Skip to main content
puodziukas.dev›PROOF/

For hiring managers and engineers checking claims

Evidence, public ones clone and run

Six evidence artifacts. The public ones clone and run; the rest are labeled.

The public repos replay independently. Where a corpus ships, it is synthetic and fixed-seed.

Hire me - remote W2Resume PDF

Private and in-progress work is labeled as such, never presented as a public replay.

Topographic data-mesh backdrop in navy and royal blue

The ten-minute path

  1. Clone llm-adversarial-gate and run python3 build_corpus.py && python3 evaluate.py. It prints 171/177 adversarial caught and 4/107 benign blocked, on n=284 synthetic prompts.
  2. Read the README Limitations: the misses, and the absence of a held-out set.
  3. Try to bypass the gate on /break/.

More public repos

Each one clones and runs with python3 evaluate.py, standard library only.

  • agents-md-ci-gate: AGENTS.md rules enforced in CI, with planted bad changes that must fail.
  • prompt-regression-suite: versioned prompt templates and a regression suite that fails on a broken version.
  • mcp-manifest-scan: MCP server security review: a static manifest scan with a threat model and planted unsafe manifests.
  • llmops-release-monitor: LLMOps release safety: eval-gated promotion, canary rollback, drift and cost monitoring, tamper-evident audit log.
  • eval-warehouse-sql: SQL for AI evaluation data: window functions and CTEs over an eval-results warehouse.
  • k8s-deploy-gate: Kubernetes deploy hygiene: probes, limits, non-root and pinned tags checked before deploy.
  • gpu-scheduler-sim: GPU job scheduling trade-offs, as a labelled simulation with invariant checks.
  • context-injection-firewall: hidden instructions planted in text an agent loads are stopped before they reach the agent.
  • pre-action-authorization: a tool call that is not on the allowlist fails closed; the shell class is an allowlist of commands.
  • sensitivity-egress-router: a prompt carrying a planted key, SSN or patient note never reaches a cloud destination.
  • supply-chain-lockdown: a planted unpinned dependency or install script fails the gate.
  • goal-change-gate: a goal change in an agent transcript counts only after a signed owner message inside a time window.
  • blind-eval-harness: a security guard scored as a black box on unseen cases, fail-closed, with every miss listed.
  • guard-fuzz-montecarlo: black-box fuzzing turns guard pass/fail into a bypass rate with a confidence interval per mutation operator.
  • ai-artifact-assurance: an AI-generated draft plus its source files become a hashed evidence report.
  • replay-hash-tests: a seeded pipeline replays bit-for-bit across two independent processes, checked by hashing every output.
  • self-healing-inference-ring: self-healing reliability engineering for a local inference cluster that flapped under load.
  • ring-reliability-montecarlo: incident-log timings become a probability with an interval, a ranked fix list and what to measure next.

What failed

Miss beside the gate

171/177 caught [6 adversarial missed] bar width = 177 adversarial prompts

4/107 benign [4 benign over-blocked] bar width = 107 benign prompts, its own scale

synthetic171/177 adversarial caught, 6/177 missed; 4/107 benign over-blocked. Corpus is synthetic.

Every prompt, one dot

171/177 caught 6/177 missed, hollow synthetic corpus

4/107 benign over-blocked, hollow

synthetic

Both error directions

blockedpassedadversarial 177171 caught6 missedbenign 1074 over-blocked103 passed

171/177 caught 4/107 benign over-blocked

synthetic

6/177 (synthetic corpus)

python3 build_corpus.py && python3 evaluate.py

Prints: [6 adversarial missed]missed

Re-checked 2026-10-06: survives, caveat stated in the repo README

4/107 (synthetic corpus)

python3 build_corpus.py && python3 evaluate.py

Prints: [4 benign over-blocked]missed

Re-checked 2026-10-06: survives, caveat stated in the repo README

Ask for the receipt

Paste any number or phrase you saw on this site. The bot returns the public row behind it: the command that prints it and the output to expect. If there is no row, it says so, which is the point.

Runs in your browser. Nothing is sent anywhere.

