Back to blog
Apr 7, 2026·Updated Apr 19, 2026·8 min read
Update (2026-04-19): The registry has grown to 38 entries since this post was written — Sprint 2 through 7 added 13 more retractions (Sobol atlas narrative, OpenEMS install failures, WM-811K v2 collapse root-cause, PROV 1 synthetic-baseline, Palace 5 GHz re-extraction partial closure, 3.7 GB of duplicate CSV artifacts, broken drift-sentinel hook, etc.). The "25" count below is faithful to this post's original Apr 7 timestamp; current canonical is 38, see RETRACTION_REGISTRY.md.

How We Found and Fixed 25 Inflated Claims in Our Own Portfolio

TransparencyIP PortfolioAudit

Most patent portfolios are graveyards of inflated claims that survive because nobody audits them. We ran a 10x deep audit on our own and retracted 25. Here is what we found, why we published it, and how the portfolio is stronger as a result.

Why audit your own portfolio

Buyers do due diligence. Real diligence — not the marketing kind. They run your code, recompute your benchmarks, read the comments in your test fixtures. If something is wrong, they find it. The only question is whether they find it from you, with full context, or from a third party while you are negotiating term sheets.

We chose to find it ourselves first. Over six weeks of dedicated audit work, ten reviewers independently re-derived every headline metric in all nine provisional patent applications. They read the source code that produced each number. They re-ran every benchmark from a clean clone. When numbers disagreed, they wrote a root-cause document.

The result was 25 retracted claims, fully documented in RETRACTION_REGISTRY.md. Below are four representative cases.

Example 1 — FNO yield model: R²=0.8725 → R²=0.07

Our v1 Fourier Neural Operator yield surrogate cross-validated to R²=0.8725 on a 10K sample dataset. We cited that number in over // [DRIFT C5: FNO v2 R²=0.87 RETRACTED; honest 100K image-max R²=0.642] 170 places: pitch decks, technical reports, the API documentation, the homepage. When we re-ran the same model on a wider 100K sample distribution that better represented the production process window, R² collapsed to 0.07 on pixel-wise prediction and -1.56 on image-max prediction.

The training distribution was too narrow. The model had memorized the specific operating point and could not extrapolate. This is textbook ML overfitting, and we walked into it because our cross-validation protocol used the same narrow distribution. The fix was to retrain on the wider distribution and accept the honest R²=0.50 pixel score that the model can actually deliver.

All 170+ references in the codebase were updated. The current honest number is what ships.

Example 2 — PROV 1 Physics Cliff: 142 dead claims

The PROV 1 patent family contained 142 claims tied to a "Physics Cliff" phenomenon: a 2.49x amplification of wafer warpage at azimuthal stiffness ratio k_azi = 0.80, observed in our 2D biharmonic FEM solver. We had 11,000 parametric runs supporting the cliff onset and amplitude, and dozens of dependent claims built on top.

When we Richardson-extrapolated the same problem in 3D — the physically correct geometry — the singularity disappeared. The cliff was a 2D plate-bending artifact, not a real physical effect. All 142 cliff family claims were retracted as scientifically dead. The biharmonic solver, the ILC controller, and the rest of PROV 1 survive. The cliff family does not.

Example 3 — PROV 5 REE selectivity: "ENTIRELY ASSUMED"

A PROV 5 claim cited 35.2x rare-earth selectivity for a Janus ligand against a baseline carbonate matrix. During the audit, a reviewer opened the source script and found a comment line three functions deep:

# WARNING: selectivity ratio entirely assumed
# from DFT trends, not measured experimentally
SELECTIVITY = 35.2  # placeholder pending wet-lab data

The number was a placeholder that escaped into the patent application because the placeholder was never replaced. The fix: the claim was retracted, the script was rewritten to require measured input data before producing a selectivity number, and the related patent family was rescoped to claims that survive without the disputed ratio.

Example 4 — Monroe-Newman 4× misquote

The PROV 6 solid-state battery dossier cited a Monroe-Newman dendrite suppression safety factor of "4×" based on the 1992 paper. The actual value in the paper, when re-derived from the published equations, is closer to 1.0× at the porosity we tested. A digit had been transposed in an early conversation and propagated. The retraction restated the gyroid result correctly: it passes Monroe-Newman at 60% porosity by a margin of 1.8×, which is still a strong result and the headline survives. The 4× misquote does not.

Why this matters

25 publicly retracted claims sounds like bad news, but it is exactly what a sophisticated buyer wants to see. It tells them:

  • The portfolio has been pressure-tested by the team that built it
  • Every surviving claim has been re-derived after the audit and still holds
  • The team is honest enough to publish negative results, which means the rest of the data can be trusted
  • Diligence will not produce surprise findings during a deal — the worst is already on the table

Industry average for public retractions is zero. We retracted 25 and the surviving 909 filed claims are stronger as a result. Self-disclosure beats discovery during DD every time.

The full retraction registry is available under NDA. Each retraction has a root cause document, a list of affected claims, and a description of what was salvaged. See the IP portfolio page for the public summary.

Back to all posts