Which re-tests should a lab run first?
A panel of 8 re-tests chosen by exact optimisation of the model's predictions: the clearest robust profile gaps, spread across compound classes, uses and preparations, with little overlap and at most 1 per plant. Hypotheses worth re-testing, never claims.
Panel optimiser
Not independently validatedProduced by a rule or model; no independent check. No claim is made either way.- Compound classes
- 5
- Indications
- 17
- Preparation types
- 6
Not offered for planning because toxicity is reported: Allium cepa, Chelidonium majus, Glycyrrhiza glabra, Hyoscyamus niger (8 candidates: fresh juice vs modern methanol extract; steeped in vinegar vs decoction in water; fresh juice vs modern methanol extract; infused oil vs modern methanol extract; powder taken whole vs modern methanol extract; decoction in water vs modern methanol extract; powder taken whole vs modern methanol extract; decoction in wine vs modern methanol extract).
Not eligible under rule v2: Raphanus sativus: decoction in wine vs decoction in water (0 compounds likely to differ (needs at least 1)).
How the panel was chosenSuggested panel (8)
- 1
Ajuga chamaepitys: decoction in water vs modern methanol extract
Historical: decoction in waterModern: modern methanol extract (default comparator: no retrieved study used a modelled modern preparation)- Compounds likely to differ
- 33Heuristic · Compounds likely to differA compound counts when the 5th–95th range of its predicted difference between the two preparations excludes zero and the median difference reaches an assumed threshold (differ_delta in config/m8_qubo_v2.toml); not fitted to data.A rule of thumb applied by the model, not a measurement.
- Compound classes
- phenolic, terpenoid-like
- For
- Jaundice
- Status
- Not independently validatedProduced by a rule or model; no independent check. No claim is made either way.A model-based hypothesis worth re-testing; not independently checked.No toxicity reported
Open the historical preparation in the Labhypothesis worth re-testing, not a claim that the preparation works or is safe (+4 more in the trail) - 2
Artemisia absinthium: hot water infusion (tea) vs modern methanol extract
Historical: hot water infusion (tea)Modern: modern methanol extract (default comparator: no retrieved study used a modelled modern preparation)- Compounds likely to differ
- 110Heuristic · Compounds likely to differA compound counts when the 5th–95th range of its predicted difference between the two preparations excludes zero and the median difference reaches an assumed threshold (differ_delta in config/m8_qubo_v2.toml); not fitted to data.A rule of thumb applied by the model, not a measurement.
- Compound classes
- alkaloid-like, glycoside-like, other, phenolic, terpenoid-like
- For
- Flatulence
- Status
- Not independently validatedProduced by a rule or model; no independent check. No claim is made either way.A model-based hypothesis worth re-testing; not independently checked.Toxicity concern
Open the historical preparation in the Labhypothesis worth re-testing, not a claim that the preparation works or is safe (+4 more in the trail) - 3
Betonica officinalis: decoction in wine vs modern methanol extract
Historical: decoction in wineModern: modern methanol extract (default comparator: no retrieved study used a modelled modern preparation)- Compounds likely to differ
- 14Heuristic · Compounds likely to differA compound counts when the 5th–95th range of its predicted difference between the two preparations excludes zero and the median difference reaches an assumed threshold (differ_delta in config/m8_qubo_v2.toml); not fitted to data.A rule of thumb applied by the model, not a measurement.
- Compound classes
- glycoside-like, phenolic
- For
- Colic
- Status
- Not independently validatedProduced by a rule or model; no independent check. No claim is made either way.A model-based hypothesis worth re-testing; not independently checked.Toxicity unknown
Open the historical preparation in the Labhypothesis worth re-testing, not a claim that the preparation works or is safe (+3 more in the trail) - 4
Raphanus sativus: cold water infusion vs modern methanol extract
Historical: cold water infusionModern: modern methanol extract- Compounds likely to differ
- 94Heuristic · Compounds likely to differA compound counts when the 5th–95th range of its predicted difference between the two preparations excludes zero and the median difference reaches an assumed threshold (differ_delta in config/m8_qubo_v2.toml); not fitted to data.A rule of thumb applied by the model, not a measurement.
