source of truth remains the

original Observatory MVP plan.

These are directional product-completion estimates, not biological validation

scores.

What changed

• The workbench is now a surface grid you can arrange. The fixed evidence layout is gone. The workbench is a 12-unit grid of draggable, resizable scientific surfaces — add them from a catalog, rearrange by drag or arrow keys, resize with handles, and every surface keeps its source, provenance, and limitations visible. Six surfaces ship now:
  • Gene and genomic neighborhood — RefSeq annotations around the locus, with an embedded IGV.js genome browser.
  • Gene structure overview — exon-intron geometry with CDS highlighting and MANE Select transcripts; clicking an exon shares the selection across surfaces.
  • CDS and amino-acid translation — local RefSeq coding sequence translated against the reviewed UniProt protein, residue by residue.
  • Promoter and proximal regulation — ENCODE cCRE elements and transcription-factor signals within ±2 kb of the TSS, with explicit limits.
  • Genomic sequence — the exact bounded GRCh38 reference, exons in uppercase, paged by transcript coordinates.
  • Protein overview — reviewed UniProt features along the backbone, with AlphaFold linkage.
• Your arrangements are saved. Layouts persist as named views through a Convex-backed store with revision tracking and conflict detection, and can be exported as JSON or shared as target links.
• The last prototype heuristic is gone. PEST regions are now computed with a tested implementation of the Rogers et al. (1986) PEST hypothesis — region count, max score, and coverage, validated by nine tests, shown everywhere PEST data appears (UI, reports, evidence bundles, assistant).
• Bug and feedback intake moved to EXO. Signed Convex boundary replaces Linear; drafts stay explicit and are never auto-filed.
• Engineering hygiene, for the record:
  • Offline Python pipeline for the five-gene fixture — SHA-256 source checksums, CDS translation validation, deterministic JSON/Markdown export, checkpoint resumability. Thirteen Python tests pass.
  • Four roadmap issues closed (#2 invite-only auth, #3 MANE Select resolution, #4 tested sequence signals, #5 offline fixture). Nine stale draft PRs closed. Zero open PRs, zero open issues.

What to try

• Open the workbench and search TP53 — six surfaces load with real data.
• Drag a surface by its handle. Resize it. Your layout autosaves; save it as a named view.
• Click an exon in gene structure and watch the selection propagate to genomic sequence and promoter regulation.
• Check the PEST regions in any report — structured data from a tested algorithm, not a percentage guess.

What's next for Field Notes 06

• Integrate the pipeline into the app. The offline Python fixture exists and is tested; ingestion and export are not yet wired into the workbench runtime.
• Local AlphaFold assets. Structures still load from the remote AlphaFold service; local acquisition with checksums and license metadata is not built.
• Broader ingestion. The five-gene fixture is the only offline slice; GENCODE/MANE, ENCODE cCRE, and disease/miRNA sources are not ingested yet.
• More surfaces. The catalog's planned surfaces (notes, report builder, metadata, wiki) are not on the grid yet.

Code Details

Surface workbench and scientific surfaces

• PR #92 · unified omnibar
• PR #93 · repair workbench geometry and chat evidence continuity
• PR #94 · dense real-data gene and protein viewer foundation
• PR #97 · configurable surface workbench architecture
• PR #99 · EXO intake cutover
• PR #100 · first three scientific surfaces
• PR #101 · regulation, sequence, and protein surfaces (IV-174)

Scientific core completion

• PR #102 · promote Field Notes 04 + 05 to production
• PR #103 · PEST-region detection + five-gene offline Python fixture

Stale PR cleanup

• Closed PRs #95, #98, #91, #90, #82, #81, #58, #18, #16 — all superseded by merged work in PRs #97–#103.

