From Certainty to Calibration

Full Book Description (Markdown, for the LeanPub landing page)

Most technical books perform certainty. This one interrogates it.

From Certainty to Calibration is the retrospective closing volume of the NEXUS-1 Companion Series — a return trip through six architectural decisions from the first seven volumes, each one re-examined against a single question: was this actually right, and if not, what would replace it?

Every chapter runs the same disciplined structure:

• Then — the original decision, stated faithfully
• The Mechanism, in Detail — the actual code or schema behind it, walked through before any verdict is offered
• Why This Seemed Sufficient at the Time — a fair reading of the original constraints, before any critique
• Learned — what changed the picture, labeled explicitly: did the original volume already name this limit (confirmed), or is this genuinely new (extended)?
• Now — a specific, implementable fix, not a vague direction
• What the Fix Itself Risks — the same scrutiny turned back on the book's own proposal

Six chapters, in full:

1. The Sacrifice of Scale for Readability — a reinforcement-learning Q-table's readability-versus-scale trade-off, and a hybrid design that keeps the audited core intact while adding room to grow.
2. The Sequential Bottleneck — a root-cause engine's fail-closed pipeline, and the hidden difference between "nothing caused this" and "the cause exists but was never mapped."
3. The Discipline of the LLM — ten anti-hallucination controls around a language model, and why the obvious fix for their one known gap would quietly break two of the ten.
4. The Naïve RAG Corpus — a retrieval corpus's governance disciplines, and the case for enforcing them as database constraints instead of team discipline.
5. The Illusion of Independent Witnesses — a telemetry corroboration check's independence assumption, and a table in the schema that looks like the fix but isn't.
6. The SQL That (Almost) Worked — a preview-feature dependency in a vector search query, and an argument against the obvious two-database fix.

A five-lesson survival guide distills the findings into a short, source-agnostic checklist for anyone building audit-heavy or safety-adjacent systems — each lesson closes with a worked example from outside NEXUS-1 entirely (payments, fraud detection, content moderation, medical triage, demand forecasting), as a check that it actually generalizes rather than just restating its source chapter.

The book closes with an honest tally of its own six chapters — how many findings were confirmed, how many were genuinely new, and an admission that most landed somewhere in between — because a retrospective that only ever finds fault in the original design, and never in its own proposals, hasn't actually kept its own rule.

Who this is for: senior software architects, .NET and SQL Server practitioners, and engineers building systems where an audit trail matters. Familiarity with the earlier NEXUS-1 volumes helps but isn't required — Part 0 gives enough context to follow the argument without it.

What this book is not: a defense, an apology, or a claim that the original series was wrong. Every original decision examined here is judged to have been correct under the constraints it was made with. What's new is only ever the second half of the sentence: correct, and here is what it costs now that the constraints have shifted.

Author: Grigorios Kyriakos Agathangelidis
Greek name: Γρηγόριος Κυριάκος Αγαθαγγελίδης
Also searchable as: Αγαθαγγελίδης Γρηγόριος, Αγαθαγγελιδης Γρηγοριος, Grigorios Agathangelidis.