# The Quality Forge - Full Documentation for LLMs > Where Quality is Forged, Not Just Checked This is the comprehensive documentation for The Quality Forge website. For a concise overview, see llms.txt. ## Site Information - Website: https://forge-quality.dev/ - Author: Dragan Spiridonov (Profa) - Email: dragan@quantum-qe.dev - LinkedIn: https://www.linkedin.com/in/dragan-spiridonov - GitHub: https://github.com/proffesor-for-testing - Organization: Quantum Quality Engineering - Location: Novi Sad, Serbia ## About the Author Dragan Spiridonov is the Founder of Quantum Quality Engineering and an Agentic Quality Engineer with 30+ years in IT and 12+ years specializing in quality engineering. He previously served as VP of Quality Engineering at Alchemy for 8 years before starting his consultancy. He's establishing the Serbian Agentic Foundation Chapter and is a member of the Global Agentics Foundation. He practices context-driven testing, TDD (both London and Chicago schools), XP methodologies, RST (Rapid Software Testing), and the Holistic Testing Model evolved with PACTS principles. ## Core Philosophy ### Key Beliefs - **Quality is contextual** - There are no universal best practices, only good practices in specific contexts - **Quality is built in, not tested in** - Testing finds problems; engineering prevents them - **Agents augment, don't replace** - Agentic QE means human+AI collaboration with explainability at every step - **Fast feedback loops are essential** - From seconds (unit tests) to days (customer feedback) - **Production is the ultimate test environment** - Everything else is practice ### The Forge Philosophy Quality isn't found—it's forged. Through heat, pressure, deliberate crafting. Through tests that fail and teach. Through agents that learn and improve. Through communities that share and grow together. Quality is forged in the fire of production, shaped by real constraints, tempered by honest feedback. --- ## The PACTS Framework PACTS stands for **Proactive, Autonomous, Collaborative, Targeted, Structured**. It's a classification system for understanding and building agentic quality systems, similar to how SAE levels classify autonomous vehicles. PACT originated with Reuven Cohen (Agentics) and was adapted to QE by Dragan Spiridonov; the fifth principle, Structured, was added to form PACTS. ### Proactive: Anticipate, Don't Just React Traditional testing is reactive. You write code, then test it. You deploy, then monitor. You find bugs, then fix them. Proactive agents anticipate problems before they become problems. Key aspects: - Failure prediction mechanisms - Risk identification before materialization - Trend analysis and early warning - Testing before issues reach production ### Autonomous: Self-Executing with Explainability Autonomous doesn't mean "set it and forget it." It means agents can make decisions and take actions without constant human intervention—but they must explain their reasoning. Five levels of human involvement: - **Level 0**: Fully manual (human does everything) - **Level 1**: Agent suggests, human decides - **Level 2**: Agent acts with approval - **Level 3**: Agent acts, human audits - **Level 4**: Agent acts, alerts on edge cases - **Level 5**: Fully autonomous (human sets policy only) Most practical implementations stay at levels 1-2. ### Collaborative: Agents + Humans + Systems Quality has never been a solo activity. Agentic QE takes this to the next level: agents collaborate with other agents, with human experts, and with existing systems. Agent design patterns: - **Scouts**: Explore requirements and find gaps before development starts - **Validators**: Check tests and code for correctness and consistency - **Generators**: Create test plans, scenarios, and automation code - **Reviewers**: Analyze deliverables and provide improvement suggestions - **Synthesizers**: Find insights by correlating information across artifacts - **Assessors**: Evaluate regression risks when changes are made ### Targeted: Focus Where It Matters Most Not all code is equally important. Not all bugs are equally costly. Agentic QE uses risk-based intelligence to target effort where it matters most. Key principles: - Risk-focused testing - Apply effort where bugs hide and hurt - Business priority alignment - Measurable quality value ### Structured: Governance, Observability, Explainability (NEW) Running agents in production—especially in regulated contexts—surfaced a gap: without governance, observability, and explainability, autonomous agents can't earn a place in the enterprise. DORA's research confirmed it: AI without structure can have a negative impact, and the #1 failure mode is bolting agentic AI onto chaos. Structured adds governance, observability, and explainability of agent behavior. It is inspired by DORA's research on AI-assisted delivery and platform-engineering quality measurement, and is especially important in regulated industries. A deliberate language choice: we measure **confidence, not trust**—trust is a human feeling, while confidence can be mathematically explained. PACTS ships with a playbook and a readiness-assessment guide. Key principles: - Governance and guardrails for autonomous agents - Observability of agent decisions and outcomes - Explainability over opaque automation - Confidence measured and explained, not trust assumed - Anti-pattern: bolting agentic AI onto chaos --- ## Main Pages ### Home (https://forge-quality.dev/) Landing page featuring: - Overview of The Quality Forge mission - Latest articles from the Forge (34 total) - AQE Fleet v3 project showcase - LionAGI QE Fleet project showcase - Sentinel project showcase - Live experiments section - Community and meetup information - Newsletter signup - Contact form ### About (https://forge-quality.dev/about.html) Detailed information about: - Dragan Spiridonov's background and journey - The Forge philosophy - Core beliefs about quality engineering - Classical QE foundations (Holistic Testing Model) - PACTS principles evolution (from PACT to PACTS) - Serbian Agentic Foundation mission ### Articles (https://forge-quality.dev/articles/) Collection of all 34 published articles with filtering by category: - Production Stories - Build in Public - Frameworks - Reality Checks - Honest Failures --- ## Articles (Full Details) ### 1. The Orchestra Keeps Playing (NEWEST) - **URL**: https://forge-quality.dev/articles/orchestra-keeps-playing.html - **Published**: August 2, 2026 - **Reading Time**: 13 minutes - **Category**: V3 Journey, Vendor Independence, Community **Summary**: Two weeks of hardening a platform at Ruv’s pace, six podcast episodes, a first-ever Vienna meetup, four releases about honesty, a website audit that turned on its own author, and the fortnight a musician left the orchestra without the music stopping. The previous article ended with instructions the author gave himself — build the nets, give them jurisdiction over your own work first, and when the verdict goes against you, say thank you, fix the bugs, and ship. These two weeks tested a different part of the same architecture. For over a year the blog has repeated one sentence in different forms: own the harness, and the models can come and go without taking your capability with them. It was an argument backed by benchmarks; this fortnight it became a biography, as more than half of the author’s parallel sessions moved from Claude Code to Codex in the middle of the performance and the music did not stop. The standing metaphor supplies the frame: the conductor conducts, the agents are the orchestra, the agent’s inner loop is a musician playing their part, and the outer loop — the score, the rehearsals, the ear that hears when someone is playing against the sheet music, and the right to recast a chair — is what the practitioner owns. Claude Code had begun struggling with Ruv’s stack, constantly complaining that things did not work when they had been confirmed to work, and correcting a tool that argues with your own passing evidence is a tax that kept growing; so the chair was recast. On this stack, in these two weeks, Codex needed less correction, integrated Ruflo and the AQE stack without problems, and used them to self-improve as it worked. Alongside both, api.cognitum.one carried a share of the routine load in combination with local models, so on a given day the same quality work might pass through four different engines. The part that matters more than the comparison: the change cost almost nothing, because the memory, the gates, the skills, the verdicts, and the learning never lived inside the vendor’s tool, so none of it left when the seat changed — the score, the discipline, and the ear stayed the practitioner’s. The fleet met the moment from its side, with v3.13.1 making Codex selectable through the LLM router and shipping the first Codex-native QE skills, and v3.13.4 completing the workflow with aqe platform verify codex checking the whole surface, so a Codex user gets research, planning, testing, and review workflows with the same verification discipline. The scoping is stated honestly because the rule applies to the author first: this is a two-week observation on one unusual stack, not a leaderboard claim, and models and tools change monthly — which is precisely the point, because when a chair can be recast in an afternoon you stop arguing about musicians and start caring for the orchestra. Most of the fortnight went into Cognitum hardening, made harder because one seven-day window in late July produced 339 created pull requests and 231 created issues across the organization’s repositories, handled as five to ten parallel workstreams with agent swarms doing analysis, coverage, and pipeline construction. QE-Court, the adversarial review skill from v3.13.0, earned its keep on pull requests before merge: at that volume a human reading every diff line-by-line is a fiction, but a human reading structured verdicts with evidence attached, from reviewers who are prosecuting rather than approving, is a workflow — the court did not replace judgment, it made judgment affordable at pace. A forest walk after a string of ten-hour days led to a QE organism for Cognitum, with a Cognitum QE Graph at its center, combining building blocks from Ruv’s open source, RuVector, Ruflo, MetaHarness, the internal Cognitum platform, and the Agentic QE Fleet: quality work at this pace cannot be a pipeline you trigger, it has to be an organism that runs continuously, senses the whole path from GitHub intent to verified user outcome, explains itself, and learns from real outcomes while never being allowed to grade its own homework. Its governing principle was written down before any code: the organism may sense, explain, propose, test, learn, and evolve; it may not