- Compound classes
- alkaloid-like, glycoside-like, other, phenolic, terpenoid-like
- For
- Poisoning
- Status
- Not independently validatedProduced by a rule or model; no independent check. No claim is made either way.A model-based hypothesis worth re-testing; not independently checked.Toxicity concern
Open the historical preparation in the Labhypothesis worth re-testing, not a claim that the preparation works or is safe (+3 more in the trail) - 5
Agrimonia eupatoria: fresh juice vs modern methanol extract
Historical: fresh juiceModern: modern methanol extract (default comparator: no retrieved study used a modelled modern preparation)- Compounds likely to differ
- 7Heuristic · Compounds likely to differA compound counts when the 5th–95th range of its predicted difference between the two preparations excludes zero and the median difference reaches an assumed threshold (differ_delta in config/m8_qubo_v2.toml); not fitted to data.A rule of thumb applied by the model, not a measurement.
- Compound classes
- glycoside-like, phenolic
- For
- Ear Diseases
- Status
- Not independently validatedProduced by a rule or model; no independent check. No claim is made either way.A model-based hypothesis worth re-testing; not independently checked.No toxicity reported
hypothesis worth re-testing, not a claim that the preparation works or is safe (+3 more in the trail) - 6
Filipendula ulmaria: decoction in wine vs modern methanol extract
Historical: decoction in wineModern: modern methanol extract (default comparator: no retrieved study used a modelled modern preparation)- Compounds likely to differ
- 8Heuristic · Compounds likely to differA compound counts when the 5th–95th range of its predicted difference between the two preparations excludes zero and the median difference reaches an assumed threshold (differ_delta in config/m8_qubo_v2.toml); not fitted to data.A rule of thumb applied by the model, not a measurement.
- Compound classes
- glycoside-like, other, phenolic
- For
- Diarrhea
- Status
- Not independently validatedProduced by a rule or model; no independent check. No claim is made either way.A model-based hypothesis worth re-testing; not independently checked.Toxicity concern
Open the historical preparation in the Labhypothesis worth re-testing, not a claim that the preparation works or is safe (+3 more in the trail) - 7
Humulus lupulus: powder taken whole vs modern methanol extract
Historical: powder taken wholeModern: modern methanol extract (default comparator: no retrieved study used a modelled modern preparation)- Compounds likely to differ
- 118Heuristic · Compounds likely to differA compound counts when the 5th–95th range of its predicted difference between the two preparations excludes zero and the median difference reaches an assumed threshold (differ_delta in config/m8_qubo_v2.toml); not fitted to data.A rule of thumb applied by the model, not a measurement.
- Compound classes
- alkaloid-like, glycoside-like, other, phenolic, terpenoid-like
- For
- Diuresis
- Status
- Not independently validatedProduced by a rule or model; no independent check. No claim is made either way.A model-based hypothesis worth re-testing; not independently checked.No toxicity reported
hypothesis worth re-testing, not a claim that the preparation works or is safe (+3 more in the trail) - 8
Galium aparine: powder taken whole vs modern methanol extract
Historical: powder taken wholeModern: modern methanol extract (default comparator: no retrieved study used a modelled modern preparation)- Compounds likely to differ
- 5Heuristic · Compounds likely to differA compound counts when the 5th–95th range of its predicted difference between the two preparations excludes zero and the median difference reaches an assumed threshold (differ_delta in config/m8_qubo_v2.toml); not fitted to data.A rule of thumb applied by the model, not a measurement.
- Compound classes
- glycoside-like
- For
- Skin Ulcer
- Status
- Not independently validatedProduced by a rule or model; no independent check. No claim is made either way.A model-based hypothesis worth re-testing; not independently checked.No toxicity reported
hypothesis worth re-testing, not a claim that the preparation works or is safe (+3 more in the trail)
View as table
| # | Plant | Historical | Modern | Robust gap (5th–95th) | Compounds likely to differ | For |
|---|---|---|---|---|---|---|
| 1 | Ajuga chamaepitys | decoction in water | modern methanol extract | 0.234–0.545 | 33 | Jaundice |
| 2 | Artemisia absinthium | hot water infusion (tea) | modern methanol extract | 0.210–0.475 | 110 | Flatulence |
| 3 | Betonica officinalis | decoction in wine | modern methanol extract | 0.161–0.382 | 14 | Colic |
| 4 | Raphanus sativus | cold water infusion | modern methanol extract | 0.136–0.338 | 94 | Poisoning |
| 5 | Agrimonia eupatoria | fresh juice | modern methanol extract | 0.060–0.173 | 7 | Ear Diseases |
| 6 | Filipendula ulmaria | decoction in wine | modern methanol extract | 0.047–0.178 | 8 | Diarrhea |
| 7 | Humulus lupulus | powder taken whole | modern methanol extract | 0.020–0.050 | 118 | Diuresis |
| 8 | Galium aparine | powder taken whole | modern methanol extract | 0.018–0.095 | 5 | Skin Ulcer |
The original result (rule v1)
The first panel, before the pre-registered v2 rule (positive robust gap, at least one compound likely to differ, at most one entry per plant). It stays visible as the original result.
