Skip to main content
Introducing The Quantum Infrastructure Cloud Learn more

Decision Reliability Index · in QCOS

Know whether an optimal plan survives tomorrow’s data.

Your optimiser returns one answer. DRI re-solves the same decision 28 times under controlled changes to its data, parameters and constraints, then scores from 0 to 100 how well the answer holds, with a report your approvers and auditors can read.

  • Method in code, covered by tests
  • Switched off in production
  • Not on sale

Illustration with invented values: a day-ahead plan re-solved 28 times under controlled changes. Three plans with the same objective value score 85, 55 and 25 out of 100.

Illustration with invented values. Equal objective value, unequal reliability. A score of 85 is not an 85 % probability.
controlled re-solves in a default run
28
indices, equally weighted into one score
4 × 25 %
score, read in three bands: proceed, caution, review
0–100
sections in the proof report, from identification to signature
11

Put DRI to work

For teams that act on an optimiser’s answer and have to defend it.

Energy, finance, logistics and manufacturing run optimisers every day. DRI is for the moment between the answer and the commitment: before the plan is dispatched, the portfolio rebalanced, the network built or the line scheduled.

Energy and utilities

Day-ahead unit commitment and dispatch

Does the committed set of units survive a 5 % demand error, or does a different set win as soon as the forecast moves?

Typical teams: Generation scheduling and trading desks at utilities and independent power producers; system operators.

DecisionDay-ahead unit commitment and dispatch
  • noise_injection
  • subsample
  • constraint_relaxation

Does the commitment survive a 5 % demand error?

Probe set · illustrative

  • noise_injectiondemand forecast, 1–20 %
  • subsamplerenewable output, 90–50 %
  • constraint_relaxationramp limits, 1 · 5 · 10 %

Default re-solves

28

10 · 5 · 5 · 3 · 5

Day-ahead unit commitment and dispatch. Probes that would test it: noise_injection, demand forecast, 1–20 %; subsample, renewable output, 90–50 %; constraint_relaxation, ramp limits, 1 · 5 · 10 %. Illustrative.

Who reads the rating

One run, four readers.

Builds the model

Optimisation and OR teams

See which inputs the answer depends on before handing it over, and which constraint is doing the deciding.

Checks the model

Model risk and validation

A repeatable test of a model’s output from a fixed probe configuration, so two runs of the same decision test the same things.

Signs the decision

Approvers and risk committees

One number from 0 to 100 and a band, with the drivers of any weakness named and ranked.

Reviews afterwards

Internal and external audit

An eleven-section report that states the hypothesis, the probes that were run and the ones that were not.

What one solve cannot see

Two schedules finish in 41 hours. Only one keeps its sequence when processing times move 5 %.

On the makespan line, two schedules from the same production-scheduling model look identical, and both meet every constraint. Pick one and see where DRI’s 28 re-solves land: one schedule stays in its basin, the other scatters into near-equal alternatives.

Production scheduling

Schedule A

Makespan 41.0 h · every constraint met

92/ 100

Proceed

  • Sensitivity94
  • Degeneracy90
  • Regime shift95
  • Solver consistency89

Under the probes

  • Processing times +5 %: the same job sequence; only start times move.
  • Due-date relaxation 1–10 %: no change of sequence.
  • Nearest alternative sequence: 3.1 % longer makespan.
  • Second solver: the same sequence.

Where the 28 re-solves landillustrative

chosen schedule
  • seed_variation × 10
  • noise_injection × 5
  • subsample × 5
  • constraint_relaxation × 3
  • parameter_sweep × 5

92 and 28 are not probabilities. A score of 92 does not mean a 92 % chance that the schedule is right; it means the schedule held under the 28 changes that were tested, and the report lists them. Schedules, scores, positions and outcomes in this example are invented to show the method; the probe groups and their counts are the generator’s defaults.

How it works

Four steps, after your solver and before your decision.

DRI is not exclusively quantum. It works with the solver you already run, classical or simulated, and could include a quantum backend when one is available. It does not replace the solver.

  1. Step 01 · classical, simulator or quantum backend

    Your optimiser answers

    DRI starts from the solution your own solver produced. It does not replace the solver and does not need a quantum computer; it calls the solver again.