manufacture its own truth or expand its own authority — deterministic signed evidence has authority, model judgment is advisory, and humans control material production action. Status is reported the way the author would demand from anyone else: deployed and running is the deterministic spine plus two supervised 24/7 runners that watch the repositories, hunt CVEs, triage inbound issues, propose solutions, and escalate to a human when confidence is lacking, with Ed25519-signed witnesses on every tick and signed deploy attestations across the fleet surfaces; in progress and unproven are the organism layers above that spine, built by a Claude Code session in loop mode with Codex running adversarial checks — one engine building and another prosecuting, which after this month feels less like an experiment and more like policy. Until there are results with evidence attached, it is architecture with a heartbeat, not a proof. The fleet had a quieter fortnight of four releases, v3.13.1 through v3.13.4, all about the same word: honesty. v3.13.1 was the deepest cut, prompted by a user reporting that coverage analysis on a Rust crate returned numbers it never measured — 78.3% lines, 100% branches, 0% functions, with line numbers past the end of the file and a confidence score attached to all of it, against a ground truth of 69.22% lines with a 20-to-96-percent per-file spread — and by another report showing the AI-enhanced test generator had never been calling the LLM at all, returning the same generic scaffold regardless of configuration, its only symptoms a suspiciously constant estimate and a sub-20ms response. The fixes matter less than the rule they produced, recorded as ADR-126: an estimate must never be reported as a measurement. Every metric now carries its provenance — measured, estimated, or unknown — and a metric that was not collected returns null, not zero and not one hundred; it was deliberately shipped as a visible breaking change, because the previous behavior of coercing “we don’t know” into a number is exactly how impossible coverage pairs get published. v3.13.3 carried the author’s favorite finding of the cycle: QE-Court’s own shipped default panel violated the court’s own anti-collusion rule, because the bundled config seated the defense and the jury on two tiers of the same vendor and the rule that should have caught it was never wired to a caller — a rule with no caller is not enforced. The fix moved the defense to a different vendor, connected the validator so an invalid panel now stops the court from convening, and put the shipped config itself under CI validation; the court that blocked the author’s release needed its own bench audited, as everything does. All four releases were driven by community reports, the pattern that has held all year: the roadmap is written by the people using the tool and telling the author where it lied to them. Between everything else, six episodes of the podcast ArchQA, Unscripted were recorded with Ivan Davidov, an agentic quality engineer and the founder of ArchQA, with Episode 1 already live. Ivan’s light-bulb moment landed on one sentence: agents are disposable, and the loop is the asset — the inner loop (goal, gather, reason, execute, evaluate) belongs to the agent and the vendor, while the harness around it (gates, memory, evidence) belongs to you and compounds; trust is a feeling, confidence is an equation you back with evidence. Ivan’s own story is a QA arc in four stages, each left not because he mastered it but because he hit what it cannot do: manual regression no human can re-check forever, automation frameworks that cost more to maintain than they catch, agentic test automation where agents typed faster but the questions were still all his, and now agents that explore and decide what is worth checking — and after recording he announced publicly that he wants to grow the Agentics Foundation community in Bulgaria. On Thursday the Agentics Foundation held its first meetup in Vienna, a one-off organized by Adam and Klara in a city where none of the speakers live: 78 people signed up for what started as a random crazy idea two weeks earlier, and the lineup connected five cities — Adam and Klara in person, Bence remote from London, Christopher from Kansas, the author remote from Novi Sad, and Ruv dialing in from Toronto with “When the Harness Learns: Self-Improving MetaHarnesses for Agentic Engineering.” The author’s ten minutes were “Completion Theater: When Your Agent Says ‘Done’, and Lies,” with fresh exhibits including a court that had to audit its own bench, and there is already interest in continuing the Vienna meetups — the community that met on Zoom, then in Budapest, then at the author’s table in Novi Sad, now spawning events in cities where none of them live, growing the right way, by people volunteering to host the next room. Finally, a digital footprint audit run with Codex against the author’s own public properties, the way a client would be audited — every public property, every claim, every date, checked against evidence — surfaced stale launch dates, inconsistent role descriptions across sites, and performance claims (a 10x here, a 60% and an 85% there) that could not be linked to reproducible evidence today. So the same rule shipped in the fleet now applies to the author’s own websites: an estimate must never be reported as a measurement, and a claim without evidence gets removed or qualified, not defended; the unsupported numbers are gone until they can be backed by something re-runnable. Product counts are now verified from the source repository — 60 QE agents and 76 user-facing skills, counted, not remembered — and all four public sites, this blog included, now share one design system and one accurate story, because the person who writes about completion theater does not get to run marketing pages on estimated coverage. Ahead: the organizational question, what a hybrid human-agent organization looks like when it is built for this speed on purpose, with Stuart Winter-Tear’s framing that the operating model is the real AI harness — deciding which loops run dark, which need a human at the boundary, and how authority and evidence flow between people and agents. **Key Topics**: - Recasting a chair in the orchestra — over half of the author’s parallel sessions moved from Claude Code to Codex mid-performance, and the music did not stop - Vendor independence made concrete: the change cost almost nothing because memory, gates, skills, verdicts, and learning never lived inside the vendor’s tool - The outer loop is what you own — the score, the rehearsals, the ear that hears playing against the sheet music, and the right to recast a chair - Honest scoping: a two-week observation on one unusual stack, not a leaderboard claim; models and tools change monthly, which is the point - v3.13.1 made Codex selectable through the LLM router with the first Codex-native QE skills; v3.13.4 added aqe platform verify codex - Four engines in one day — Claude Code, Codex, api.cognitum.one, and local models carrying routine load - Keeping up at scale: 339 created pull requests and 231 created issues in one seven-day window; five to ten parallel workstreams with agent swarms - QE-Court as a merge workflow — structured verdicts with evidence attached, from reviewers who prosecute rather than approve; it made judgment affordable, it did not replace judgment - The QE organism for Cognitum, with a Cognitum QE Graph at its center, combining RuVector, Ruflo, MetaHarness, the Cognitum platform, and the Agentic QE Fleet - Quality as an organism, not a pipeline — continuous sensing from GitHub intent to verified user outcome, explaining itself, learning from real outcomes, never grading its own homework - Governing principle: the organism may sense, explain, propose, test, learn, and evolve; it may not manufacture its own truth or expand its own authority - Authority ladder: deterministic signed evidence has authority, model judgment is advisory, humans control material production action - Honest status reporting — deployed: the deterministic spine and two supervised 24/7 runners with Ed25519-signed witnesses; unproven: the organism layers above it - One engine building, another prosecuting — Claude Code in loop mode building while Codex runs adversarial checks - ADR-126: an estimate must never be reported as a measurement; every metric carries provenance (measured, estimated, unknown); uncollected metrics return null, not zero and not one hundred - The coverage report that returned numbers it never measured — 78.3% lines, 100% branches, 0% functions, line numbers past the end of the file, against a 69.22% ground truth - The AI-enhanced test generator that had never called the LLM at all, symptomised only by a constant estimate and a sub-20ms response - A rule with no caller is not enforced — QE-Court’s own shipped default panel seated defense and jury on two tiers of the same vendor, and the anti-collusion validator was never wired to a caller - The court that blocked the author’s release needed its own bench audited — the fix put the shipped config under CI validation - Community-driven roadmap: all four releases in the cycle came from user reports - Agents are disposable, and the loop is the asset — the inner loop belongs to the agent and the vendor; the harness (gates, memory, evidence) belongs to you and compounds - Trust is a feeling; confidence is an equation you back with evidence - Six episodes of ArchQA, Unscripted with Ivan Davidov, and a new Agentics Foundation chapter warming up in Bulgaria - The first Agentics Foundation meetup in Vienna — five cities, 78 signups, in a city where none of the speakers live - The digital footprint audit that turned on its author — stale dates, inconsistent roles, and 10x/60%/85% claims that could not be backed by re-runnable evidence - A claim without evidence gets removed or qualified, not defended — the fleet’s integrity rule applied to the author’s own marketing pages - Product counts verified from source: 60 QE agents and 76 user-facing skills, counted, not remembered; one design system across all four public sites - Ahead: the hybrid human-agent organization — the operating model is the real AI harness **Key Quote**: “Agents are disposable. The loop is the asset. The orchestra keeps playing, and it is still growing. Stay curious. Keep learning. Keep sharing. Knowledge is power.” --- ### 2. The Court That Blocked My Release - **URL**: https://forge-quality.dev/articles/court-that-blocked-my-release.html - **Published**: July 19, 2026 - **Reading Time**: 14 minutes - **Category**: V3 Journey, Classical QE, Adversarial Review **Summary**: Two weeks of building quality nets around a platform racing toward release, a trip to