Roadmap issues closed

• #2 · invite-only authentication · Clerk integration live
• #3 · MANE Select resolution · implemented for five fixture genes
• #4 · sequence signal calculations · PEST-region detection replaces prototype
• #5 · five-gene offline Python fixture · pipeline/ package created

Evidence you can inspect, and a simpler way in

Start here

• Open the Observatory workbench. If you are not signed in, Observatory will take you to its first-party sign-in screen and return you to the workbench afterward.
• Approved collaborators can create or return to their account with their approved email address and a one-time verification code. A separate invitation link and mandatory multi-factor setup are not required.
• This remains a private research preview. Observatory is not for diagnosis or treatment decisions.

Release ledger

• Current delivered edition: Stable 4, published to production on 2026-09-04 through PR #73, with the landing-page hotfix completed through PR #77.
• Application package version: 0.1.0. Field Notes editions describe product releases without implying that the application has reached 1.0.
• Included release boundary: `c4d0886..24fa243`.
• Reviewed staging cutoff: merge `84a1543`. Production release: merge `24fa243`.
• Contribution count: 14 implementation and operational PRs + 1 Field Notes delivery PR + 3 production promotion PRs = 18 merged PRs since Field Notes 03.

The bridge from Field Notes 03

• Field Notes 03 established the five-gene Scientific Core, cost-aware AI routing, bounded interface actions, a more coherent workbench, and the first Exobase delivery record. Field Notes 04 makes those foundations more useful: scientific files become inspectable evidence, protein structure confidence and provenance become visible, AI answers carry sources, the workbench interaction system is tighter, and approved collaborators can provision themselves directly.

Progress against the original MVP specification

The source of truth remains the

original Observatory MVP plan.

These are directional product-completion estimates, not biological validation

scores. The change column measures progress since Field Notes 03.

What changed

• Approved collaborators can now create an Observatory account directly with an approved email and one-time verification code. Clerk verifies the identity, then Observatory enforces the exact approved-email boundary, so access no longer depends on locating an invitation email.
• Scientific files are treated as first-class conversation evidence. FASTA, structured text, and other supported inputs receive bounded format inspection, limitations, and retention status without storing the raw file.
• Protein structure inspection now shows per-residue AlphaFold pLDDT confidence, a readable four-band distribution, model and database versions, file integrity, licensing, and explicit interpretation limits.
• Assistant responses can carry source citations back into the evidence field, making it easier to move from an answer to the underlying scientific record.
• Observatory chat now uses a live OpenRouter-backed model path with an approved low-cost candidate set, a one-cent request ceiling, and a ten-dollar monthly testing cap. The evidence fallback remains available for provider failures.
• The workbench interaction system has been tightened across the convo list, omnibar, evidence navigation, resizable panels, responsive tables, and scientific visualization controls.
• Observatory now has a fail-closed Exo ID authorization adapter and staging deployment evidence, while retaining the app-level approved-email gate until shared EXOBASE sessions and just-in-time grants are actually operational.

What to try

• Sign in at the workbench link above with an approved email address; confirm that the one-time verification-code flow returns you to the workspace without an invitation step or mandatory MFA setup.
• Open the TP53 protein fold and inspect the pLDDT confidence distribution, model version, checksum, license, and linked AlphaFold record.
• Attach a small scientific file and inspect what Observatory records, what it does not retain, and how that evidence persists with the convo.
• Ask a scientific question and follow the cited sources rather than treating the assistant response as unsupported prose.

What is already forming for Field Notes 05

• Replace transitional app-specific eligibility with audited EXOBASE just-in-time provisioning for the exact app:observatory grant, without making today’s approved-user access depend on unfinished shared SSO work.
• Turn bounded agent actions and scientific intake into an inspectable workflow runner that can connect approved tools while preserving intermediate evidence, parameters, and provenance.
• Add durable, versioned scientific indexing beyond the current five-gene slice and independently validate calculations outside the interface.
• Run a researcher-question benchmark against the live assistant, including clarification, off-topic redirection, unsafe biological requests, citation quality, latency, and cost.

