gates

RELEASE VALIDATION

How a release is validated

Before anything ships, one script runs the whole set of checks below against the repository, the trained models and the live routes. The script is in the repository; so is every file it reads. You can run it against our bytes without asking us for anything.

Run it yourself

bash scripts/audit/verify_gates.sh

That script declares 165 checks. The count is read from the script itself, not written into this page — the record beside it carries the script's hash and the one-line command that reproduces the count.

/receipts/dashboard/v1/validation/checks.json

Nothing drifts quietly

Where two records in the repository describe the same thing, they are compared and any disagreement stops the release. A figure that was withdrawn cannot reappear unannotated, and a simulation whose inputs have moved on cannot be cited as current.

Every artefact is the artefact it claims to be

Content hashes on every manifest entry, a sniffer that refuses synthetic data presented as measured, a detector for citations pointing at files that do not exist, and a freshness check on the evidence behind each claim.

The physics heads stay calibrated

Thermal, crosstalk, power-delivery and mechanical correction heads are re-measured against their held-out data, as is the cross-physics head and the temperature-aware impedance head. A head that has drifted out of tolerance blocks the release rather than quietly serving.

Photonics is held to the same bar

The photonic surrogates must load and emit, the photonic design-rule suite must catch known-bad geometries, the comparison against published silicon-photonic references must hold, and an adversarial sweep must stay inside the declared envelope and flag what falls outside it.

A surrogate is checked against the truth it approximates

Scattering-parameter surrogates are compared against the boundary-element solve off the training grid; the multi-fidelity calibration is checked on geometries neither model saw; the inverse design stack is checked against its objective.

The checks themselves are attacked

Fault injection plants a defect that each check must catch, so a check that can no longer fail is found and fixed. Adversarial input fuzzing across every endpoint must produce no unhandled error, and physically impossible scattering parameters must be rejected rather than returned.

What this page does not tell you

  • How many of those checks pass. The outcome of a run depends on which corpora and which solvers the machine running it has, so a score taken on one machine and printed on a web page is not a measurement of anything a reader can use. Run the script and read your own.
  • Whether an independent party agrees. These are our own checks on our own work. No third-party audit of them has been commissioned.
  • Anything about fabricated hardware. Every result these checks guard is a computed one.

Operational envelope and disclosures

  • Evaluation workspace, not a released product.
  • Every result here is computed. We claim no fabricated through-glass-via measurement anywhere on this platform.
  • Comparisons against the literature are against values other people published, not against measurements we took.
  • The checks that guard a release are our own. No third-party audit of them has been commissioned.
  • Cadence, Ansys, Keysight, Synopsys, NVIDIA and every other company or tool named here are named only as formats and interfaces we read and write: ChipletOS has no commercial, partnership, supply, evaluation or endorsement relationship with their makers.
What state each figure on this page is in

This workspace binds no figure; figures and their receipts are on chipletos.com.

  1. LOADING none here; a figure read in the browser is listed here until its receipt is read.
  2. UNSEARCHED none here; a claim class nobody has searched for is listed here by name.
  3. SEARCHED_NONE_FOUND none here; a search that ran and found nothing is listed here with its scope and date.
  4. NO_DATA_AVAILABLE none here; a receipt that holds no value for its row is listed here.
  5. SEARCH_FAILED none here; a receipt that is not served, or does not hash to the registry, is listed here.
  6. UNRESOLVED_IDENTITY none here; two receipts about different objects offered as one figure are listed here.
  7. UNDERPOWERED none here; a figure whose sample is below the site minimum is listed here with its sample size.
  8. CONTRADICTORY none here; two receipts giving different values for one quantity are listed here, both shown.
  9. LICENSE_RESTRICTED none here; a receipt that cannot be published is listed here with its commitment.
  10. STALE none here; a receipt older than the freshness bound is listed here with its age.
  11. NOT_APPLICABLE none here; a claim class that does not apply to this page is listed here.
  12. ANALYSIS_FAILED none here; a generator that exited non-zero is listed here with its exit code.