← Feed

Posts

• View post

  COBOL MAIN MENU overview ISPF and the command line have advantages Claude and I are building out an ISPF-style main menu written in IBM COBOL 1.2.0 for Linux. It acts as a data-driven launcher for a suite of independent COBOL or python sub-applications, each living in its own sub-directory. A key design focus for this CLI application is to reduce friction for doing small tasks. Instead of loading a browser, finding a bookmark, loading the web-page, and reading through… https://felixquinihildeb...

• View post

  Why COBOL main menu search is instantaneous I asked Claude and it explains as follows: Why COBOL search is instantaneous Why not just use Python + pandas, SQLite FTS, or ripgrep? The answer is in what the COBOL + IBM STL approach eliminates entirely. No parsing, no allocation, no dispatch Every record in every .DAT file is fixed-width. COBOL declares the exact byte layout in the DATA DIVISION at compile time: 05 SEC10K-TICKER PIC X(10). ← starts at byte 1 05… https://felixquinihildebet.wordpre...

• View post

  COBOL MM RSS feed based daily report The following was generated by Claude using the COBOL MM RSS feed processing as the input data: what you do think about the readability and quality of articles selected? Daily Review: Wednesday, 2026-04-22 AI & Google Cloud Next Google's annual cloud conference delivered its most hardware-forward announcements in years: two new TPUs purpose-built for training and inference — a direct challenge to Nvidia at the infrastructure layer… https://f...

• View post

  Using local LLM as modules in a procedural workflow COBOL MM has two menu items using this approach with tangible benefit I am currently generating a SEC 10-K risk analysis menu module in COBOL MM. It is interesting because it uses Gemma 4:2eb, a very small LLM, to perform semantic analysis of each ticker's risk analysis section of the 10-K filing in a systematic manner with grading for factors. Conducting this workflow is not quick: Gemma takes about an hour if not… https://felixquini...

• View post

  COBOLMM with Gemma4 processing and Claude final analysis of TSLA 10-Q Gemma 4 analysis is not quick but there is no token cost I am testing out our 10-Q analysis menu item which follows the 10-K risk factor analysis but for 10-Q (good timing as the 10-Q just came out). While I will need to review the 10-Q manually, just by viewing the impressions and reactions I have read from the current reporting I do think this Gemma 4 procedural workflow has outstanding potential.… https://felixquinihildeb...

• View post

  COBOLMM with Gemma 4 local summary and SEC filing analysis instead of Claude Saving tokens and keeping data local yet obtaining equal quality results point to data structure and procedural workflow as being more important than model benchmarks For the purposes of a daily RSS summary, I did not notice much of a drop off in quality, readability or utility when using Gemma 4 instead of Claude. For the SEC 10-Q, 10-K, and risk analysis using Gemma 4, I cannot see any… https://felixquinihildebet.wo...

• View post

  COBOLMM SEC 10-Q analysis with Gemma 4:31b Utterly fascinating to see the divergence This entire report is generated by COBOLMM with no editing by me (not sure why some financial info is missing and will need to double check): SEC Gemma Q — Local LLM (gemma4:31b) Report + Financials Context: INTC: Q1-2026 → Q3-2025 → Q2-2025 → Q1-2025 → Q3-2024 → Q2-2024 → Q1-2024 → Q3-2023Generated: 2026-04-27 07:40 Score Data… https://felixquinihildebet.wordpress.com/2026/04/27/cobolmm-sec-10-q-analysis-with...

• View post

  COBOLMM SEC 10-Q TSLA analysis I am stunned Gemma 4 can pick up these points from SEC filings Created by COBOLMM, without edit or review by me except for my highlighting (double check financials as cash flow reporting needed processing from SEC filing): SEC Gemma Q — Local LLM (gemma4:31b) Report + Financials Context: TSLA: Q1-2026 → Q3-2025 → Q2-2025 → Q1-2025 → Q3-2024 → Q2-2024 → Q1-2024 → Q3-2023… https://felixquinihildebet.wordpress.com/2026/04/27/cobolmm-sec-10-q-tsla-analysis/

• View post

  COBOLMM continues to impress Local LLMs bring unexpected intelligence to procedural workflows I continue to test COBOLMM and chase down minor bugs: IBM COBOL seems to have an issue with SIGSEGV originating inside libvfs_lsq_64r.so, IBM COBOL's LINE SEQUENTIAL file I/O driver. It occurs during the VFS driver's OPEN INPUT initialisation phase for a large file (63 MB, 7120+ records) on a LUKS-encrypted filesystem with… https://felixquinihildebet.wordpress.com/2026/05/08/cobolmm-co...

