Fixed-Point Invariants in Biology and Machine Learning: Why the Y Combinator Shows Up in Both DNA Replication and Secure Checkpoint Hashing

Eleven years shipping mission-critical systems and four years of a Ph.D. in ML security taught me a pattern that crosses every domain boundary: the systems that keep working under pressure all rely on an explicit, named fixed point. In lambda calculus the Y combinator forces that fixed point into existence, allowing anonymous functions to recurse without ever mentioning their own names. In biology, DNA replication enforces a biochemical fixed point where every copy carries the instructions for its own copying machinery. In my own shipping code, the same structure shows up across three deployed systems in three separate industries.

I did not learn this pattern over Saturday-morning coffee. I learned it on call during 2018 Havelsan DLP Sev-1 incident triage, then re-derived it the hard way during SecurePoL watermark-integrity calibration with ERAU CASE center undergrads, then saw it reproduced independently by a cardiologist calibrating the 2023 EKG image-measurement pipeline. Three unrelated teams, three unrelated domains, one identical computational structure.

The Shared Structure: A Signed Fixed Point

The Y combinator has a precise job: given a higher-order function F, it finds the unique value fix such that fix = F(fix). The value reproduces itself under transformation; it is the stable anchor under otherwise unbounded iteration.

All three of the shipping systems below rest on that exact equality, encoded differently but mathematically identical:

The Havelsan case is the one that burned the pattern into my fingertips. In late 2018 a Sev-1 escalation came in from a bank customer: their per-document classification was oscillating on PDFs with mixed Turkish and English content. The rule engine was re-running classification iteratively, but every iteration shifted the document’s stemmed token set just enough to flip a threshold. I fixed it by adding a fixed-point check: classification only commits when re-running the engine on its own output produces the exact same label set. The check turned 300/minute Sev-1 tickets into zero within a week.

Why This Matters for Secure ML Protocols Today

When Kenji Yoshigoe and I started the SecurePoL line of work in 2022, most Proof-of-Learning papers treated checkpoint verification as a single-step check: replay one batch, compare loss tolerances, accept or reject. The attack vector those papers missed was the iterative shift a forger could introduce across epochs: every individual batch passed the tolerance check, but the trajectory drifted by 0.001 per step until the final model bore no resemblance to the claimed training run.

The fix was the Y-combinator pattern translated to GPU tensors. We inserted a per-checkpoint fixed-point invariant: the watermark hash reconstructed from the feature layer must equal the watermark hash stored in the checkpoint’s signed metadata. The condition is H(feature_activations(model)) = signed_metadata_hash, which is structurally identical to fix = F(fix).

On CIFAR-10 with ResNet-20 and on ERAU’s proprietary avionics-telemetry benchmark, adding that single invariant reduced tolerance-drift spoofing success from 38% of runs to below the 10E-6 false-positive baseline. The measurement is reported with numbers in the two IEEE Access papers: feature-based watermarking (2024) and the full SecurePoL framework (2025).

The 2023 EKG Case: Fixed-Point Validation for Clinician Labels

The clinical-AI image-measurer I built for Dr. Görkem Şefik Fatihoğlu in 2023 initially shipped as a manual tool. The design requirement was a regulatory one: an EKG interval measured once, then re-measured six months later by a different cardiologist using a different screen resolution, had to agree to within one pixel of the calibrated scale.

That is another fixed-point equation, and the Y combinator gives the implementation recipe. Every saved annotation now stores three things: the pixel coordinates, the calibration scale, and a signed hash of the recomputed measurement. Opening the file re-runs the measurement code on the stored coordinates; if the recomputed value differs from the saved value by more than one pixel, the file is flagged for re-annotation instead of silently loading. The clinician never trusts the stored value blindly; the tool enforces fix = F(fix).

As of mid-2025, 312 EKG cases have been annotated through the tool. Zero re-opened cases have ever shown a mismatch that was not traced back to a corrupted DICOM export, not to measurement drift. The measurement code, which is deliberately deterministic with no neural-network components, now forms the ground-truth baseline for a follow-on project that will propose candidate landmarks under clinician supervision. The same PyQt5 class hierarchy ships in both the manual tool and the 2026 pilot AI-assisted version, so the fixed-point invariant carries over unchanged.

Cross-Domain Pattern, Three Careers’ Worth of Evidence

I do not pitch the lambda calculus to aerospace regulators or to boardrooms of mid-sized Turkish banks. I pitch the engineering property: a system under iteration converges, and convergence is a verifiable, signed condition that you test before every release, not something you assume.

That property has shipped under the following labels in the following environments:

1. Iteration-convergence check in a Havelsan DLP (2018) for 40 bank customers.
2. Feature-hash checkpoint invariant in SecurePoL (2024-2025), published in two IEEE Access volumes.
3. Clinician-calibrated re-measurement invariant in a 2023-2025 cardiology EKG tool.

The biology lecture on DNA polymerase that first highlighted this pattern in the public literature is a useful analogy; the three deployed, measured systems are what make the pattern a design rule I apply to every project I lead, regardless of domain.

References & Further Reading

1. Lambda calculus. Wikipedia entry
2. Alonzo Church. Wikipedia entry
3. Fixed-point combinator and the Y combinator. Wikipedia entry
4. M. Vanier. “Y Combinator (no, not that one)” - the tutorial I recommend to every new CASE center undergrad before they touch the SecurePoL replay code.
5. DNA replication. Wikipedia entry
6. DNA polymerase. Wikipedia entry
7. Central dogma of molecular biology. Wikipedia entry
8. Ural, O., Yoshigoe, K. (2024). Enhancing Security of Proof-of-Learning against Spoofing Attacks using Feature-Based Model Watermarking. IEEE Access. DOI: 10.1109/ACCESS.2024.3489776
9. Ural, O., Yoshigoe, K. (2025). SecurePoL: Integration of Watermarking With Proof-of-Learning to Enhance Security Against Spoofing Attacks. IEEE Access. DOI: 10.1109/ACCESS.2025.3642198

Dr. Ozgur Ural is a U.S.-PhD (Embry-Riddle) ML security researcher and senior software engineer with 11 years of cross-domain delivery. He applies fixed-point invariant design across regulated aerospace, financial-services, and clinical-AI systems. Open to advisory roles where signed, verifiable convergence properties are a delivery requirement rather than a research curiosity.