Code Details

Scientific intake and workbench completion

• PR #60 · typed scientific file intake · merge `c58e9a3`
• PR #61 · finalized workbench interaction system · merge `08fd466`
• PR #62 · first-class scientific inspection cards · merge `eee41df`

Confidence, provenance, and sourced answers

• PR #63 · confidence-aware AlphaFold inspection · merge `6472a5a`
• PR #64 · mobile confidence-overlay polish · merge `6190266`
• PR #65 · versioned AlphaFold provenance · merge `4172857`
• PR #66 · source citations in assistant answers · merge `209d54e`

Identity, delivery, and approved access

• PR #67 · fail-closed Exo ID App grants · merge `96946d8`
• PR #68 · Railway checksum-build recovery · merge `f967a00`
• PR #69 · direct approved-account provisioning · merge `1e2e47f`
• PR #71 · preserved first-party authentication return paths · merge `3e4b3be`

Live routed AI

• PR #72 · live OpenRouter assistant path with bounded model policy · merge `9f4bb67`

Stable 4 delivery and landing-page hotfix

• PR #70 · Field Notes 04 renderer and owner-preview delivery · merge `2017192`
• PR #73 · Stable 4 production promotion · merge `b8ac984`
• PR #74 · landing-page density, hierarchy, and footer refinement · merge `0a023d9`
• PR #75 · first production landing-page hotfix · merge `8cafb87`
• PR #76 · headline-wrap and hero-density correction · merge `84a1543`
• PR #77 · final production headline hotfix · merge `24fa243`

A more capable Observatory, without losing the evidence

Release ledger

• Current delivered edition: Stable 2, published from Field Notes 02 on 2026-08-15 at merge `8be8e26`.
• Release edition: Stable 3, approved for owner-preview and collaborator delivery on 2026-08-23.
• Application package version: 0.1.0. Field Notes editions describe product releases without pretending the pre-release application has reached 1.0.
• Included staging boundary: `8be8e26..cd3610d`.
• Reviewed staging cutoff: merge `cd3610d`. Production release: merge `8f4c9fc`.
• Contribution count: 21 implementation PRs + 3 production-release PRs = 24 merged PRs between the Stable 2 boundary and this Stable 3 candidate.

The bridge from Field Notes 02

• Field Notes 02 ended with a persistent conversation-centered workspace, saved analyses, comparisons, portable evidence bundles, responsive controls, and first-party authentication. Field Notes 03 begins at that exact boundary and adds the scientific identity layer, real-time backend foundation, evidence-grounded AI routing, bounded agent actions, inspectable methods, offline five-gene evidence, reliability work, the public organization page, and Exobase-managed delivery metadata.

What changed

• Observatory now feels much closer to a real scientific workbench. The conversation, active workspace, evidence panels, protein structure, source links, saved analyses, and exports work as parts of one research surface instead of a collection of separate demos.
• The assistant path can answer broader biological questions while staying grounded in the evidence available to Observatory. When a live model is configured, a cost-aware router chooses between faster and deeper paths, applies privacy and spending limits, and falls back to deterministic evidence when model access is unavailable.
• The assistant can now propose bounded interface actions, such as navigating to a target or opening an evidence view. Actions remain inspectable, and consequential method runs still require confirmation rather than happening invisibly in the background.
• The scientific core is more reproducible. Five initial genes now carry checksummed local MANE Select coding sequences and reviewed UniProt context, with validation that fails closed when source identities or translations do not agree.
• The workbench loads more reliably. Optional scientific lookups no longer hold the entire interface in a skeleton state, conversations keep a readable fixed rhythm, and the responsive layout has been tightened for smaller screens.
• Observatory now has a public organization page and waitlist, a cleaner first-party sign-in experience, and an Exobase-managed control-plane record for its Page, runtime, deployment evidence, and release boundaries.

What to try

• Ask a general question about TP53, BRCA1, CFTR, EGFR, or DMD, then follow the response into the underlying genomic, protein, or structural evidence.
• Ask Observatory to move to another target or open a relevant evidence view, and notice that the proposed interface action remains visible and bounded.
• Save an analysis, return to it later, and export the evidence trail rather than relying on the conversation alone.