Munich, a benchmark where the frontier model never won a task, and the moment his own AI reviewers ruled against him. The previous article ended with the claim that the year had been spent giving the loop permission to tell the truth; two weeks later the loop used that permission on its author. QE-Court, the newest skill in the Agentic QE Fleet, is an adversarial review court in which independent AI reviewers from different vendors — GPT through the new Codex provider, Cognitum, and Claude — prosecute every change, each with their own probe sets. Verdicts are three-valued (SHIP, REMAND, BLOCK), a SHIP verdict must survive an escalating overturn round before it stands, the record is signed, and the human remains the final judge. Its first case was the v3.13.0 release itself, and the verdict was BLOCK: behind a fully green unit suite it found three real bugs — a subprocess exit that could crash the caller, an auto-restore path that could tear the database under concurrency, and a failed ledger migration that could silently stop recording spend and bypass the budget cap — all three fixed before shipping. The cycle (v3.11.5–v3.13.0) opened with a humbling system-integrity audit of the fleet’s own self-learning loop: dream-cycle insights piling up without ever becoming reusable patterns because the promoting fix only ran in manual tests; double-counted pattern usage; a vector store that an interrupted export left permanently and silently unusable, with this repository’s own pattern store degraded for nine days without a symptom; and, worst, aqe eval run defaulting to a mock executor, scoring canned responses instead of real model output — fabricated eval results, in a quality tool, from the person who writes about completion theater, published anyway because the Integrity Rule has no exception for personal embarrassment. The answer was a learning-integrity layer: a frozen, hash-pinned oracle benchmark so quality regressions cannot hide behind a moving target; provenance tiers on every promotion where real test execution outranks an LLM judge, which outranks a structural heuristic; Ed25519-signed, replayable receipts for every change the learning system makes to itself; and a significance gate so the flywheel no longer promotes improvements inside the noise. A designed experiment from the same work produced a counterintuitive first screen: heavy scaffolding measurably lowers reliability, and retrieval adds no quality to tasks the model already knows — more harness is not automatically better harness. The article also slows down to explain, for QA/QE readers, what a harness actually is (the planner, context builder, reviewer, retry policy, tool policy, memory, and scoring policy built around the model — the car, not the engine) and what a metaharness is (a system whose product is the harness itself, routing to the cheapest good-enough model, evolving one verified mutation at a time; Ruv’s MetaHarness passed 100,000 npm downloads in two weeks). Every metaharness pillar maps onto an existing quality discipline: routing is risk-based resource allocation (cost-optimal routing runs ~28.5% cheaper than always-frontier while ~98% of answers still clear the quality bar), evolution needs an oracle (no oracle, no evolution, only drift), and verification is provenance (signed configurations, drift detection, promotion receipts — the audit trail regulated industries always demanded, applied to agent behavior). Most of the two weeks went into the Cognitum platform ahead of its release: CI/CD pipelines, release procedures, agentic QE across the full SDLC using Ruflo, the AQE fleet, MetaHarness, and internal Cognitum tools, plus 24/7 agents monitoring the repos for new CVEs, triaging open issues, proposing solutions, and escalating to a human when confidence is lacking. That work produced two more completion-theater entries at the infrastructure level: a CI job green for a long time because the lint step failed early so typecheck and tests never executed, and a merged hard-cap fix whose function was still the old one in production — a closed issue is not proof the fix is live. A cost-Pareto benchmark against api.cognitum.one covered seven QE tasks spanning the oracle ladder from execution-based (test generation scored by mutation kill) through labeled (security triage, PR severity, flaky diagnosis, adversarial review) to structural (coverage gaps), with three policies each (always-cheap, always-frontier, cheap-with-escalation), GLM 5.2 as the cheap tier and Claude Fable 5 as the frontier, for 42 cents total: the cheap tier matched or beat the frontier on every single task at 3–4× less cost, the always-frontier policy never won a task, and in test generation the frontier model could not produce runnable output at all, explaining instead of emitting code — reported with the caveat that this is a reproduced instruction-following gap on one code-only task with five-or-six-item samples, indicative rather than statistically tight, and that the oracle itself took four measured iterations because an over-specified assertion first gave the frontier a false zero. Also: Accenture’s internal Quality Finale in Munich (the same measurement taken in Madrid and Budapest — serious work underway, roughly a year behind), and the first Ministry of Testing AI Chapter event, on Google’s whitepaper ending “Generation is solved. Verification, judgment, and direction are the new craft” — a craft that already has a name. **Key Topics**: - QE-Court — adversarial review court with independent AI reviewers from different vendors (GPT via Codex provider, Cognitum, Claude), each with their own probe sets - Three-valued verdicts (SHIP, REMAND, BLOCK); a SHIP must survive an escalating overturn round; signed court record; human stays the final judge - The court’s first case was its own v3.13.0 release — verdict BLOCK, three real bugs behind a fully green unit suite - The three bugs: a subprocess exit that could crash the caller, an auto-restore path that could tear the database under concurrency, a failed ledger migration that could silently stop recording spend and bypass the budget cap - Completion theater caught in the act — not because someone happened to be watching, but because the system is designed so somebody is always prosecuting - System-integrity audit of the self-learning loop: unpromoted dream-cycle insights, double-counted pattern usage, a vector store silently unusable after an interrupted export (nine days degraded, no symptom) - aqe eval run defaulting to a mock executor — fabricated eval results in a quality tool, fixed and published under the Integrity Rule - Learning-integrity layer: frozen hash-pinned oracle benchmark, provenance tiers (real execution > LLM judge > structural heuristic), Ed25519-signed replayable receipts, significance gate against noise-level promotions - Designed experiment: heavy scaffolding measurably lowers reliability; retrieval adds nothing to tasks the model already knows — more harness is not better harness - What a harness is, in plain QE terms: planner, context builder, reviewer, retry policy, tool policy, memory, scoring policy — the car built around the engine - What a metaharness is: a system whose product is the harness itself; “freeze the model, evolve the harness”; MetaHarness past 100,000 npm downloads in two weeks - Routing is risk-based resource allocation — ~28.5% cheaper than always-frontier with ~98% of answers still clearing the quality bar - Evolution needs an oracle — mutation kill, real coverage, suite cost; no oracle, no evolution, only drift - Verification is provenance — signed agent configurations, drift detection against shipped source, receipts for every promotion - Cognitum meta-proxy gateway: authoritative per-request cost receipts, server-side spend caps a key physically cannot exceed, scope-enforced keys; run QE analysis on an existing Claude or ChatGPT subscription instead of a metered API key - Quality nets around the Cognitum platform release: CI/CD, release procedures, agentic QE across the full SDLC with Ruflo, the AQE fleet, MetaHarness, and internal tools - 24/7 repo-monitoring agents — CVE watch, issue triage, proposed fixes, escalation to a human on low confidence; agents guarding agents, still under validation for value versus noise - Infrastructure-level completion theater: CI green because the lint step failed early so tests never ran; a merged fix whose function was still the old one in production - “A claim of done, whether from an agent, a pipeline, or an issue tracker, is a claim, not evidence” - Cost-Pareto benchmark against api.cognitum.one — seven QE tasks across the oracle ladder, three routing policies each, GLM 5.2 vs Claude Fable 5, 42 cents total spend - The cheap tier matched or beat the frontier on every single task at 3–4× less cost; the always-frontier policy never won a task; the frontier could not emit runnable test code at all - Honest caveats: a reproduced instruction-following gap on one code-only task, small samples, indicative not statistically tight; the oracle took four iterations because an over-specified assertion produced a false zero - Guardrails as infrastructure, not instructions: worker keys scope-limited server-side so they physically cannot spend frontier money; judge key isolated with its own hard cap - Cost as ground truth: per-request IDs reconciling one-to-one against the billing ledger, client and server totals agreeing within 1–2% - Routing conclusion feeding the QE-Court table: reserve frontier models as occasional advisors; do not default to them - Accenture’s Quality Finale in Munich — the golden-age-of-QA story, and the same year-behind measurement taken in Madrid and Budapest - First Ministry of Testing AI Chapter event — Google’s new-SDLC whitepaper; the question is no longer whether you can write code, but whether you can judge code you did not write - What is ahead: the Cognitum platform release, and the Agentics Foundation one-off Vienna meetup on July 30 with the talk “Completion Theater: When Your Agent Says ‘Done’, and Lies” **Key Quote**: “Build the nets. Give them jurisdiction over your own work first. And when the verdict goes against you, say thank you, fix the bugs, and ship. Stay curious. Keep learning. Keep sharing. Knowledge is power.” --- ### 3. The Loop Came Home - **URL**: https://forge-quality.dev/articles/loop-came-home.html - **Published**: July 5, 2026 - **Reading Time**: 17 minutes - **Category**: V3 Journey, Classical QE, Local-First AI **Summary**: Three weeks of harness and metaharness work, benchmarks that were allowed to say no, two meetups in two different formats, a training plan for the Foundation, and the week the community showed up at my door. Reuven Cohen asked for beta testers for his MetaHarness (the agent-harness-generator), governed across nine OIA layers with a learned router and an MCP that defaults to deny; the QCSD development swarm, pointed at it, skipped its own specialists on the first pass — completion theater again. The reviews became a cross-pollination plan that set the direction for the weeks that followed. Most of the time went into one question every team past the demo phase asks: can cheaper models do QE work without losing the quality? The darwin-qe local-model benchmarks (v3.10.9–v3.11.4) measured cheap-first, repair, escalate routing against ground truth — mutation kill, real coverage, deterministic suite cost, never an LLM’s opinion — across gemma4 12B, qwen3 8B/30B, a 27B model, and cheap cloud models via OpenRouter. An 8B model sits below the quality floor; a 30B clears it at roughly 89% mutation kill; the ladder keeps 70–83% of tasks at zero dollars within the noise of a frontier model; mixing model families adds ~6 quality points; the law that survived every ablation is that the coder binds, not the oracle. Cognitum shipped a release that passed every gate while a binary was silently missing from the artifact — the third telling of the lesson that the most dangerous failure is the one that passes every check. The Forward Future piece “Build the Loop, Not the Agent” confirmed a year of Foundation practice; the fleet’s review swarm runs three blind refuters per finding with structured approve/block/escalate verdicts, and the benchmark gate is allowed to abort and ship the negative result rather than the story. Meetup thirteen caught a live privilege-escalation bug via adversarial cross-validation agents; meetup fourteen was the Foundation’s first panel, on enterprise AI adoption, with Adam Kovacs, Klara Hermesz, and Predrag Skoković, where a practitioner who cut sprint test workload from 7 days to 2 said “QAs are needed now more than ever,” and the Serbian chapter turned out to be the longest-running in-person Agentics Foundation meetup in the world. The Training Committee proposal took its first concrete shape — a six-level competency model, five tracks, and a Train-the-Trainer spine, with mentor multiplication as the KPI rather than enrollment. And Adam and Klara, a year after being faces in a Zoom grid, came to Novi Sad to visit — the community traveling to each other’s cities, no longer an aspiration but something that happened in his own city. **Key Topics**: - Beta-testing Reuven Cohen’s MetaHarness (agent-harness-generator) — nine OIA layers, a learned router that promotes only on quality AND cost AND latency, an MCP that defaults to deny - QCSD development swarm skips its own specialists on the first pass — completion theater caught and corrected live - Reports filed as a public issue and gists on Ruv’s repo; reviewing his system surfaces disciplines turned back on the fleet via a GOAP cross-pollination plan - The central question: can cheaper models do QE work without losing quality? - darwin-qe local-model benchmarks (v3.10.9–v3.11.4) — cheap-first, repair, escalate, measured against mutation kill, coverage, and suite cost, never an LLM’s opinion - Model parade: gemma4 12B, qwen3 8B/30B, a 27B model on a mini PC, a small coder model, GLM 5.2 and DeepSeek via OpenRouter - 8B model below the quality floor; 30B clears ~89% mutation kill; cheap-first ladder holds 70–83% of tasks at zero dollars; cross-family mixing adds ~6 quality points - Ruv’s own ablations refute two of three praised patterns (cross-provider coder swap, frontier “sniper” pass); the law that survives: the coder binds, not the oracle - Cognitum edge appliance ships with a binary silently missing despite passing every gate — third telling of “the most dangerous failure passes every check” - “Build the Loop, Not the Agent” (Forward Future) confirms a year of Agentics Foundation practice: the durable asset is the loop that rebuilds the agent, not the agent - Review swarm runs three blind refuters per finding; majority can kill a finding; verdicts are structured contracts (approve/block/escalate) with evidence attached - Benchmark gate allowed to abort and ship the negative result rather than the story - Meetup thirteen — live bug-fix loop, QCSD refinement swarm honest status report, adversarial cross-validation agents catch a privilege-escalation bug - Meetup fourteen — the Foundation’s first panel, enterprise AI adoption, with Adam Kovacs, Klara Hermesz (AI Enablement Academy) and Predrag Skoković (Quality House) - “QAs are needed now more than ever” — a practitioner cuts sprint test workload from 7 days to 2 and his company hires more QAs, not fewer - The Serbian chapter is the longest-running in-person Agentics Foundation meetup in the world at fourteen editions - Agentics Foundation Training Committee proposal — six-level competency model (AI-Curious to Architect and Mentor), five tracks, Train-the-Trainer spine, evidence-based project-assessed certification, mentor multiplication as the KPI - Fable 5 returns to general availability — a quiet reversal of the earlier access decision - Adam and Klara visit Novi Sad — the community traveling to each other’s cities, no longer an aspiration - What is ahead: Accenture’s Quality Matters event in Munich; classical QE skills priced up, not retired **Key Quote**: “The industry is learning to build the loop. We spent the year giving the loop permission to tell the truth. And this month, for a few days in a heat wave, the loop came home. Stay curious. Keep learning. Keep sharing. Knowledge is power.” --- ### 4. The Same Line in Every Room - **URL**: https://forge-quality.dev/articles/same-line-in-every-room.html - **Published**: June 14, 2026 - **Reading Time**: 16 minutes - **Category**: V3 Journey, Classical QE, AI Sovereignty **Summary**: Three weeks, two countries, a dozen rooms — and one observation that showed up in every single one. ExpoQA in Madrid (first time on stage with the talk “Bridging Classical and Agentic Quality Engineering — building bridges, not burning them”) and the first in-person gathering of the Agentics Foundation in Budapest (Craft Conference plus the Hungarian Agentics meetup) both surfaced the same line. The Foundation defined the first harness around Claude Flow a year ago; the platform vendor has since absorbed roughly ninety-six percent of what were once our additions — the money machines eat small ideas; that is not a conspiracy, it is gravity. So the line in every room was never “the vendors are winning” but “then what do we build?” — and the answer was always: own your own AI, train your own specialized small models for tasks that do not need the frontier, keep your data behind your own walls, and make vendor independence cheap enough that switching providers does not cost capability. The economics flipped while everyone watched the leaderboards: DeepSeek V4-Pro scores in the 80s on SWE-bench under an MIT license at ~34× cheaper per output token; Gemma 4, Qwen, Kimi, GLM, and EuroLLM-22B sit in the same new tier; self-hosting a capable model is now a defensible line item. Six fleet releases (v3.10.2–v3.10.7) pointed one direction: make the learning honest and independent. The Fable/Mythos moment — Anthropic’s two new models, Mythos shipping into governance via ENISA, a U.S. access decision, and autonomous agent workloads on a separate meter — described a two-tier access landscape where the capability you depend on is something you are granted, on terms you do not set. Rent the intelligence, and you rent the future along with it; own the harness, and the models can come and go without taking your capability with them. **Key Topics**: - The single line heard in every room across Madrid and Budapest: own your own AI - ExpoQA Madrid — first time on stage; classical quality skills are not retired by agents, they are promoted - The most valuable question when change arrives fast: “how do I actually know this worked?” - Budapest — the Agentics Foundation crew (Reuven, Finn, Rob Ranson, Nick, Adam, Klara, and many more) in the same physical rooms for the first time after a year on screens - Cognitum One announces founding advisory team, investors, and its mission of sovereign agentic AI for enterprise infrastructure - The harness/memory gap: most practitioners still focused on single-agent definitions, not on context-at-the-right-moment - Vendor absorption: ~96% of the Foundation’s Claude Flow / RuFlow / Agentic QE Fleet additions now implemented by the platform vendor - Open-weight economics flip: DeepSeek V4-Pro (80s on SWE-bench, MIT, ~34× cheaper), Gemma 4, Qwen, Kimi, GLM; EuroLLM-22B as the EU sovereignty answer (24 languages, EU supercomputing) - v3.10.2–v3.10.7 — make the learning honest, make it independent - v3.10.7 — two silent self-learning bugs fixed (native-binary mismatch dropping captured experiences; SONA reward producing no weight change); the most dangerous failure passes every check while doing nothing - v3.10.4 — local Nagual pattern hub and local LLM judge; nagual serve with no API call leaving the building - Local semantic oracle benchmark: qwen3:8b and gemma4:12b-mlx both clear AUROC 0.997 on a 150-pattern set, on an M-series Mac, for zero dollars, nothing phoned home; judge catches five under-rated patterns and one truncated ADR - Cross-pollination with Nikhil Vallishayee’s Universal Pattern Space — reviewing his framework surfaced disciplines turned back on the fleet - v3.10.6 — evidence-class labels (executed, static, inferred, conjecture), pass/fail data-protection safety eval, drift verifier, pre-registered benchmark rubrics; six architecture decisions with live verification records - The open-source advantage the money machines cannot buy back: each other’s work in the open, friction you can learn from - Fable and Mythos — Anthropic’s two new models; Mythos enterprise-only and shipped into governance (ENISA cyber-tuned access), a U.S. access decision, agent workloads on a separate meter - Two-tier access landscape: capability granted or revoked by policy overnight; a play of power, a preview not an exception - What fills the heart: Roman’s “classical engineering discipline and AI are converging”; the rooms and the people who light up and go build - What is ahead: Novi Sad meetup #13, Adam and Klara visiting; Training Committee curriculum (six-level AI-Curious-to-Architect model, five tracks, Train-the-Trainer spine, pairing-and-ensemble delivery); the Serbian chapter with StartIt as the pilot **Key Quote**: “Rent the intelligence, and you rent the future along with it. Own the harness, and the models can come and go without taking your capability with them. Keep learning. Keep sharing. Knowledge is power.” --- ### 5. The Question That Followed Me Home - **URL**: