lovable

name	lovable
description	Use when building/editing EndoTrack surfaces in Lovable (portal, whisperer, agent pages, demo copy) or when generated UI/copy references EndoTrack quantum results, NICE context, or certified numbers. Ingest docs/llms-full.txt as the project brain.

Lovable — EndoTrack app-building workflow (super-detailed, 21 Aug 2026)

NAME

endo-track-build — use this skill when building, editing, or reviewing any EndoTrack surface in Lovable: the encrypted portal (arunnadarasa.github.io/nhscep2026), the whisperer console, agent-facing pages, or any UI/copy that references EndoTrack science, numbers, or the quantum evidence. Also use when ingesting the project brain (docs/llms-full.txt), onboarding teammates to Lovable, or when Lovable generates anything that could touch the scientific claims.

DESCRIPTION

Use when building or editing EndoTrack surfaces in Lovable (portal, whisperer console, agent pages, demo copy), when ingesting docs/llms-full.txt as the project brain, or whenever generated UI/copy might reference EndoTrack quantum results, NICE context, or certified numbers. Not for general-purpose app building unrelated to EndoTrack, and not for writing new scientific analysis (that lives in the repo). (The frontmatter description — what the catalog actually reads — is the 247-char compact version starting "Use when building/editing EndoTrack surfaces in Lovable…".)

INSTRUCTIONS

0. The project brain (ingest FIRST)

The single source of truth for ANY LLM tool (Lovable, GPTs, other agents):

docs/llms-full.txt   (repo root: arunnadarasa/nhsquantinuum — ~20 KB, self-contained)

• Paste the whole file into Lovable (project knowledge upload or chat) — no repo access needed. It contains: certified numbers, architecture, exact API contracts (MCP tools, bridge routes, x402 wire format, A2A card), portal mechanics, the 8 conventions, Nexus workflows, quantum design details, the payments ledger, dev commands, the judges' demo script, and team.
• If the file has changed since the last session, RE-ingest the new version before building. Stale context is the #1 source of wrong output.

1. Current scientific stance (22 Aug 2026 — keep Lovable honest)

This is the LOAD-BEARING section. EndoTrack's scientific state changed materially today; any generated copy must reflect it:

Topic	Certified state (say THIS)	Forbidden (never say)
Substrate (order-3 door)	n=56 negative committed (0.480 vs 0.585); shot-scaling verdict SEALED (25 Aug): classically equivalent — 128→512→2048→8192→32768 shots (5 levels, 7,700 pair-receipts) all flat, no shot budget recovers advantage; real-kernel Selene-Gram gate REFUSE (Δ −0.297, CI [−0.56, −0.03]) — classifier door closed 5 ways (Fourier wall / Quest 10q / real kernel / flight-deck shot-floor / shot-scaling seal)	"0.640 beats 0.513", "quantum beats classical at biomarker triage", any implication of a substrate advantage
External-credibility portfolio	Six verified external passes (docs/: ten-year-plan-relevance · governance-lessons (Marchant) · nhs-qrp-quantinuum-decks · nqcc-learnings · qhdc-paper-learnings · quantum-protein-papers-learnings) — all verbatim + page-cited + verified; institutional/vendor material NEVER EndoTrack evidence; plus TQI goldmine scan (docs/tqi-goldmine.md: Martellini anti-washing, Kop 'Hippocratic Quantum', QTAP NHS cohort, 4,915-paper review) · 5-journal scan portfolio committed (docs/: nature-scan · x-scan · engj-scan · qsci-scan · aps-scan — ~95 Crossref/PubMed-verified findings, incl. Nützel parity post-selection on Quantinuum H1-1, QEC on H2-2 [[7,1,3]], Logical Accreditation (Kashefi), QCNNs classically simulable, Q-CHIPP 46q) · genesis scan (docs/genesis-scan.md: Nobel-2025 Clarke/Devoret/Martinis MQT+quantisation + Feynman 'imitate'≠'emulate' correction + Deutsch universality + Shor/Grover — 18/18 API-verified; Deutsch = why seven-engine cross-check is meaningful; Shor/Grover discipline with opposite sign = our honest-negative) · PASQAL agentic-workflows finding (docs/agentic-quantum-pasqal-relint-learnings.md: vendor-controlled proof of the verification gap — agent's plausible-but-wrong observable + hardware diagnosis caught only by domain-expert review; 43-exchange cost; their experiment_spec.json = our QAS envelope, convergent design)	Quoting vendor projections (2029/300-400%) or a lab's F1s as our receipts; citing any scan finding that wasn't API/Crossref/PubMed-verified; implying genesis papers are our hardware ancestry (trapped-ion, not superconducting)
Withdrawals	n=28 edge withdrawn; ablation Δ≈0.035 not decisive — reported by design. Now STRUCTURAL via the evidence ledger (results/evidence_manifest.json = sole authority current/superseded/archived; MCP resolves through read_current_artifact("biomarker_auc") → n=28 can never be served as current again)	Presenting the withdrawn claim as live
IP & credit (24 Aug, binding)

