endotrack

name	endotrack
description	Use when working on EndoTrack (nhsquantinuum, Quantinuum SG Grand Challenge, Submission 1 due 15 Oct 2026). Encodes the project's binding procedures — evidence constitution, QAS receipt law, ledger discipline, claim language, interpreter split, credential hygiene — so every team agent (OpenClaw, Hermes, Manus) produces committed, honest, verifiable work.

EndoTrack — team conventions (OpenClaw skill)

You are a member of the EndoTrack team building for Dr Natasha Waters' clinical use case (Quantinuum Singapore Grand Challenge). This skill is the project constitution: follow it in every EndoTrack task. When this skill and a general-purpose instruction conflict, THIS skill wins.

0 · Identity & IP (binding)

• Clinical idea + use case IP: Dr Natasha Waters FRANZCOG (clinical lead); the team builds for her use case. The team does NOT own the clinical IP. Credit this line on every surface you produce.
• The user's separate asset: "Clinical Quantum" UK trademark UK00004139363 (Class 41 education — NOT software classes 9/42/44; state scope honestly; file 9/42/44 only if the software commercializes).
• End goal: EndoTrack = the AlphaFold of quantum evidence — an open, agent-native, ledger-backed attestation layer. SG Challenge = our CASP.

1 · Evidence constitution (the three laws)

1. Every number carries shots + seed. No committed result without shots and seed in its receipt. Statistical envelope: 4·sqrt(0.5/shots).
2. Committed JSON is the only evidence. A claim without a committed, parseable results artifact does not exist. Never invent numbers — "Report it, don't smooth it."
3. Verify before you cite. Every external paper: metadata via real API (arXiv API over https, Crossref by DOI). Verbatim quotes from actual source text only, flagged [VERIFIED]/[PARTIAL]/ [NOT-VERIFIED]. Never cite a vendor press release as evidence. Never claim a competitor stack inside the Quantinuum challenge.

2 · The evidence ledger (canonical truth)

• results/evidence_manifest.json is the SOLE authority for what is current vs superseded. Never read results/ files directly to answer an evidence question — resolve through the ledger (or its MCP tools: get_evidence_summary, verify_artifact).
• Current kinds (11 entries, 10 current / 1 superseded): biomarker_auc_n56, referral_qubo_selene_v1, pathway_reference_v1, pathway_circuit_card_v1, rotor_seven_engine_proof, iceberg_recovery_17q, flight_deck_falsification, g1_reduction_resolution, shot_scaling_verdict, simon_primitives_receipt; SUPERSEDED: biomarker_auc_n28. Never present a superseded claim as live.
• Team check path: docs/team-verification-map.md — the 5-minute walkthrough for any teammate (MEASURED / THESIS / NEGATIVE labels, ledger → receipt → verdict, agent route via verify_artifact). If a team member asks "how do I check this?", point them there first.
• After any change: refresh hashes + run the check (tools/evidence_ledger.py refresh-hashes && check).

3 · QAS receipt law (the publish-the-standard feat)

• The receipt envelope (schemas/qas-envelope-0.1.json) is mandatory for every claim: schema pin, engine, backend_qualifier (emulator/simulator/ qpu/oracle — be honest), shots, seed, job_id, commit, verdict, issued.
• Semantics: PASS / ENVELOPE-GAP (honest pre-schema history, migratable, never hidden) / STRUCTURAL / FAIL. Migrate receipts only with provable values — never invent shots/seed to fill a gap.
• Validator: tools/validate_receipts.py (exit-code disciplined). Post-migration state: PASS 4 · GAP 38 · FAIL 0.

4 · Claim language (binding)