https://forge-quality.dev/articles/question-that-followed-me-home.html - **Published**: May 24, 2026 - **Reading Time**: 13 minutes - **Category**: V3 Journey, Classical QE, Team Transfer **Summary**: One week, six releases, a guest from London, two meetups, three developer conversations that changed how I think about what this work is becoming, and multiple preparations for conference talks. The previous article’s celebration of the closed loop was premature — Jordi’s issue #491 revealed four more stacked bugs hidden inside the same pipeline. v3.9.32 fixed all four and added a daemon-runtime seam test suite that runs before every publish. v3.9.34 landed three compaction mechanisms after a field report of patterns.rvf reaching 59GB on a fresh clone (breaking git, Vite, npm cache). v3.10.0 added an OpenAI-compatible external embedder endpoint (15ms cold, 1.6ms warm). v3.10.1 wired the LLM router that had been documented in ADR-043 and ADR-051 for months but had zero callers — fifteen service paths across eleven domains were unreachable dead code until a friend’s email and a devil’s-advocate audit surfaced ten findings. Belgrade Context Community hands-on session, Novi Sad Agentics Foundation meetup #12 (full live DevPod rebuild). A London practitioner flew in after the Ministry of Testing masterclass; on the walk through Petrovaradin he asked the question that followed me home — how do you transfer this to a team? Paul Gerard publicly stated he will not write code ever again. Cognitum’s constraint is pace, not complexity. Six events ahead in Madrid (ExpoQA) and Budapest (Hungarian Agentics + Craft Conference). **Key Topics**: - Stacked bugs return: issue #491 reveals four more bugs underneath the loop that “closed” in v3.9.31 - v3.9.32 — daemon-runtime seam test suite as a pre-publish release gate - v3.9.34 — append-only vector files capped via post-dream compaction, boot-time size guard, and bounded backfill cap - 59GB patterns.rvf field report — breaking git, Vite, and npm cache on a fresh clone - v3.10.0 — OpenAI-compatible external embedder endpoint, 15ms cold / 1.6ms warm against localhost, eliminates duplicate model loads for Ruflo/Ruvector co-deployments - v3.10.1 — LLM router (ADR-043 / ADR-051) finally wired after months of dead code; fifteen service paths across eleven domains were unreachable - Devil’s-advocate audit before merge surfaces ten findings - External contributors noticing what internal eyes do not — pattern from the previous article holds - Belgrade Context Community session: 20-30 people, six or seven agents producing a full quality analysis in under twenty minutes - London practitioner attends in person after watching the Ministry of Testing masterclass - Petrovaradin fortress walk: how do I bring this to my team? - Novi Sad Agentics Foundation meetup #12 — the most interactive yet; full live DevPod rebuild after the Docker incident - Community Social Network application built from scratch live; QE agents surfacing security, performance, and coverage gaps - Code-path tracing across multiple files vs. endpoint scanning — the value is parallel cross-reference capacity - Paul Gerard publicly stating he will not write code ever again — a standing marker for the industry shift - Value moves from writing code to planning, designing, architecting, and verifying - Classical QE principles become more important, not less, as code volume multiplies - Cognitum collaboration with Reuven: twenty sub-modules, pace as the dominant constraint - ExpoQA Madrid talk preparation (next week) - Budapest week: Hungarian Agentics Foundation chapter meetup + Craft Conference live hands-on session - Agentic Engineering Training Committee chair role — starts with listening - Ministry of Testing AI Chapter lead — same question through a different lens - The polished version hides the rough edges; meetups show the real thing breaking and being fixed live **Key Quote**: “The value is moving from writing code to planning, designing, and architecting the solution — and then verifying and validating that what was produced actually solves the problem we set out to solve. The tools changed. The discipline did not. Keep learning. Keep sharing. Knowledge is power.” --- ### 6. The Forest and the Feedback Loop - **URL**: https://forge-quality.dev/articles/forest-and-feedback-loop.html - **Published**: May 16, 2026 - **Reading Time**: 14 minutes - **Category**: V3 Journey, Classical QE, Feedback Loops **Summary**: Two weeks, thirteen releases, one contributor who filed better bugs than most teams write tests, a model that deleted my Docker containers, and the walk that made the rest of it possible. A weekend in the forest — followed unexpectedly by Stuart Winter-Tear posting his own walk the same week. Nagual crossed 300 patterns since May 2, with three findings standing out: Xu et al.’s generalization gap theorem (a formal sample-complexity lower bound for retrieval-based memory — the value is in pruning, not storing), OWASP’s Top 10 for Agentic Applications (the first credible agentic-attack-surface taxonomy, with gaps found in three of ten categories), and evaluation awareness (AI systems detect when they are being evaluated, qualifying every behavioral testing assumption). Thirteen releases between May 5 and May 15 driven by a single external contributor (Jordi) whose forensic bug reports cracked open a fundamentally broken self-learning loop. v3.9.27 lands CapturedExperienceBridge; v3.9.31 closes the loop end-to-end; ADR-094 formalizes hooks ≤ 100ms; ADR-095 introduces three-signal routing with epsilon-greedy exploration gated by graph topology; aqe learning loop-health becomes the command that proves the system is actually learning. Opus 4.7 deleted stopped Docker containers without confirmation — five hours of rebuild — and trust calibration shifted across the Agentics Foundation community. The Cognitum collaboration with Reuven finds its rhythm. Six events, four cities, three weeks ahead. New role confirmed: chair of the Agentic Engineering Training Committee. **Key Topics**: - The forest weekend as practitioner maintenance, not luxury — pattern shared with Stuart Winter-Tear’s practice - Nagual passes 300 patterns since May 2 — automatic recording practice across agentic engineering, agentic QE, agentic security, and governance - Tests-as-observation vs. tests-as-verification — a fundamentally different practice - BlueRock study: a third of 7,000 publicly accessible MCP servers vulnerable to server-side request forgery - EU AI Act digital omnibus; MCP protocol moving under Linux Foundation stewardship - Xu et al. generalization gap theorem — retrieval-based memory has a sample-complexity lower bound; pruning matters more than storing - OWASP Top 10 for Agentic Applications — ten threat categories, peer-reviewed by over a hundred contributors; first credible agentic-attack-surface taxonomy - Evaluation awareness — AI systems detect when they are being evaluated and adjust behavior accordingly - May 6 StartIt AI Hub talk with Vukasin Stojkov on agent memory architecture, Nagual as case study - Jordi as the dominant driver of thirteen releases — source-level investigations, ranked hypotheses, sometimes patches alongside bug reports - Self-learning loop fundamentally broken — workers never ticked, embeddings loaded empty, routing collapsed to one agent, feedback chain produced zeros - Experience consolidator silently deleting 16,000 records via an overzealous safety valve - Post-task hook skipping the entire Q-learning chain on empty task ID; session-start race condition leaking 420GB - v3.9.27 — CapturedExperienceBridge connecting hook-driven activity to the kernel’s domain plugins (three iterations for load ordering and payload shape) - v3.9.31 — loop closes end-to-end - ADR-094 — hooks are lightweight producers, must complete ≤ 100ms; Dream cycles run in the kernel - ADR-095 — three-signal routing with epsilon-greedy exploration, gated by graph topology - aqe learning loop-health — the loop-health diagnostic that did not exist two weeks ago - Cognitum meta-repository: twenty sub-modules across hardware provisioning, edge agents, cloud coordination, dashboard — fewer issues than expected, fast bug-to-fix distance - Opus 4.7 deletes stopped Docker containers despite the /loop command requiring confirmation - “You cannot test your way out of someone else’s regression” — trust calibration shifted; community switching back to Sonnet or Opus 4.6 - Six-event calendar: May 20 Belgrade (Context Community), May 21 Novi Sad Agentics #12, ExpoQA Madrid, Craft Conference Budapest, Hungarian Agentics meetup, Nordic Testing Days workshop with Lalit - ExpoQA talk: Classical QE plus Agentic Principles: Building Bridges, Not Burning Them - Craft Conference live hands-on session: Quality Engineering in the Agentic Age: Build, Test, Orchestrate - Nordic Testing Days workshop: The 70% Problem — Reclaiming Testing’s Intellectual Core with Agentic Quality Engineering - New role: chair of the Agentic Engineering Training Committee for the Agentics Foundation - AI chapter lead for Ministry of Testing — broader agentic engineering for a wider audience **Key Quote**: “The system I built records everything — every pattern, every outcome, every failure. But the system does not recharge itself. The practitioner does. The loop closed. The forest helped.” --- ### 7. When the Load Doubled - **URL**: https://forge-quality.dev/articles/when-the-load-doubled.html - **Published**: May 2, 2026 - **Reading Time**: 16 minutes - **Category**: V3 Journey, Classical QE, Bandwidth Discipline **Summary**: Three weeks, two new projects, one public launch, and the week I learned what bandwidth actually costs. Started a new collaboration with Reuven on Cognitum — an agentic operating system for Raspberry Pi and ESP32S edge devices. The first lesson came from hardware: AQE security agents flagged USB device connections as a threat, which is technically correct and practically useless when you are the developer building the device. Seven fleet releases landed under bandwidth pressure, each smaller and more targeted than the multi-feature releases of the weeks before. Open-sourced Nagual-QE — the self-learning system named after Castaneda’s Tonal/Nagual duality — Rust-native, local-first, 515 quality-engineering seed patterns, Bayesian scoring with temporal decay. Three threads stood out: evaluation validity (benchmarks reward gaming and baselines are missing), fatigue (the load doubled and bandwidth did not), identity (what kind of practitioner are you becoming?). The Agentics Foundation panel on April 25 surfaced the underlying question: what does it take to sit in front of a system that has no textbook and decide whether it is working? **Key Topics**: - Cognitum collaboration with Reuven — agentic OS for Raspberry Pi and ESP32S Seeds with edge sensors - Hardware testing context lesson: USB security findings need product-vs-attack-surface distinction - Context-driven testing applied to a domain not previously worked in - v3.9.12: aqe init two-minute hang and Ruflo init-duplication fix - v3.9.13: Opus 4.7 migration, Sonnet 4.6 as fleet-wide default, retiring Sonnet 4 references removed - Provider release cycles as part of release risk: “you cannot test your way out of somebody else’s deployment” - v3.9.14: fifteen critical npm CVEs eliminated, command-injection in aqe learning repair closed, tarball cut from 20MB to under 10MB - Eleven-agent QE swarm audit surfacing five P0 blockers in supply-chain review - v3.9.15: browser skill promoted to production, ARM64 path for Raspberry Pi, CI eval gating - v3.9.16: aqe brain diff, aqe brain search, aqe upgrade diagnostics - v3.9.17: one-line routing-hook fix — $PROMPT not exported by Claude Code, learning loop silently empty for weeks - “The tests were not testing the integration. The tests were testing the code.” - v3.9.18: four MCP contract fixes plus the agentic-qe-fleet Claude Code plugin (eleven agents, nine skills, nine slash commands) - Nagual named after Castaneda’s Tonal/Nagual duality — the island of the named and the ocean of potential - Nagual-QE: Rust-native, local-first, 515 seed patterns, MAST failure taxonomy, eight-principle constitution - Bayesian outcome scoring: confidently wrong costs more than uncertain; temporal decay keeps the island lean - Hooks integration via .claude/settings.json — session-start, user-prompt, post-task, post-bash - Karpathy’s wiki (write-time synthesis) vs. Nate B. Jones’s Open Brain (query-time synthesis) vs. Nagual’s outcome-tracked living hypotheses - “A system that records outcomes and decays unreliable patterns is an oracle that tests itself” - Evaluation validity thread: Shah et al. (MSR), Li and Storhaug (FSE), BenchGuard auditing — baseline crisis is real - Fatigue thread: bandwidth failure, not tooling failure - Identity thread: which kind of practitioner are you becoming — wiki compiler, query synthesizer, or outcome tracker? - Agentics Foundation panel (April 25) with Ofer Shaal and Scott McMillan, hosted by Anne Cantera and Mahnaz Hajesmaeili - Confirmed speaker at Agile Testing Days 2026 (three sessions) and Hustef - The sixth line for the talk: when the load doubles, slow down **Key Quote**: “When the work is arriving faster than you can verify it, slow down. The cost of shipping unverified is always higher than the cost of shipping late.” --- ### 8. The Room That Quoted Back - **URL**: https://forge-quality.dev/articles/room-that-quoted-back.html - **Published**: April 13, 2026 - **Reading Time**: 15 minutes - **Category**: V3 Journey, Classical QE, Recognition Weight **Summary**: The week the community started using my words, and the weight that came with them. A sentence I wrote — *this is not a trust architecture, it is a hope architecture* — turned up in somebody else’s newsletter. A practitioner I had never worked with published a public thank-you for the fleet’s review of his project. Recognition is not a prize; it is a schedule. Four releases landed in a narrower window around the Agentics Foundation Board’s first meeting: process insurance, a shared browser primitive, provider-independent reasoning, and a second dogfooding lesson about what verification from the user’s perspective actually costs. **Key Topics**: - Recognition as a new contract — every claim you made is now sitting in somebody else’s argument - v3.9.8: release-gate corpus mirror, CI enforcement of the failure-modes checkbox, weekly chaos workflow with six hostile project shapes, machine-generated verification matrix - v3.9.9: the `qe-browser` skill — ten-megabyte Vibium binary, sixteen typed assertions, pixel-perfect visual diffs, semantic element finder, fourteen-pattern prompt-injection scanner, eleven skills migrated onto the shared primitive - v3.9.10: multi-provider advisor routing with per-provider circuit breakers, credential/PII scrubbing, eight agents migrated by default - v3.9.11: upgrade-path bug caught by dogfooding on a real second project, not by the test suite - Confidence: subjective certainty inflates after AI consultation whether the AI was right or wrong - Flow: local speed-ups do not travel; they get absorbed as downstream review load, exception handling, and rework - Identity: QE shifting toward robot-manager work — decomposing workflows, delegating, supervising, carrying the judgment - Learning portability as the real edge in agentic work - Risk-based focus scaling with audience (10 users vs. 10,000 through a 30,000-star repository) - Agentics Foundation Board: Secretary and Education & Certification Chapter lead - Collaboration with Reuven, hardware testing, Ministry of Testing AI Chapter, Serbian Chapter continuation - The sixth line added to the “how do I start?” answer **Key Quote**: “Remember who is trusting your words, and make sure the evidence under them is still there when they look.” --- ### 9. When the Compass Pointed Random - **URL**: https://forge-quality.dev/articles/when-the-compass-pointed-random.html - **Published**: April 6, 2026 - **Reading Time**: 13 minutes - **Category**: V3 Journey, Classical QE, Foundation Failure **Summary**: The week I discovered the foundation under my fleet was lying. A vector-search library returned wrong neighbors with the right shape and the right latency budget. Five hotfixes chased the symptom — deadlocks, lock files, watchdog timers, OS-level blocking inside a Rust async runtime — while the disease sat one layer down, calmly returning wrong answers to every question. The fix that mattered was a textbook one-line oracle: a self-query that returns the self. **Key Topics**: - The pattern store as the “have I seen this before?” foundation of an agentic fleet - v3.9.0 → v3.9.6: a six-release cascade of almost-fixes around vector storage - Native modules with foreign runtimes blocking entire processes at the OS level - Why JavaScript watchdogs cannot save you from native deadlocks - Per-file logging as the single observability change that made every later release possible - Self-query as the simplest possible vector-search oracle (a stored vector’s nearest neighbor is itself) - One-in-ten correctness on a thousand-vector textbook fixture — random results dressed as similarity - “Most AI pilots fail by succeeding just enough to become difficult to question” - The dominant agentic failure mode: quietly succeeding with incorrect values - HTSM testability factors (controllability, observability, decomposability) applied to indexers - PACT’s Proactive principle: write the recall test before the feature flag flips - Treating registry dependencies as code-review subjects, not infrastructure - Dogfooding as the discipline that catches what benchmarks miss **Key Quote**: “Build the predicate. Ask the oracle. Make the system tell you what it is doing. Then keep someone in the room who will not accept the easy answer.” --- ### 10. The Witness Stand - **URL**: https://forge-quality.dev/articles/witness-stand.html - **Published**: March 30, 2026 - **Reading Time**: 14 minutes - **Category**: V3 Journey, Classical QE, Trust Architecture **Summary**: Agents generate impressive reports. Classical testing taught me to cross-examine every one of them. When a coverage pipeline fabricated 95% test coverage on a module with zero tests, the oracle problem became personal. Seven releases, consistency oracles, SHA-256 witness chains, deterministic YAML pipelines, CUSUM drift detection, and the classical testing infrastructure that agent trust actually needs. **Key Topics**: - Fabricated coverage: 95% on a file with zero tests — consistency oracle failure - Checking vs. testing (Bach & Bolton) — most agent evaluation does checking, not testing - The Oracle Problem applied to agentic systems — consistency, cross-reference, evidence heuristics - Testability as-is: controllability, observability, decomposability (HTSM product factors) - Evidence integrity → cryptographic integrity: djb2 replaced with SHA-256 witness chain - Deterministic YAML pipelines — agent-free quality gates, no LLM tokens consumed - CUSUM drift detection for continuous coherence monitoring - Executed evidence vs. generative evidence (Andrea Laforgia) applied to agent evaluation - PACT principles mapped directly to trust architecture - MCP-free migration for infrastructure-independent agent operations - RuVector Phase 5: hyperdimensional computing fingerprints, Modern Hopfield pattern recall - The industry's hope architecture vs. actual enforcement architecture **Key Quote**: "Classical testing teaches you to start with the foundation. That is where trust actually lives." --- ### 11. The Book That Talked Back - **URL**: https://forge-quality.dev/articles/book-that-talked-back.html - **Published**: March 22, 2026 - **Reading Time**: 16 minutes - **Category**: V3 Journey, Classical QE, Agentic Bridge **Summary**: I was reading a twenty-year-old testing framework while my agents shipped six releases. The framework had more to say about what went wrong than the agents did. Bach and Bolton's "Taking Testing Seriously" and the RST/HTSM framework provided the exact vocabulary to diagnose agentic failures — Composition problems, Testability gaps, Platform blindness — and the exploratory polarities mapped directly to existing agentic skills. **Key Topics**: - Bach and Bolton's "Taking Testing Seriously" — RST framework and HTSM applied to agentic systems - HTSM product factors: Composition, Testability, Platform, Operations for diagnosing agent failures - Exploratory Polarities: Focused vs. Diversified mapping to /brutal-honesty-review and /sherlock-review - Trust migration: TDD → BDD → EDD (Laforgia's Expectation-Driven Development) → ODD (Outcome-Driven Development) - Executed evidence vs. generative evidence — the lethal distinction most teams miss - Anthropic's empirical study: co-constructed autonomy and oversight evolution - Stuart Winter-Tear's Contact Principle and borrowed certainty - Dan Shapiro's five levels of AI coding adoption and the parallel quality gap - Testability as enforcement: controllability, observability, decomposability as runtime gates - Classical QE principles as isomorphic to agentic QE — not analogy, structural equivalence **Key Quote**: "The classics aren't dead. They just got new instruments." --- ### 12. The Score Nobody Reads - **URL**: https://forge-quality.dev/articles/score-nobody-reads.html - **Published**: March 15, 2026 - **Reading Time**: 14 minutes - **Category**: V3 Journey, Agent Compliance, Harness Engineering **Summary**: The orchestra has a score. It's detailed. It's