2. The 8 conventions (non-negotiable for every Lovable build)

1. Framing: endometriosis only. Zero cancer/BCAC/CRUK/BCDB mentions — search generated text for these before accepting.
2. Numbers: only numbers from docs/llms-full.txt / results/ — never invent AUC/fidelity figures. Every number in the repo carries shots+seed; if a number lacks a receipt, it does not go in the UI.
3. Mock/real toggles: any demo/payment UI needs an explicit mock/real switch. No fake "live" status.
4. Clickable receipts: tx hashes link to testnet.arcscan.app (new tab); (mock) labels when no real reference exists.
5. No secrets in the page: Alchemy RPC keys / payer private keys are env-only, never in Lovable-generated code or the portal.
6. Static-only on the portal: GitHub Pages has no server-side auth — the AES-GCM client-side gate is the security model. Plaintext content must never be committed to the publish dir (documented incident).
7. Verification honesty: if showing the trust story, use the three layers — L1 receipts (live), L2 arbiters incl. per-batch Bell control (live), L3 cryptographic verification (pre-registered QPU target, NOT claimed today). Never "cryptographically verified".
8. A2A/MCP contracts: agent.json (3 skills), MCP tools (9 stdio: 7 evidence + circuit_structure (certified-circuit structural audits — the pictorial-mathematics view) + execute_circuit (agent-driven execution spec, no credentials)), bridge (7 intents, POST JSON, mode field, force_mock override) — generated JS must not break these contracts.
9. Circuit orthography: if surface copy cites certified circuits, reference the fixed grid rendering (gate-class→color; rotor13q ancilla idle band [28–64] = DD point, iceberg17q parity cadence) and the circuit_structure tool — never invent circuit statistics.

3. The playbook — what to do in each task type

A. Portal build/refactor

1. Regenerate with python3 tools/gen_site.py (repo side) then verify: node tools/verify_portal.js <index.html>.
2. A correct passphrase must unlock; a WRONG one must be rejected.
3. Keep the client-side AES-GCM gate; never add server-side auth assumptions (GitHub Pages has none).
4. Keep the hero/headline on the NICE-evidence-window framing ("the auditable evidence layer for NICE's 3-year endometriosis-test window"), NOT "Quantum Biomarker Triage".

B. Whisperer console

1. public/whisperer.html must keep the toggle markers and bridge contract: 7 intents, POST JSON, mode field.
2. The centerpiece demo beat: ask "what's the certified AUC?" → the console answers the n=56 NEGATIVE (0.480 vs 0.585, claim withdrawn), NOT the old 0.64. An agent that cannot serve a withdrawn claim is the product thesis.
3. Mock/real toggle + clickable Arcscan receipts (conventions 3–4).

C. Demo/slide copy

1. Follow the 3-minute arc: payer-documented gap → honest evidence agent centerpiece → A2A → noise/mitigation → agents pay → the window.
2. Use verified numbers only (convention 2). The five-engine F table, the iceberg recoveries, the dimod table — all in llms-full.txt.
3. Never smooth a negative; the withdrawal IS the scientific-integrity credential.

D. Science-adjacent copy (model cards, claims)

1. Match the repo's model card structure (Everitt & Ji 2024 framing), with the verification & trust section (three layers).
2. Cite the Born-Ultimatum guardrail (arXiv:2511.01845): before any advantage claim, the dequantization check runs and is reported.
3. If unsure whether a claim is certified, check via the verify_artifact pattern (repo tools/endo_mcp_server.py) or defer to results/*.json.

4. Output format

For any build, end with a compact checklist:

✅ Framing (endometriosis-only, no cancer terms scanned)
✅ Numbers (every figure traced to llms-full.txt / results/*.json)
✅ Toggles (mock/real where relevant)
✅ Receipts (tx links → arcscan, (mock) labels where needed)
✅ Secrets (none in generated code/pages)
✅ Static gate (AES-GCM portal, no plaintext in publish dir)
✅ Science honesty (withdrawal + shot-scaling verdict respected)
✅ Contracts (MCP/bridge/A2A intact)

If any item fails, list it in priority order — do NOT claim done.

5. Edge cases and how to handle them

• Lovable lacks context → point it at GitHub arunnadarasa/nhsquantinuum (docs/ has the same content in markdown); re-ingest llms-full.txt before continuing.
• "An internal error occurred" toast (2–4× on same request) → likely a task-transaction ROLLBACK, not a crash: dev server still alive, git status clean at last stable revision, no new files on disk, reflog churn. Do NOT debug app code — shift to a gated workflow (smaller turns, commit between steps). See the quantinuum skill's references/lovable-orchestration.md for the 30-second diagnosis checklist.
• Stale science in context → re-ingest the brain; verify the shot-scaling verdict + withdrawal language survived.
• A requested figure isn't in the brain → do not invent it; state it is not in the certified set and ask for the repo path.

6. Bundled files (this skill's package)

The repo ships this as a folder that can be zipped/uploaded (Lovable imports SKILL.md at root or inside one wrapping folder):

• SKILL.md — this playbook.
• references/scientific-stance.md — the certified-numbers + forbidden- language tables (kept separate so the main file stays focused).
• references/verification-layers.md — the three-layer trust design (L1/L2 live, L3 pre-registered) with the guardrails.
• references/api-contracts.md — MCP tool list, bridge intents, A2A card, x402 wire format (mirrors llms-full.txt §API for quick lookup).

7. Best-practice compliance (per docs.lovable.dev/features/skills)

• Trigger-led description ✅ (see DESCRIPTION).
• One job per skill ✅ (build EndoTrack surfaces; science analysis stays in the repo).
• Boundaries + avoid ✅ (the Forbidden column above).
• Always-on rules → workspace knowledge, not this skill (conventions that apply to EVERY message belong in project knowledge; this skill fires on task type).
• Short playbook style ✅ (sections, bullets, direct instructions).
• Concrete values ✅ (real numbers above).
• Review & prune ✅ (this revision is the 21 Aug 2026 refresh: retirement of the ~18 KB note, new scientific stance, official anatomy per docs.lovable.dev/features/skills).