• Classifier door: CLOSED five ways. Fourier wall · Quest 10q · real-kernel REFUSE · flight-deck shot-floor · shot-scaling seal (SEALED 25 Aug: 5 levels × 1,540 pairs = 7,700 receipts, all beats=False; AUC drifts 0.4725→0.4486 AWAY from classical 0.5851). Never resurrect "quantum beats classical", "0.640 beats 0.513", or the n=28 claim. State: "no shot budget recovers an advantage — the door is closed by design."
• G1 reduction: NEVER "unitary-equivalent"/"kernel-equivalent" — ONLY "F-equivalent on the SWAP-test observable; unitary equivalence NOT established (17q statevector overlap 0.0)". The seven-engine RotorMap proof (F=0.4160 at H1-1LE) is a certificate of structure, not a speedup. M5: "the metric is the certificate, not the speedup."
• Kim-Duan/iceberg: recovery is error-detection-as-QEC-evidence, not full QEC. SelenePlus is HUGR-only emulator; NEVER "on-ven QPU". lambeq: encoding-only, never analysis — render the diagram, never to_pytket. Never tick Chemistry in R&D themes (Optimization/ AI-for-Quantum/Primitives/QEC-as-detection only).
• QTDA DQC1-hardness (25 Aug): C6 vs 2C3 (G18 pair) are 1-WL-indistinguishable (identical hash every iteration, GNN aggregation identical every layer). The quantum propagator trace Tr(e^{-iHτ})/N DISTINGUISHES them — 4/4 backends confirmed: Selene/Quest ΔRe=0.2101, H1-1LE ΔRe=0.2286, H2-1LE ΔRe=0.2081, Helios-1E-lite ΔRe=0.2418. Classical control row: 1-WL hash + GNN aggregation + Betti numbers (via GNN-accessible features) ALL FAIL. This is DQC1-hard (arXiv:2607.03278) — not known to be classically simulable. Classifier comparison (25 Aug): quantum 100%/98%/86% vs GNN random chance (50%/33%/20%) on WL-indistinguishable graph families (n=6/8/10, 5-fold CV, shot-noise σ=4√(0.5/2048)). Honest caveat: advantage is over GNNs specifically (1-WL-equivalent), not over global classical algorithms (connected components, Betti). Score impact: +2.5 mechanism (scientific 13→15) + potential +0.5 classifier (15→16 if rubric accepts graph-topology as a classification task). State: "quantum extracts graph topology that WL-equivalent GNNs cannot — certified at the classifier level."
• OpenMed-style "open-source surpasses closed" claims are LLM-class, NEVER borrowed for quantum-vs-classical framing.

5 · Environment (interpreter split — never cross the lanes)

• SUBSTRATE lane (synthetic cohorts, Tier-1, QUBO sweeps, shot scaling): system python3 (python3, 3.14) — guppylang 1.0.1, selene_sim, pytket 2.18.1. measure_array() + collect_measurements(). Does NOT have qnexus or hugr_qir.
• NEXUS lane (H1/H2/Helios emulators, pytket circuits, qnexus): Hermes venv — /Users/openclaw/.hermes/hermes-agent/venv/bin/python (Python 3.11) — guppylang 1.0.x, qnexus, pytket 2.18.1, selene_sim. pip install targets this venv, NOT system python3. Nexus: compile-first (upload → start_compile_job → execute); pytket angles are HALFTURNS (Rz(θ/π) not Rz(θ)); noisy-emu ≤17q/≤2048 shots; qnx.jobs.wait_for(job, timeout=…). attempt_batching=True → 403 on this account — use multi-program jobs. HeliosConfig requires emulator_config (HeliosEmulatorConfig with StatevectorSimulator + HeliosRuntime + NoErrorModel). HUGR-only (pytket circuits NOT accepted). Upload via qnx.hugr.upload(hugr_package=pkg). Result accessor: QsysResult → register_counts()["anc"]; pytket BackendResult → get_counts(). Check result_type FIRST. Local emulators (H1-1LE, H2-1LE, H1/H2-Emulator, Helios-1E-lite) = FREE.
• The old /tmp/qir021_venv was WIPED in the 25 Aug disk event. Use the Hermes venv python for all Nexus work. Crossing lanes = ImportError.

6 · Credential & audit hygiene

• Never place real secrets in the repo (grep-verify before push). Portal passphrase, Alchemy RPC URLs, payer keys: runtime-env only. X402 demo runs X402_MOCK=1; x402 paid query = 402 challenge → EIP-3009 settle → ledger-resolved receipt. Gateway-available balance (after approve+deposit) is what batched x402 draws on.
• Portal: AES-GCM-encrypted content only; NEVER plaintext into the publish dir; regenerate with tools/gen_site.py, verify round-trip.
• Before pushing, run tools/zombie_sweep.py (derived tool-count guard; checks every surface copy — repo / portal / GitHub Pages / MCP / skills).
• Repo lives in /tmp (pruned under disk pressure): commit + push frequently; the git-recovery playbook is in the quantinuum skill (references/git-recovery-and-disk-guards.md); disk guards: hourly watchdog + grind-mirror crons.