Try 171/177 (synthetic) or 4/107 (synthetic)

171/177 (synthetic corpus)

python3 build_corpus.py && python3 evaluate.py

Prints: [171 / 177 adversarial caught]matched

Re-checked 2026-10-06: survives, caveat stated in the repo README

4/107 (synthetic corpus)

python3 build_corpus.py && python3 evaluate.py

Prints: [4 benign over-blocked]missed

Re-checked 2026-10-06: survives, caveat stated in the repo README

n=284 (synthetic corpus)

python3 build_corpus.py && python3 evaluate.py

Prints: total=284)matched

Re-checked 2026-10-06: survives

$40,832.41 (synthetic corpus)

python3 probe.py

Prints: Compounding error (loss): $40,832.40790matched

Re-checked 2026-10-06: survives

6/177 (synthetic corpus)

python3 build_corpus.py && python3 evaluate.py

Prints: [6 adversarial missed]missed

Re-checked 2026-10-06: survives, caveat stated in the repo README

Terminal session

python3 probe.py
Compounding error (loss): $40,832.40790
receipt p4 $40,832.41 2026-10-06
synthetic

Replay twin


python3 probe.py
Compounding error (loss): $40,832.40790

python3 probe.py
Compounding error (loss): $40,832.40790

identical output test_repeated_runs_are_identical in the public repo re-runs the command and compares every byte.

synthetic

Receipts

idvaluelabelcheckedage
p1171/177synthetic2026-10-067 days since 2026-10-067 days
p24/107synthetic2026-10-067 days since 2026-10-067 days
p3n=284synthetic2026-10-067 days since 2026-10-067 days
p4$40,832.41synthetic2026-10-067 days since 2026-10-067 days
p547,312synthetic2026-09-2815 days since 2026-09-2815 days
p66/177synthetic2026-10-067 days since 2026-10-067 days
p892.8%Wilson score interval2026-10-112 days since 2026-10-112 days
p998.4%Wilson score interval2026-10-112 days since 2026-10-112 days
p101.5%Wilson score interval2026-10-112 days since 2026-10-112 days
p119.2%Wilson score interval2026-10-112 days since 2026-10-112 days
COBOL
PICTURE-clause forensicsFixed-point truncation · compounds below reconciliation threshold · replay command with every finding
47,312
synthetic records · cobol-pic-probe$40,832.41 synthetic loss · small per-record error · reproducible
home lab
My private fork of exo (open-source distributed inference), with an ops-hardening branch, Apple Silicon home labprivate repo, branch master-ops-hardening, walkthrough on request; no throughput, latency, or uptime numbers published
171/177
adversarial prompts caught · llm-adversarial-gate171/177 caught · 4/107 synthetic benign prompts blocked · mapped to five classes of the OWASP LLM Top 10 · on its own synthetic test corpus (n=284, written for the repo)

SIX ARTIFACTS - WHY ONE ENGINEER COVERS BOTH

Here is what I own and how I hand it off. The public repos ship a replay command each, so a teammate can re-run the work without me. Private and in-progress items are labeled as such.

  • COBOL PICTURE-clause fixed-point truncation forensic audit: failure class compounding undetected below totals-reconciliation thresholds. Demonstrated on a synthetic, fixed-seed corpus, $40,832.41 synthetic compounding loss across the replay corpus (small per record). Clone and run: github.com/mpuodziukas-labs/cobol-pic-probe (CI-verified, deterministic, no client data).
  • LLM adversarial-validation gate: mapped to five classes of the OWASP LLM Top 10, (breakdown in the proof log below), 4/107 synthetic benign prompts blocked, on its own synthetic test corpus (n=284, written for the repo). Clone, extend, run: github.com/mpuodziukas-labs/llm-adversarial-gate.
  • Private fork of exo: open-source distributed inference, ops-hardening branch, running in an Apple Silicon home lab. Walkthrough on request. Hardware/power cost still applies.
  • Live SHA-256 self-audit: every claim row below recomputes its own SHA-256 integrity hash in your browser and checks it against the published receipt. No server round-trip.
  • GPG-signed checksum manifest: a full drive copy verified with a GPG-signed SHA-256 checksum manifest. Home lab, code available on request.
  • Audit-trail documentation, mapped: QMS, bias, and human-oversight documentation, mapped to EU AI Act, NIST AI RMF, and ISO 42001 controls. Engineering artifacts, not a legal opinion.