What's up next

• Pressure-test the assistant with real researcher questions, including cases where it should ask for clarification, redirect an off-topic request, or refuse unsafe biological optimization.
• Turn inspectable methods and agent actions into a coherent workflow system that can connect approved scientific tools without hiding intermediate evidence or provenance.
• Expand the versioned evidence pipeline beyond the initial five-gene core and keep improving protein-first exploration, source linking, and portable research projects.
• Use invited-user sessions to prioritize the remaining MVP workbench gaps before broadening access from the waitlist.

Code Details

Scientific identity and integrated product foundation

• PR #30 · locked source metadata and gene identity · merge `1088c0e`
• PR #34 · Convex real-time reference backend · merge `ef3bdd7`
• PR #35 · conversation benchmark and Exobase AI findings · merge `6530b27`
• PR #36 · public landing page and waitlist foundation · merge `f8f0e1a`
• PR #33 · guided intake and integrated assistant refresh · merge `626f954`

Landing, authentication, and workbench refinement

• PR #38 · landing-page hierarchy · merge `1a196c9`
• PR #40 · responsive workbench controls and grid · merge `c421f80`
• PR #41 · responsive authentication surfaces · merge `e61ac78`
• PR #42 · convo-centered workbench · merge `7f77a11`
• PR #43 · standardized panels and evidence inspector · merge `f3f0080`
• PR #44 · bounded scientific lookup loading · merge `3a19a4f`
• PR #45 · removed optional-source latency from initial render · merge `60ac099`

Scientific workflows, AI, and reproducible evidence

• PR #46 · scientific objects, methods, and inspectable runs · merge `82c26b6`
• PR #47 · fail-closed evidence-based AI routing · merge `89e57a9`
• PR #48 · bounded agent-controlled interface actions · merge `83687b4`
• PR #49 · local five-gene coding sequences · merge `2121c51`
• PR #50 · refined purple scientific workbench · merge `e442d0e`
• PR #51 · offline reviewed UniProt context · merge `b4b83f2`

Organization page and Exobase delivery

• PR #52 · Structured Liquidity organization landing page · merge `0b972db`
• PR #53 · Exobase v1.6 enrollment · merge `a64fa82`
• PR #55 · tighter mobile landing-page narrative · merge `cd3610d`

Production release PRs

• PR #37 · public landing and workbench release · merge `0d77392`
• PR #54 · Exobase-managed Observatory release · merge `4098ab4`
• PR #56 · mobile landing-page release and Stable 3 cutoff · merge `8f4c9fc`

Research that remembers where you left off

What changed

• Observatory is now one continuous research workspace. Conversations, genomic context, protein structure, source provenance, comparisons, and report drafts stay together instead of living in separate gene screens.
• Conversations can be searched, grouped into projects, moved between folders, completed, archived, and restored, making it easier to pause an investigation and return to it later.
• Scientific workspaces can now be saved and reopened with their active target, comparison target, visible evidence modules, conversation history, and source versions intact. The saved-workspace library also supports renaming and archiving older work.
• Reports now have concise and detailed preview formats. Portable evidence bundles include SHA-256 integrity checks, while deterministic rerun audits distinguish source-version changes from changes in the evidence itself.
• The workbench is cleaner at a glance: navigation and content align to one surface, sidebar controls use familiar icons, and the About page now carries the deeper product and reference-data context.
• Field Notes delivery now keeps the owner preview, release approval, and broader collaborator delivery as separate checkpoints.
• Please sign in again using the private review link below so your session uses Observatory's updated first-party access flow. If you have not created an account yet, use the newest invitation email first; returning users can sign in with the method already connected to their account.