  2. Step 02 · five probe groups

    The decision is re-solved 28 times

    Each probe changes one thing by a known amount: the random seed, noise on the inputs, a share of the data, a constraint, a parameter. Every probe records what it changed and which backend solved it.

  3. Step 03 · each 0–100, each weighted 25 %

    Four indices measure the answers

    Sensitivity, degeneracy, regime shift and solver consistency each answer one question about how the re-solved answers relate to the original.

  4. Step 04 · proceed · caution · review

    A score, a band and a report

    The weighted score falls in one of three bands. The report names the fragility drivers, recommends actions and lists the evidence.

    score = 0.25 × sensitivity + 0.25 × degeneracy + 0.25 × regime + 0.25 × consistency

The four indices

Each index asks one question of the 28 answers.

Every value in this panel is a definition from the DRI code: what feeds the index, how it becomes a score and where its warning starts. Each rating also carries a confidence value and names its top fragility drivers, so the number never travels without the reason behind it.

How far does the result move when the input moves?

  • seed_variation
  • noise_injection
  • subsample

Sensitivity

score = (1 − sensitivity) × 100

The mean deviation of the objective from the baseline across seed, noise and subsample probes, plus its spread, capped at one. Above 0.7 a small change in the input produces a large change in the answer.

Where the warning starts

Above 0.7, a small change in the input moves the answer a long way.

Three readings, illustrative scores

What 85, 55 and 25 would mean for the person who signs.

The bands are fixed in the code. The scores here are examples, chosen to sit in the middle of each band.

  • Band 70–100 · Proceed

    The plan holds

    The answer stayed put across the probes, few alternatives came close, and the solvers agreed. Act on it, and keep the report.

    Illustrative score

    85out of 100,Proceed

  • Band 40–69 · Caution

    The plan holds, with named weaknesses

    Typically one index pulls the score down, for instance several near-equal alternatives. The report names it and recommends what to check before acting.

    Illustrative score

    55out of 100,Caution

  • Band 0–39 · Review

    Do not act before a review

    The answer moves with its inputs, changes structure or depends on the solver. A proof run on this decision concludes FRAGILE_BLOCK.

    Illustrative score

    25out of 100,Review

The detail behind a rating

The probes, the drivers and the report.

Defaults from the code, not from a slide.

The generator builds the probe set from a fixed configuration. Seed count can be set from 1 to 100 and relaxations from 1 to 20; backends default to classic and simulator.

  • seed_variation: 10 seeds drawn from base seed 42. Tests solver randomness.. Default count 10.
  • noise_injection: 1, 5, 10, 15, 20 % noise on the inputs. Tests measurement error in the data.. Default count 5.
  • subsample: 90, 80, 70, 60, 50 % of the data. Tests dependence on particular records.. Default count 5.
  • constraint_relaxation: 1, 5, 10 % relaxation. Tests constraints that bind hard.. Default count 3.
  • parameter_sweep: points around the baseline parameters. Tests parameter dependence.. Default count 5.
  • scenario: named what-if cases. Tests structured alternatives.. Default count as defined.

Available today

A method with code behind it, not a service on sale.

The scoring, the probe generator and the proof report exist in the QCOS codebase and are exercised by tests. The capability is switched off in production. This page explains what the method does so that a buyer can judge it; it does not offer a rating, a pilot or a price.

Not today

Ratings

The DRI routes answer 503 capability_disabled in production. No rating has been issued to a customer and none is on sale.

Not today

Quantum backends in a rating

The generator can include quantum backends, and real QPU execution is switched off across QCOS. The method has been exercised on classic solvers and simulators.

Not today

A guarantee

A rating measures how a decision behaves under the probes that were run. It says nothing about perturbations that were not run, and the report lists which were.

Next step

Bring a decision you would like to see measured.

If your organisation acts on optimisation results and would want a reliability rating before it does, the method team will walk through the indices against your problem class. No rating is issued today; the conversation is about what one would measure.

Updates

Hear when DRI ratings become available

One message when the capability is switched on, and nothing before.

Header animation computed by SoftQuantus from synthetic optimisation landscapes; the abstract images are 3D renders made by SoftQuantus. No stock footage. Index definitions, probe defaults, bands and report structure are taken from the DRI code in QCOS. Patent application FR2513440 filed; no patent has been granted.