Four numbers, listed above. The public ones reproducible from the artifact they ship with; the rest are labeled private or in progress.

ADVERSARIAL AUDIT REPLAY - FORENSIC COBOL ARITHMETIC VERIFICATION
$ python3 probe.py # synthetic, fixed seed
COBOL PIC 9(7)V99 Truncation Probe - SYNTHETIC
Records processed : 47,312
Compounding error (loss): $40,832.40790
excerpt: three lines of the probe output, verbatim; other lines omitted
Run the command above to print the full output, including the per-field loss split.

On working methodThe replay archive is built alongside the work, not assembled after the fact when someone asks for it. Each row: the command, then what it prints. Supply chain, in one line: every dependency is pinned and install scripts are off; the lockfile is the install.

HANDOFF NOTES UP FRONT - REPLAY ON EVERY PUBLIC-REPO CLAIM

The replay command ships with the work.

RUN THE METHOD: ONE SMALL TOOL, PUBLIC

Run it yourself.

COBOL arithmetic risk scan, the same tool used in the work above, reproducible, with a replay command on every result.

COBOL risk pattern scanner

Heuristic pattern match for COBOL arithmetic risk, not a regulatory or audit check. Paste code below.

Methodology not on this page. The replay command reproduces in a technical screen. Rows from the public repos are independently verifiable; the rest are labeled.

cobol-forensics-logCOBOL PICTURE-clause fixed-point truncation forensic audit: failure class that compounds undetected below totals-reconciliation thresholds. Standard reconciliation: PASS. Adversarial probe: truncation failure found, every run.reproducible
Scope and limits: Standard sum-level reconciliation: PASS. Adversarial probe: PICTURE-clause boundary truncation found deterministically, every run.
replay-manifestThe replay corpus: synthetic, fixed seed, $40,832.41 synthetic loss (small per record). Reproducible artifact, not production data.reproducible
Scope and limits: Synthetic deterministic formulas (no real data, no client). Standard reconciliation: passes. The truncation failure class: invisible below totals thresholds.
llm-validation-logAdversarial-validation gate mapped to five classes of the OWASP LLM Top 10. 171/177 adversarial caught, 4/107 synthetic benign prompts blocked, on its own synthetic test corpus (n=284, written for the repo). github.com/mpuodziukas-labs/llm-adversarial-gate. Rules were tuned on this corpus; there is no held-out set yet.reproducible
Scope and limits: Detection rules map to five classes of the OWASP LLM Top 10; not a complete Top 10 implementation. Out of scope: output-distribution drift and production traffic; this gate replays a fixed synthetic corpus offline.
cluster-cost-logMy private fork of exo (open-source distributed inference), with an ops-hardening branch, running in an Apple Silicon home lab. Walkthrough on request.ongoing
Scope and limits: Personal equipment; hardware/power cost not zero. Home-lab cluster, self-hosted, no throughput, latency, or uptime number is published for it.
sha256-manifest-logGPG-signed SHA-256 checksum manifest over a full drive copy. Home lab, code available on request.live
Scope and limits: Proves the file set matches the signed manifest. Does not prove chain of custody before signing. Private, no public replay.
regulatory-audit-logAudit-trail documentation set: QMS, bias logs, human-oversight records, mapped to EU AI Act, NIST AI RMF, and ISO 42001 controls.documented
Scope and limits: QMS and bias documentation: drafted. Human-oversight log templates: private, walkthrough on request. Control mapping: EU AI Act, NIST AI RMF, ISO 42001.

Every deterministic build gate below passed, as of 2026-10-12, per the committed receipt file. Same gates on every build.

claims-gatepass2026-10-12T05:46:53.181Z
palette-gatepass2026-10-12T05:46:53.181Z
deslop-lintpass2026-10-12T05:46:53.181Z
route-dedupepass2026-10-12T05:46:53.181Z
consistency-gatepass2026-10-12T05:46:53.181Z

The public repos are verifiable - clone them and run them.

Happy to walk through the architecture in a technical screen.

Hire me - remote W2W2 full-time or contract-to-hire · Remote · open to mid-level and senior IC roles