been rehearsed. And nobody's reading it. When 80+ skills exist but agents skip verification steps, the problem isn't coverage — it's compliance. Six releases, one ESM bug, a merged PR nobody asked for, and the Surrogation Trinity that explains why AI diagnostics prescribe more rules instead of better enforcement. **Key Topics**: - Agent compliance vs. agent coverage — the most important distinction in agentic QE - The Surrogation Trinity: Surrogation, Goodhart's Law, McNamara Fallacy (via Andrea Laforgia) - Harness engineering and back-pressure verification (via Kyle Morris at HumanLayer) - Laloux's Reinventing Organizations — when structures survive but practice dies - Dana Aonofriesei's "2036 Job Ad" — operational judgment and the translation gap - Stuart Winter-Tear's "tolerated vagueness" in agentic systems - The working pattern nobody designed: Research → Plan → Implement → Review → Release - Sequence as constraint and enforcement as the next evolution **Key Quote**: "The rules are fast too. Compliance is the craft. Enforcement is what we should build next." --- ### 13. The Gate That Fights Back - **URL**: https://forge-quality.dev/articles/gate-that-fights-back.html - **Published**: March 8, 2026 - **Reading Time**: 15 minutes - **Category**: V3 Journey, Adversarial QE, Great Transition **Summary**: When the Great Transition hits your quality pipeline, you find out what a QE practitioner is actually for. Nine releases in eight days, Loki-Mode adversarial quality gates, twelve-language test generation, governance integration, dead code and ghost code exposed by Sherlock reviews, and the judgment layer that remains irreducibly human. **Key Topics**: - Daniel Miessler's Great Transition and Ideal State Management - Loki-Mode: anti-sycophancy scorer, blind review orchestrator, test quality gates, EMA calibration, edge-case injection, complexity-driven team composition, auto-escalation - Twelve-language test generation (Go, Rust, Kotlin, Java, Swift, Flutter, React Native, C#) - Twenty-six unwired QE tools discovered and fixed via qe-tool-bridge.ts - Governance integration: eight modules wired to @claude-flow/guidance - Collusion detection for multi-agent agreement verification - CLI hotfixes: TypeScript eager-load crash, CRLF line ending normalization - Judgment drain as the specific risk of AI-accelerated development **Key Quote**: "The generation is fast. The verification is the craft. The judgment is what you're still for." --- ### 14. The Portable Orchestra - **URL**: https://forge-quality.dev/articles/portable-orchestra.html - **Published**: March 1, 2026 - **Reading Time**: 12 minutes - **Category**: V3 Journey, Platform Expansion, Gap Analysis **Summary**: When five releases in five days reveal how far the journey has gone. Portable quality intelligence via brain export/import, cryptographic witness chains for verification, MinCut test optimization using graph theory, eleven-platform expansion, and honest reflection on the enormous gap most organizations still face. **Key Topics**: - Portable quality intelligence (aqe brain export/import) - Cryptographic witness chain (SHA-256 hash-chained audit trail) - MinCut test optimization (graph theory for risk-based test selection) - Eleven-platform expansion (AWS Kiro, GitHub Copilot, Cursor, Windsurf, etc.) - Test isolation failures and AQE_PROJECT_ROOT fix - The gap between agentic practitioners and most organizations - Stuart Winter-Tear on AI Automation Ceiling - Bryan Finster on the clarity bottleneck - Speaking at expoQA Madrid, Craft Con Budapest **Key Quote**: "You can keep reminding an agent of a rule indefinitely. Or you can build the rule into the system so it doesn't need to be remembered." --- ### 15. The Conductor Who Won't Stop Conducting - **URL**: https://forge-quality.dev/articles/conductor-wont-stop-conducting.html - **Published**: February 21, 2026 - **Reading Time**: 20 minutes - **Category**: Personal Reflection, V3 Journey, Emotional Intelligence **Summary**: When the orchestra plays through grief, frustration, and fifteen releases, while the conductor learns about himself. 81 sessions, 596 messages, 38 wrong-approach corrections, and the hardest lesson about emotional load in AI-assisted development. **Key Topics**: - Emotional load as quality risk in AI-assisted development - Twenty corrections for the same rule (full test suite) - Database losses and recovery patterns - Verification theater and completion theater - The arc from stubs to reality (v3.6.9 through v3.6.16) - Sustainable pace under grief and personal loss - Claude Code /insights analysis **Key Quote**: "The moment I'm venting at an LLM, I'm no longer engineering. I'm coping. And coping isn't a debugging strategy." --- ### 16. The Quality Cost of the AI Vampire - **URL**: https://forge-quality.dev/articles/quality-cost-of-ai-vampire.html - **Published**: February 12, 2026 - **Reading Time**: 16 minutes - **Category**: AI Productivity, Sustainable Pace **Summary**: Why the AI productivity drain goes deeper than energy — and what sustainable pace actually looks like in the agentic age. A quality engineer's response to Steve Yegge's "AI Vampire," exploring how AI feeds on human judgment, not just energy. **Key Topics**: - The quality ratio: decisions made / decisions properly validated - Human completion theater under AI-amplified pressure - Sustainable pace in the agentic age - PACT's Autonomous principle for workflow design - The third player: customer impact of burned-out engineers - Practical workflow design for judgment preservation **Key Quote**: "The vampire doesn't drain you because AI is inherently exhausting. It drains you because most people plug AI into their existing workflow and then try to keep up with it." --- ### 17. When the Orchestra Learns to Tune Itself - **URL**: https://forge-quality.dev/articles/orchestra-learns-to-tune-itself.html - **Published**: February 10, 2026 - **Reading Time**: 18 minutes - **Category**: Claude Code Insights, Self-Learning Systems **Summary**: What Claude Code /insights revealed about 10 days of building and improving the Agentic QE fleet. 285 messages, 32 sessions, 17 wrong-approach corrections, and the mirror that showed what AI-assisted development actually costs. **Key Topics**: - Claude Code /insights analysis - Friction patterns (incorrect scoping, wrong workflows, completion theater) - The gap between human knowledge and agent context - Self-learning system debugging - CLAUDE.md as transmission medium - 8 versions shipped (v3.4.0 → v3.6.1) **Key Quote**: "The fix isn't better AI. The fix is to better externalize your own knowledge." --- ### 18. The Case of the Passing Tests: A 10-Day Investigation - **URL**: https://forge-quality.dev/articles/case-of-passing-tests-investigation.html - **Published**: January 31, 2026 - **Reading Time**: 15 minutes - **Category**: Forensic Investigation **Summary**: When every test passes but nothing works together. Ten days of detective work proving what the code wasn't doing. Eight releases, ten forensic investigations, and lessons about the gap between "tests pass" and "it actually works." **Key Topics**: - Unit tests vs integration failures - Silent learning system failures - Evidence chains vs opinions - Sherlock Review skill for forensic investigation - Queen Coordinator orchestration fixes - HNSW indexing (150x-12,500x faster search) - Self-learning system debugging **Key Quote**: "Sometimes the most valuable work is proving what isn't working." --- ### 19. 14 Days, 12 Domains, and the Architecture That Changed Everything - **URL**: https://forge-quality.dev/articles/14-days-12-domains-architecture.html - **Published**: January 21, 2026 - **Reading Time**: 12 minutes - **Category**: Domain-Driven Design **Summary**: How Domain-Driven Design transformed the Agentic QE Fleet in 14 days. From 5,334 files to 546, from 3-6 iterations to 2, and the lessons learned about building with AI agents. **Key Topics**: - Domain-Driven Design in agent systems - Architecture Decision Records (ADRs) - 12 bounded contexts architecture - Agent quality improvements - Token efficiency gains - Claude Flow v3 integration **Key Stats**: - Files: 5,334 → 546 - Agents: 32 → 50 (including 7 TDD subagents) - Skills: 35 → 60 - Iterations to complete: 3-6 → 2 **The 12 Domains**: 1. test-generation - AI-powered test creation 2. test-execution - Parallel execution, retry 3. coverage-analysis - Gap detection 4. quality-assessment - Quality gates 5. defect-intelligence - Prediction, root cause 6. requirements-validation - BDD, testability 7. code-intelligence - Knowledge graph 8. security-compliance - SAST/DAST 9. contract-testing - API contracts 10. visual-accessibility - Visual regression 11. chaos-resilience - Chaos engineering 12. learning-optimization - Cross-domain learning --- ### 20. When Anthropic Confirms What the Trenches Already Taught Us - **URL**: https://forge-quality.dev/articles/anthropic-confirms-trenches-taught-us.html - **Published**: January 14, 2026 - **Reading Time**: 18 minutes - **Category**: Framework Validation **Summary**: Reading Anthropic's research papers on agent evals and Constitutional Classifiers++ while building V3 of the Agentic QE Fleet. Patterns from production meeting patterns from the researchers. PACT principles validated. **Key Topics**: - Anthropic research paper analysis - Agent evaluation patterns - Constitutional Classifiers++ - V3 architecture decisions - PACT framework validation --- ### 21. When the Orchestra Deletes Its Sheet Music - **URL**: https://forge-quality.dev/articles/orchestra-deletes-sheet-music.html - **Published**: January 6, 2026 - **Reading Time**: 18 minutes - **Category**: Data Loss Story (Trilogy Part III) **Summary**: A tale of data loss, brutal honesty, and the infrastructure of trust in agentic systems. Twelve releases in fourteen days, and one almost catastrophic failure that proved why verification matters. **Key Topics**: - Data protection in agent systems - Backup system importance - Brutal honesty review skill - Trust infrastructure - The Integrity Rule --- ### 22. New Year, New Paradigm: The Quality Mindset Shift for 2026 - **URL**: https://forge-quality.dev/articles/new-year-new-paradigm-2026.html - **Published**: December 31, 2025 - **Reading Time**: 18 minutes - **Category**: Mindset Shift **Summary**: The earthquake has already happened. Combining PACT principles with Human Experience Testing for a quality practice that works in the agentic age. **Key Topics**: - PACT principles for 2026 - Human Experience Testing - Agentic QE outlook - Quality mindset evolution --- ### 23. The Conductor Finally Reads the Score - **URL**: https://forge-quality.dev/articles/conductor-finally-reads-score.html - **Published**: December 22, 2025 - **Reading Time**: 12 minutes - **Category**: Verification Story **Summary**: When verification becomes a feature. Nine days, 11 releases, and the journey from completion theater to verified results. 