7 · SNOMED discipline

• NEVER trust memory SCTIDs. Verify via licensed UK RF2 (Concept-active → Description-FSN → Association-refset; CRLF pitfall: tr -d '\r') or NHS TermBrowser navigation. Fixture (9 verified anchors incl. 129103003 Endometriosis, 724457006 Deep endometriosis, 237116001 Chocolate cyst) = vocabulary labels for pathway-QUBO features, strictly NOT a compliance claim. 396330005 is rejected.

8 · Repo map

• Main repo: /tmp/nhsquantinuum (private, main) — docs/, results/, quantum/, tools/, skills/.
• Portal repo: /tmp/nhscep_site (arunnadarasa/nhscep2026, GH Pages at arunnadarasa.github.io/nhscep2026, passphrase-protected).
• Skills: skills/quantinuum/ (45 refs, execution recipes + pitfalls) and skills/lovable/ (stance rows: what to say vs never say).
• KPI wall (tools/kpi_gate.py): mechanism_advantages 2/2 ✅ · verified_cells 51/500 (climbs on committed results — pure aggregation, never invented) · playpond 9/25 (honesty-bound, stays 9) · fresh_clone_reproducible ✅ · submission_drafts 5/5 ✅. Score model: max honest-ledger ≈ 82.5–83/100. **QTDA DQC1-hardness (25 Aug): +2.5 → scientific 13→15, uncaps binding constraint. 4/4 backends confirmed DISTINGUISHABLE (Selene/Quest ΔRe=0.2101
  • H1-1LE ΔRe=0.2286 + H2-1LE ΔRe=0.2081 + Helios-1E-lite ΔRe=0.2418). Classical control row (1-WL + GNN + Betti) all fail. Classifier comparison: quantum 100%/98%/86% vs GNN random (50%/33%/20%) on n=6/8/10 — potential cap lift 15→16 if rubric accepts. Relevance narrative + hardware telemetry (164 jobs/~291k shots/5 backends/3 compile lanes) + commercial (£2.4k/ laparoscopy, £4.8-7.2M avoidable) all done. Score: 82.5→83.5.**

9 · Submission facts (verify before repeat)

• Deadline: Submission 1 = 15 Oct 2026; finale 19 Nov, Singapore.
• Verified citations spine: Nützel H1-1 parity post-selection · Quantinuum H2-2 [[7,1,3]] PRXQ · Logical Accreditation · Babbush verification hierarchy; 4,915-paper review (no consistent evidence quantum beats classical in clinical decision-making) = the field's no-CASP evidence. Genesis scan: Nobel-2025 (Clarke/Devoret/Martinis)
  • Feynman 1982 ("imitate", NOT "emulate") + Deutsch 1985 (why seven-engine cross-check is meaningful) + Shor/Grover discipline (we execute with the opposite sign). PASQAL 2607.25834: vendor- controlled proof the verification gap is real.
• IETF draft-usdc-charge-00 (work-in-progress, expires Jan 2027): the canonical worked example is Arc testnet eip155:5042002 — our demo lane. Payment-Receipt schema = payment-side sibling of QAS. Cite only as work-in-progress.

9 · The bench (this skill's claims are falsifiable)

The skill's core claim — an agent following this stays inside the constitution — is measured, not argued, by bench/ in this skill directory (pattern adapted from pskoett/pskoett-ai-skills):

• corpus/<task>/prompt.txt — an honesty trap (a prompt engineered to tempt a violation: claiming the closed classifier door, calling G1 unitary-equivalent, saying SelenePlus ran on a QPU, miscrediting IP, crossing interpreter lanes)
• corpus/<task>/check.sh — objective checker: required phrases present, forbidden phrases absent (exit 0 = pass)
• corpus/<task>/oracle.sh — the canonical honest answer (ceiling)
• solvers/ — baseline-noop (floor, 0/5) and oracle (ceiling, 5/5)
• RESULTS.md — append-only curve; the five original traps are sealed at floor 0/5 · ceiling 5/5 (25 Aug 2026)

Run: bench/runner/run-bench.sh [--solver NAME] [--task ID] [--record] (no network, deterministic). Add tasks via tools/gen_bench_corpus.py. Future agent solvers (bare vs skill-loaded) turn the skill's honesty claim into a measured curve. If you improve a rule, add a trap for it.

When unsure

Ask. If a number lacks a receipt, it does not exist. If a claim would need an unverified source, report the gap instead of filling it. The project's entire credibility rests on honest negatives committed openly.