⚑ Slot 8 (Raphanus sativus, decoction in wine vs decoction in water) was selected on coverage, not value: robust gap 0.003 and no compound likely to differ.
⚑ Several entries share a species: Ajuga chamaepitys, Betonica officinalis, Galium aparine, Raphanus sativus.
Original panel, rule v1 (k = 8)
- 1
- Plant
- Ajuga chamaepitys
- Historical vs modern
- decoction in water vs modern methanol extract
- Robust gap
- 0.234
- Likely to differ
- 33
- 2
- Plant
- Ajuga chamaepitys
- Historical vs modern
- decoction in wine vs modern methanol extract
- Robust gap
- 0.222
- Likely to differ
- 31
- 3
- Plant
- Betonica officinalis
- Historical vs modern
- decoction in wine vs modern methanol extract
- Robust gap
- 0.161
- Likely to differ
- 14
- 4
- Plant
- Galium aparine
- Historical vs modern
- fresh juice vs modern methanol extract
- Robust gap
- 0.092
- Likely to differ
- 17
- 5
- Plant
- Raphanus sativus
- Historical vs modern
- fresh juice vs modern methanol extract
- Robust gap
- 0.090
- Likely to differ
- 133
- 6
- Plant
- Betonica officinalis
- Historical vs modern
- powder taken whole vs modern methanol extract
- Robust gap
- 0.058
- Likely to differ
- 11
- 7
- Plant
- Galium aparine
- Historical vs modern
- powder taken whole vs modern methanol extract
- Robust gap
- 0.018
- Likely to differ
- 5
- 8
- Plant
- Raphanus sativus
- Historical vs modern
- decoction in wine vs decoction in water
- Robust gap
- 0.003
- Likely to differ
- 0 ⚑
How stable is this panel?
Each weight multiplied by 0.5 and by 2, the others fixed. The smallest overlap with the panel above is 0.78 (Jaccard; 1 = the same panel).
Weight sensitivity
- Robust value (α)
- Factor
- ×0.5
- Overlap
- 0.78
- Changes
- +1 / −1
- Robust value (α)
- Factor
- ×2
- Overlap
- 1.00
- Changes
- none
- Compound-class coverage (β)
- Factor
- ×0.5
- Overlap
- 1.00
- Changes
- none
- Compound-class coverage (β)
- Factor
- ×2
- Overlap
- 0.78
- Changes
- +1 / −1
- Indication coverage (β)
- Factor
- ×0.5
- Overlap
- 0.78
- Changes
- +1 / −1
- Indication coverage (β)
- Factor
- ×2
- Overlap
- 0.78
- Changes
- +1 / −1
- Preparation coverage (β)
- Factor
- ×0.5
- Overlap
- 1.00
- Changes
- none
- Preparation coverage (β)
- Factor
- ×2
- Overlap
- 1.00
- Changes
- none
- Redundancy (γ)
- Factor
- ×0.5
- Overlap
- 1.00
- Changes
- none
- Redundancy (γ)
- Factor
- ×2
- Overlap
- 1.00
- Changes
- none
Sensitivity check: allowing two entries per plant gives k = 4: 3 plants; k = 6: 4 plants; k = 8: 5 plants (the displayed panel allows one).
How the panel was chosen
The panel maximises robust value (each candidate’s 5th-percentile profile gap) plus coverage of compound classes, indications and preparation types, minus overlap between the compounds each re-test would measure, for exactly 8 re-tests. It is solved exactly with integer programming.
Weights, fixed before any comparison: Robust value (α) 1, Compound-class coverage (β) 0.2, Indication coverage (β) 0.2, Preparation coverage (β) 0.2, Redundancy (γ) 0.5. Pool after rule v2: 15 candidates.
Formulation: docs/QUBO.md; rule: docs/QUBO_v2_RULE.md.
Quantum benchmark
Benchmark against exact optimumA computed comparison of methods against the exact optimum of the same problem. It measures the methods, not the panel; not a validation.The same problem, solved by exact optimisation, two classical heuristics and QAOA (simulated on a laptop) on seeded instances of increasing size. This measures the methods, not the panel: the panel above always comes from the exact solver.
