feat(eln): scientific spine (Strain→Experiment→Hypothesis→Result) + perception→HuggingFace flywheel

[pskeshu]

ELN scientific spine — Strain → Experiment → Hypothesis → Result

Adds a co-authored electronic-lab-notebook scientific spine to gently as optional,
file-based overlays, plus the perception-annotation → HuggingFace flywheel. Turns
data-collection into claim-testing and gives the human a write-path that isn't chat.

Origin: docs/product-ideation/ENTITY-EVOLUTION.md (Fable's entity-evolution assessment)

• Hari Shroff's Plan for Agentic Microscopy Paper. Design brainstormed → spec → plan.

Built + tested (backend, 28 tests)

• Spine stores (FileContextStore, agent/{strains,experiments,hypotheses,results}/):
  Strain (create/get/list/resolve with bare-string back-compat), Experiment
  (arms × controls, link_session_to_arm), Hypothesis (proposed/supported/refuted/
  inconclusive), Result (accuracy/measurement/derived + experiment_ref filter).
• gently/eln/ package: accuracy (per-timepoint predicted-vs-ground-truth →
  accuracy Result), hf_connector (approval-gated push; HF_TOKEN from env; default
  repo pskeshu/gently-perception-benchmark; lazy datasets; never auto-pushes),
  strain_resolution (non-mutating display shim), seed_loader.
• Ground-truth authoring route POST /api/embryos/{id}/ground_truth (require_control,
  range-based → set_ground_truth) — replaces the localStorage-only Agree/Disagree
  dead-end.

The flywheel

gently-perception predicts → in-ELN ground-truth (assisted batch-confirm: agent
pre-screens ranges, human confirms uncertain frames) → accuracy Result → approval-gated
export to HuggingFace → retrain. Bidirectional: the model's calls become the human's
annotation queue; the corrections become the next training set.

Real seed programs (dogfood, not stubs)

seed/programs/: stage-classification (nuclear+BF fusion, hatching), temperature-
hatching (Richard's K-pump thermo-paralytic → hatching-time distributions), mutant-
screen (WT vs mutant ID), dopaminergic-outgrowth ("catching outgrowth", BZ555
dat-1p::GFP, expression-triggered high-cadence via dopaminergic_signal.py). A test
loads all four into a fresh store.

Docs

• Spec: docs/superpowers/specs/2026-07-02-eln-scientific-spine-design.md
• Plan: docs/superpowers/plans/2026-07-02-eln-scientific-spine.md
• Usage eval: docs/product-ideation/ELN-USAGE.md (biologist journey over time, iterative
  stories, story-flow specs, design implications)

Remaining (in the plan, flagged for in-browser verification)

Vitals stage-strip range-authoring UI + assisted batch-confirm, experiment inspector +
export button, agent tools, and the ELN US- story flows (after #76 lands the crawler
harness on the base). Deliberately deferred because they need the running app + a Chrome
MCP UI/UX audit before shipping.

Non-goals (YAGNI)

Setpoint→Actuator, Embryo→Specimen rename, multi-tenant, re-inventing the organism
plugin, gently/dataset migration.

🤖 Generated with Claude Code — Fable 5

[pskeshu]

Follow-up commits:

Safety fix — inverted confirmation guards (39d9ff7). The UI confirmed reversible map edits (Remove/Move embryo, Stop) but fired the two specimen-harming actions unconfirmed. Now toggleLaser (OFF→ON), setTemperature, and setLightsheetTemperature ask an explicit confirm naming the effect on live embryos; turning the laser OFF stays friction-free. require_control unchanged.

Ground-truth authoring UI (b00d41f, 61fc478) — the annotation flywheel's visible write path. Replaces the localStorage-only Agree/Disagree with a canonical-stage picker pre-filled with the model's predicted stage + a Confirm that POSTs to /api/embryos/{id}/ground_truth, shows a persisted GT: {stage} badge, and a visible 'Save failed' hint on rejection. The GT route is now store-flexible (works without an agent). Verified in-browser via the real render path (screenshots attached in the session).

Still remaining (in the plan): assisted batch-confirm, live accuracy readout, experiment inspector + export button, agent tools, and the ELN US- story flows (after #76 lands the crawler on the base).

[pskeshu]

Now visible in the UI (the earlier commits were backend/data; these make it show):

• scripts/load_seed_programs.py + launch_gently --seed — load the 4 seed research programs into the store. A fresh --seed boot opens with real notebook content.
• Plans tab now renders the ELN scientific spine per campaign: Strains (chips w/ genotype), Experiments (cards; arms with strain chip + condition; controls distinguished), Hypotheses (status-badged) — via new GET /api/campaigns/{id}/eln.
• Verified in-browser (fresh origin, native integrated render): the Temperature program shows RK-KP1 (Richard's twk-18 K-pump strain) + WT-HIS, 3 experiments (paralysis validation, 2×2 protocol×imaging, thermo-cycling optimization) with arms/conditions/controls, and 4 proposed hypotheses.

Run it: python scripts/load_seed_programs.py then restart, or launch_gently --seed.

[pskeshu]

Flywheel push landed (per-embryo timelapse → HuggingFace, the active project's terminal step):

• gently/eln/export_service.py — collect_embryo_annotations pairs each prediction with the ground-truth stage covering its timepoint (range-based) into HF rows; annotation_summary gives {n_predictions, n_annotated}.
• POST /api/embryos/{id}/export (require_control) — approval-gated (confirm + HF_TOKEN) push via export_service + hf_connector; GET /api/embryos/{id}/annotation-summary. Shared _gt_store() resolver (works agent or benchmark mode).
• UI: each Vitals strip shows 'N/M annotated' + a Push to HF button (disabled when nothing annotated; inline outcome: Pushed N / Set HF_TOKEN / Log in / Failed). Verified in-browser.
• 38 ELN tests green.

Per the plan-mode↔ELN fold study (committed), the flywheel is model-agnostic so it ships independent of the fold. Next: ELN agent tools, then the fold (Experiment→phase-role).