What's up next

• Learn from the first invited testers: where research context gets lost, what feels trustworthy, and what needs to be clearer before sharing a report.
• Replace the remaining lightweight intent responses with a model-backed, evidence-grounded assistant that explains the sources behind its conclusions.
• Deepen collaborator-ready reports and comparisons without turning prototype heuristics into claims of biological validation.
• Expand the versioned evidence pipeline and make reruns easier across local, portable, and hosted workspaces.

Code Details

• Unified workbench and Observatory staging origin: PR #15, PR #17
• Field Notes delivery and release gates: PR #19, PR #23
• Saved workspaces and deterministic comparison: PR #24
• Portable evidence bundles and rerun audits: PR #25
• Conversation-centered workbench: PR #27
• Saved-workspace library: PR #29
• Sidebar and navigation polish: PR #31
• First-party sign-in and sign-up: PR #32

Changed

• Rebuilt the workbench around searchable, project-organized conversations, including drag-and-drop filing, distinct complete and archive states, a unified title-and-conversation panel, and a focused export entry point.
• Simplified the composer and evidence controls with familiar icon-only actions, cancellable response generation, styled selectors, and less repeated target context throughout the interface.
• Expanded About into a product and reference-data guide, and moved source attribution out of the workspace footer into that dedicated page.
• Replaced the dedicated feedback card and header search form with familiar navigation plus a persistent multifunction command composer.
• Added intent routing for gene questions, bug reports, product feedback, direct gene navigation, report commands, voice input, and attached evidence files.
• Promoted import, report generation, raw export, Word-compatible export, and print-to-PDF to explicit styled actions.
• Replaced the native model dropdown with Structured Liquidity engine controls.
• Made the AlphaFold viewer's existing drag/touch interaction explicit and added styled reset and auto-rotation controls plus responsive resizing.
• Added mobile-specific navigation, command-composer, action, report, and 3D-viewer layouts.
• Introduced a modular analysis runtime that separates the scientific evidence workspace from the interpretation engine and output format.
• Added optional Ollama model discovery and switching for report generation, while retaining a deterministic Observatory evidence report.
• Added working raw-evidence JSON and Markdown report exports.
• Documented how background indexing, typed ingestion workers, portable data workspaces, and an optional OpenWebUI engineering console fit the product without replacing Observatory's interface.
• Renamed the company, platform, application, repository, service identifiers, and primary domain to Observatory and observatory.bio.
• Replaced the landing-page-led prototype with a product-first scientific workspace.
• Applied the Structured Liquidity design language in light mode with rigid containment, liquid glass depth, sharp geometry, flat offset shadows, and a single purple accent.
• Reframed Observatory as a commercial scientific-analysis platform rather than a family prototype.
• Established the long-term product direction around connected visualization, reproducible workflows, scalable analysis, scientific extensibility, and lower-cost access to high-quality tools.
• Added a staged roadmap from the private scientific MVP through workflow, commercial-scale, and scientific-innovation platform phases.
• Clarified that private testing is the initial validation strategy, not the product's eventual audience or scope.

Security

• Hardened private staging responses with no-store, no-referrer, framing, content-type, permissions-policy, and search-index exclusion headers.
• Replaced direct staging-credential equality checks with constant-work comparison and documented the limits of shared Basic authentication.

The scientific core is taking shape

What changed

• Observatory now has versioned identity evidence for TP53, BRCA1, CFTR, EGFR, and DMD, connecting genomic, protein, and evidence views around the same source-backed record.
• Reports and exports now carry provenance from checksummed MANE 1.5 and UniProt 2026_02 snapshots, with live-source drift detection.

Why it matters

• This moves Observatory beyond a visual prototype. Scientific results can now be traced to pinned sources, checked for drift, and reproduced as the workbench evolves.

Private by design

• Staging now uses invite-only Clerk access with first-party sign-in, per-user controls, origin validation, and fail-closed deployment behavior.

What we are observing next

• Richer research workflows, persistent projects, and clearer paths from evidence to a shareable report.

How you will hear from us

• Observatory Field Notes now travel through code-based Resend delivery. Samir receives the exact preview first, and family delivery remains paused until he approves it.