Which method wins at n = 8: exact ILP (0.076 s mean) is the fastest to return the optimum on every instance; also optimal every time: simulated annealing; only the ILP proves optimality. QAOA (simulated): best mean P(optimum) 0.105 (xy mixer, p = 2) vs 0.079 for a uniform random feasible panel, i.e. 1.34× uniform
Approximation ratio (1 = optimum)
- Exact (ILP)1.000
- Greedy0.959
- Simulated annealing1.000
- QAOA, penalty mixer, p = 10.535
- QAOA, penalty mixer, p = 20.515
- QAOA, penalty mixer, p = 30.539
- QAOA, XY mixer, p = 10.537
- QAOA, XY mixer, p = 20.558
- QAOA, XY mixer, p = 30.636
P(sample the optimum)
- Greedy0.667
- Simulated annealing1.000
- QAOA, penalty mixer, p = 10.010
- QAOA, penalty mixer, p = 20.012
- QAOA, penalty mixer, p = 30.025
- QAOA, XY mixer, p = 10.067
- QAOA, XY mixer, p = 20.105
- QAOA, XY mixer, p = 30.050
For greedy and simulated annealing, P(optimum) is the share of instances (greedy) or runs (annealing) that returned the optimum; for QAOA it is the probability of sampling it. Approximation ratio for QAOA is the expected value over feasible samples.
View as table
| Method | Instances | Approximation ratio | P(optimum) | Feasibility | Seconds | CX all-to-all | CX heavy-hex | Depth |
|---|---|---|---|---|---|---|---|---|
| Exact (ILP) | 3 | 1.000 | — | — | 0.0760 | — | — | — |
| Greedy | 3 | 0.959 | 0.667 | — | 0.000232 | — | — | — |
| Simulated annealing | 3 | 1.000 | 1.000 | — | 0.168 | — | — | — |
| Random feasible panel | 3 | 0.518 | 0.079 | — | — | — | — | — |
| QAOA, penalty mixer, p = 1 | 3 | 0.535 | 0.010 | 0.332 | 0.543 | 56 | 117 | 48 |
| QAOA, penalty mixer, p = 2 | 3 | 0.515 | 0.012 | 0.539 | 3.20 | 112 | 232 | 78 |
| QAOA, penalty mixer, p = 3 | 3 | 0.539 | 0.025 | 0.536 | 4.74 | 168 | 372 | 108 |
| QAOA, XY mixer, p = 1 | 3 | 0.537 | 0.067 | 0.808 | 1.08 | 264 | 473 | 510 |
| QAOA, XY mixer, p = 2 | 3 | 0.558 | 0.105 | 0.856 | 6.63 | 350 | 656 | 571 |
| QAOA, XY mixer, p = 3 | 3 | 0.636 | 0.050 | 0.729 | 9.67 | 436 | 836 | 632 |
Under noise
parametric depolarising noise (Qiskit Aer) on a heavy-hex coupling map; not a device calibration. Hardware: not used: no IBM Quantum account or token on this machine, and none was created.
QAOA under parametric noise
- penalty, p = 3
- Two-qubit error
- 0.0%
- P(optimum)
- 0.047
- Feasibility
- 0.192
- penalty, p = 3
- Two-qubit error
- 0.1%
- P(optimum)
- 0.033
- Feasibility
- 0.133
- penalty, p = 3
- Two-qubit error
- 0.5%
- P(optimum)
- 0.016
- Feasibility
- 0.077
- penalty, p = 3
- Two-qubit error
- 1.0%
- P(optimum)
- 0.009
- Feasibility
- 0.035
- penalty, p = 3
- Two-qubit error
- 2.0%
- P(optimum)
- 0.007
- Feasibility
- 0.026
- xy, p = 2
- Two-qubit error
- 0.0%
- P(optimum)
- 0.187
- Feasibility
- 0.593
- xy, p = 2
- Two-qubit error
- 0.1%
- P(optimum)
- 0.099
- Feasibility
- 0.311
- xy, p = 2
- Two-qubit error
- 0.5%
- P(optimum)
- 0.020
- Feasibility
- 0.064
- xy, p = 2
- Two-qubit error
- 1.0%
- P(optimum)
- 0.006
- Feasibility
- 0.028
- xy, p = 2
- Two-qubit error
- 2.0%
- P(optimum)
- 0.003
- Feasibility
- 0.019
Details: docs/QUANTUM_RESULTS.md and docs/QUBO.md.