• View post

  IBM COBOL for Linux display bug finally fixed it took days as it was masked by other things Claude documents the fix: Second Bug: Display Output Corruption — Investigation Log (2026-05-13) Summary IBM COBOL for Linux 1.2.0 has an internal output state corruption bug that manifests when programs produce large volumes of DISPLAY output (hundreds of articles / thousands of lines) via a PTY. Symptoms include: Separator lines ( https://felixquinihildebet.wordpress.com/2026/05/13/ibm-cobol-for-linux...

• View post

  So MPDOK is fast. What is going on? Gemini provided the project concept explanation and Claude and I built it This is the project idea explained by Gemini: The Project Concept: "High-Fidelity Multi-Sensor Field Reconstruction" Imagine you have a network of N = 8,192 irregularly spaced weather stations or environmental sensors scattered across a geographic region. Each sensor tracks multiple independent, continuous variables over time: Temperature, Atmospheric Pressure,… https...

• View post

  MPDOK uncovered an issue with k=3 risk estimates apparently used in Nvidia's excellent cuFolio Markowitz portfolio optimiser and Basel's failure to get beyond 1-hop. Claude and Gemini seem to agree there is a problem. I hope I am wrong and would be grateful for review by those involved in portfolio risk assessment. If MPDOK is correct, are looking at another avoidable 2008 financial crisis? https://felixquinihildebet.wordpress.com/2026/05/19/the-next-2008-financial-crisis/

• View post

  MPDOK macro contagion explained It reminds me of reinforcement learning: picking up latent structure through the data alone NotebookLM does a great job of summarising MPDOK results. Note page 8's example of the blind spot (avoided by using the signed model which incorporates positive and negative shocks). The_Global_Price_TopologyDownload https://felixquinihildebet.wordpress.com/2026/05/20/mpdok-macro-contagion-explained/

• View post

  What is the concept search engine in COBOLMM? stash, MPDOK semantic search and now the concept search engine: I can finally find things Say you do a universal search using our utility stash in COBOLMM. It has indexed Kindle clippings, podcast transcripts and audio files, web clippings of articles you saved, and your PDFs and images. “Hardware shortage” must be a common idea. In 200 ms you find no hits. https://felixquinihildebet.wordpress.com/2026/06/28/what-is-the-concept-search-engine-in-cob...

• View post

  RANSAC scanner on Android surprising original verification scanner; could it help curb Pokemon card fraud or verify watch certificates of authenticity? The RANSAC Android application currently has a fast video scanner and a more precise photo scanner running on a Google Pixel 9 Pro XL. The video scanner can instantly identify the original bill with 90%+ readings while other genuine bills are under 90% to 70%. Curiously, non-original but still genuine $20 bills score… https://felixquinihildebet...

• View post

  GP engine lab results Fable 5 is great for our labs Fable 5 was able to debug and work through our Gaussian Process engine lab. I should take a step back and underscore just how unique (to my knowledge) this is: we ran the exact same lab that previously took an 8 V100 GPU cluster to complete using older consumer hardware. We cannot find mention of this being done before. The FP64 kernel matrix on 3droad normally takes over 1.2 TB. Specifically, we took experiments from… https://felixquinihilde...

• View post

  The irony of a paper titled “Exact Gaussian Processes on a Million Data Points” Gemini explains why MPDOK outperforms the paper arXiv:1903.08114 1. The "Exact" Illusion: Why Wang et al. Had to Approximate To understand why they compromised, we have to look at what BlackBox Matrix-Matrix (BBMM) and Conjugate Gradient (CG) methods actually do. The baseline paper uses an iterative Krylov solver (Conjugate Gradient) to solve linear systems like K−1yK^{-1}y without ever… https://f...

• View post

  ban the confidently wrong! https://felixquinihildebet.wordpress.com/2026/07/15/laplace-gpc-vs-svm-using-our-gp-engine/

• View post

  Engineering GP-CVar Synthesis NotebookLM helps to summarise our gp portfolio optimisation module (edited) in our COBOLMM Portfolio studio (now using cuFolio, MPDOK and gaussian process) 1. Introduction: Beyond Historical Averages The "expected return" of a stock is often treated as a simple historical average—a naive assumption that if an asset returned 10% annually over the last decade, it will continue to do so. However, in the Portfolio Studio, we treat financial… https://...

• View post

  Status report: from first time radix sort to uncertainty aware decisions Written on 2026-08-29 by Claude and edited by Fraser: our record of progress 1. The origin It started with radix sort for CUDA Fortran — months of it, hand-written, learning GPU programming the hard way, and at the end it was correct and slow; it it was comprehensively beaten by Thrust. From that initial disappointment, a lot of lessons: A working mental model of the machine — memory hierarchy,… https://felixquinihildebet...