Research

A plan measured in evidence

What we will test, how we will judge it, and what we will refuse to claim.

Maturity roadmap

Eight phases, P0 to P7. Each closes with a test, not a date.

P0FramingCurrent
Scope
Define the questions, the claims registry, the evaluation rules and the limits.
Exit test
Reviewers can restate the scope, claims and rejection rules without ambiguity.
P1Synthetic benchmarksPlanned
Scope
Build synthetic models with known truth and fixed baselines.
Exit test
Baselines reproduce from a clean checkout and truth metrics are computed on them.
P2Solver validationPlanned
Scope
Check the wave solver against analytic cases and published references.
Exit test
The solver matches reference solutions within stated, pre-registered tolerances.
P3Open-data inversionPlanned
Scope
Invert openly licensed datasets and compare with published results.
Exit test
Results on open data are reproducible and compared with the baselines like for like.
P4Uncertainty testsPlanned
Scope
Test whether the stated uncertainty ranges contain the truth.
Exit test
Uncertainty coverage is measured on held-out cases and reported, failures included.
P5Joint inversionPlanned
Scope
Combine data types in controlled experiments.
Exit test
Each added data type shows added information in a controlled experiment, or is dropped.
P6Independent reviewPlanned
Scope
External reviewers examine the methods, the data handling and the claims.
Exit test
Reviewers sign off, or list the claims that must be withdrawn.
P7Pilot readinessPlanned
Scope
Prepare a pilot with a partner, on data they are licensed to share.
Exit test
Both sides accept a pilot protocol with pre-agreed success and stop criteria.

Programs A, B and C advance alongside the phases, each behind its own entry criterion.

A · Advanced seismic imaging and scientific computing
Entry: the solver is validated on synthetic cases (P2), then each method is compared with its baseline.
B · Multi-physics geophysical intelligence
Entry: seismic baselines exist (P3), then each extra data type is tested for added information (P5).
C · Nano, quantum and experimental sensing
Entry: a written feasibility and marine-limits review for each concept; nothing is built until a concept passes it.

How any result will be judged before it is shown.

Every method is compared with a baseline on the same data, with the same inputs and the same compute budget.

Test data are separated from training and tuning data by survey and by geology, not by random split.

We reject images without quantitative truth metrics.

Every claim on the site, with the evidence behind it. A claim appears here only once evidence exists.

The ledger is empty on purpose. It fills as results are validated and reviewed, and until then every claim on the site stays at concept.

What this program cannot do, and what remains to be shown.

No guarantee of discovery
Better images can inform decisions. They do not create resources or promise that any is present.
Resolution decays with depth
The deeper the target, the coarser the detail the data can support.
Inversion is non-unique
Different models can fit the same data, which is why we report ranges, not single answers.
Field data access is licence-bound
Real data belong to their owners. We work on synthetic and open data until a partner is licensed to share more.
AI can hallucinate structure
That is why learned components are tested against held-out geology and cannot certify results.
Drilling remains necessary
Only direct sampling confirms what is in the ground.
Shallow-water component tests are not proof of deep-water performance
A sensor or method that works in shallow water must still be shown to work in deep water.

No publications yet. Papers, datasets and code will be listed here when they are released.