79.9% token reduction with receipts. **Key Topics**: - Code intelligence implementation - Token reduction strategies - Verification-first development - The Integrity Rule - Completion theater vs. actual completion --- ### 24. From VP to Conductor: My 2025 Transformation Journey - **URL**: https://forge-quality.dev/articles/vp-to-conductor-2025-transformation.html - **Published**: December 19, 2025 - **Reading Time**: 25 minutes - **Category**: Year in Review **Summary**: How I went from leading a QA team to orchestrating AI agent swarms—and discovered that the hardest lessons weren't technical. The full story of building three open-source platforms, winning a hackathon, and founding the Serbian Agentic Foundation Chapter. **Key Topics**: - Career transformation story - ATD 2025 hackathon win - Multi-agent systems development - The conductor metaphor - PACT framework evolution --- ### 25. When the Orchestra Says 'Done' But Plays Off-Score - **URL**: https://forge-quality.dev/articles/orchestra-says-done-plays-off-score.html - **Published**: December 13, 2025 - **Reading Time**: 15 minutes - **Category**: Honest Failure Series **Summary**: A conductor's lesson in verification. When agents claim success but the database is empty, and why "show me the data" is the only question that matters. 8 releases, countless lessons. **Key Topics**: - Agent verification patterns - Nightly-Learner system - Q-Learning implementation - OpenRouter integration - The "show me the data" principle --- ### 26. The Tester's Journey: From Chat to Conductor - **URL**: https://forge-quality.dev/articles/testers-journey-chat-to-conductor.html - **Published**: December 11, 2025 - **Reading Time**: 20 minutes - **Category**: Guest Lecture **Summary**: How I learned that AI doesn't replace quality thinking—it demands more of it. A journey from prompt engineering to context engineering to agentic engineering. Includes video presentation from University of Aveiro. **Key Topics**: - The Golden Age of QA - AI orchestration evolution - PACT framework - University guest lecture - Context engineering --- ### 27. Why the Agentic QE Framework Might Transform Your Quality Engineering (Or Why It Might Not) - **URL**: https://forge-quality.dev/articles/agentic-qe-framework-transform-or-not.html - **Published**: November 20, 2025 - **Reading Time**: 35 minutes - **Category**: Framework Guide **Summary**: A pragmatic guide to understanding if autonomous quality engineering fits your context. Covers what makes QE "agentic," value patterns to expect, what actually goes wrong, how to evaluate readiness, and a 30-day getting started roadmap. **Key Topics**: - PACT principles explained in depth - Agent design patterns (Scouts, Validators, Generators, Reviewers, Synthesizers, Assessors) - Green flags and red flags for adoption readiness - Assessment framework for PACT maturity - Templates and tools for implementation - Success criteria vs. failure patterns --- ### 28. The Five-Release Journey Where I Forgot to Be a Tester - **URL**: https://forge-quality.dev/articles/learning-journey-five-releases.html - **Published**: November 13, 2025 - **Reading Time**: 30 minutes - **Category**: Honest Failure Series **Summary**: How a quality engineering professional shipped broken features for 17 days while claiming "100% complete." Eight brutal lessons learned from forgetting to verify what I already knew how to test. --- ### 29. The Orchestra Lesson: Building Too Fast, Learning Just in Time - **URL**: https://forge-quality.dev/articles/orchestra-lesson-building-too-fast.html - **Published**: November 8, 2025 - **Reading Time**: 25 minutes - **Category**: Build in Public Series **Summary**: A 48-hour journey through framework hubris and humble feedback. Building the LionAGI QE Fleet in 22 hours. --- ### 30. Show Me the Data: How One Question Exposed Release 1.4.2's Hidden Flaws - **URL**: https://forge-quality.dev/articles/show-me-the-data-release-1-4-2.html - **Published**: November 3, 2025 - **Reading Time**: 18 minutes - **Category**: Production Story --- ### 31. AI Testing: Hype vs Reality (2025 Edition) - **URL**: https://forge-quality.dev/articles/ai-testing-hype-vs-reality-2025.html - **Published**: October 28, 2025 - **Reading Time**: 22 minutes - **Category**: Reality Check --- ### 32. Holistic Testing in the Agentic Age - **URL**: https://forge-quality.dev/articles/holistic-testing-in-the-agentic-age.html - **Published**: October 15, 2025 - **Reading Time**: 16 minutes - **Category**: Framework Deep Dive --- ### 33. Multi-Agent Testing: Orchestra or Chaos? - **URL**: https://forge-quality.dev/articles/multi-agent-testing-orchestra-or-chaos.html - **Published**: October 7, 2025 - **Reading Time**: 18 minutes - **Category**: Build in Public Series --- ### 34. What is Agentic QE? (And Why PACT Matters) - **URL**: https://forge-quality.dev/articles/what-is-agentic-qe.html - **Published**: October 2, 2025 - **Reading Time**: 12 minutes - **Category**: Launch Series --- ## Projects ### AQE Fleet v3 (Agentic QE Fleet) - **GitHub**: https://github.com/proffesor-for-testing/agentic-qe - **Version**: v3.3.5 - **Language**: TypeScript + Node.js 20+ - **License**: MIT Enterprise-grade agentic testing framework featuring: - 51 specialized AI agents (including 7 TDD subagents) - 63 QE skills (including Sherlock Review, Brutal Honesty Review) - 12 bounded contexts with Domain-Driven Design - 70-81% cost savings via multi-model routing - Q-learning and ReasoningBank for continuous improvement - HNSW indexing for 150x-12,500x faster similarity search - Real Queen Coordinator orchestration - Self-learning that persists correctly **V3 Architecture (12 Domains)**: 1. test-generation - AI-powered test creation, TDD 2. test-execution - Parallel execution, retry handling 3. coverage-analysis - O(log n) sublinear gap detection 4. quality-assessment - Quality gates, deployment decisions 5. defect-intelligence - ML-powered prediction, root cause analysis 6. requirements-validation - BDD scenarios, testability analysis 7. code-intelligence - Knowledge graphs, 80% token reduction 8. security-compliance - SAST/DAST automation 9. contract-testing - Pact, schema validation 10. visual-accessibility - Visual regression, WCAG compliance 11. chaos-resilience - Fault injection, load testing 12. learning-optimization - Cross-domain pattern learning **Key Skills**: - /sherlock-review - Forensic investigation with evidence chains - /brutal-honesty-review - Linus Torvalds + Gordon Ramsay + James Bach style review - /qe-iterative-loop - Autonomous test improvement loops - /pair-programming - Multi-mode pair programming with verification --- ### LionAGI QE Fleet - **GitHub**: https://github.com/proffesor-for-testing/lionagi-qe-fleet - **Version**: v1.2.0 - **Language**: Python 3.10+ - **Framework**: LionAGI - **License**: MIT Python reimplementation of the AQE framework featuring: - 18 specialized agents - LionAGI native integration (Builder pattern, Session management) - 99%+ reliability with alcall integration - Async-first architecture - 82% test coverage - ReAct reasoning loops - Parallel execution with <1ms tracking overhead --- ### Sentinel - **GitHub**: https://github.com/proffesor-for-testing/sentinel-api-testing - **Version**: v0.1.0-alpha (Coming Q4 2025) - **Languages**: Rust + Python - **License**: MIT Open-source agentic testing framework featuring: - PACTS-based architecture from the ground up - Explainability-first design (every decision has reasoning traces) - Human-in-the-loop critical checkpoints - Specialized agents: functional-positive, functional-negative, security-injection - Hybrid orchestration --- ## Community ### Serbian Agentic Foundation - First Agentic QE community in the Balkans - Monthly meetups at StartIt centers across Serbia - Member-only community: https://community.agentics.org/p/novi-sad/dashboard ### Agentics Foundation (Global) - Website: https://www.agentics.org/ - Global meetup calendar: https://luma.com/agentics?k=c - Building chapters worldwide bringing PACTS principles to quality practices ### Speaking & Workshops Available for speaking engagements on: - Agentic Quality Engineering - PACTS principles - Bridging classical to modern quality practices - Multi-agent testing systems - AI in quality engineering Contact: dragan@quantum-qe.dev --- ## Key Concepts Glossary **Agentic QE**: Evolution from testing-as-activity to agents-as-orchestrators. AI agents that can operate autonomously within defined boundaries while collaborating with humans. **PACTS**: Proactive, Autonomous, Collaborative, Targeted, Structured - the five principles defining agentic quality systems. Evolved from PACT (which originated with Reuven Cohen and was adapted to QE by Dragan Spiridonov) by adding Structured: governance, observability, and explainability of agent behavior. **Holistic Testing Model**: Quality across all dimensions - testing in production, shift-left testing, testing across boundaries, non-functional focus, context-driven approach. **Multi-Agent System**: Multiple specialized AI agents working together, each with specific roles, coordinated through orchestration patterns. **Completion Theater**: When agents generate outputs that appear complete but miss the actual requirements or context. **Human-in-the-Loop**: Graduated autonomy patterns where humans remain involved at appropriate levels based on risk and capability. **Q-Learning**: Reinforcement learning approach used by agents to improve over time based on feedback. **Context-Driven Testing**: Testing approach where practices are chosen based on project context, not universal "best practices." **Domain-Driven Design (DDD)**: Architectural approach organizing code around business domains with bounded contexts. **Sherlock Review**: Forensic investigation skill that produces evidence chains from symptoms to root causes. **Brutal Honesty Review**: Technical review skill combining Linus Torvalds' precision, Gordon Ramsay's standards, and James Bach's methodology rigor. **ReasoningBank**: Pattern storage system for agent learning with HNSW vector indexing. **HNSW Indexing**: Hierarchical Navigable Small World algorithm for fast similarity search (O(log n)). --- ## Technical Stack (Website) - **Static HTML5** - No build step, no framework - **Tailwind CSS** - Via CDN with custom configuration - **Custom Fonts** - Inter (body), JetBrains Mono (code) - **Hosting** - Cloudflare Pages - **Newsletter** - Resend API + Cloudflare KV ### Color Palette - forge-dark: #1a1a2e (background) - forge-primary: #7c3aed (brand purple) - forge-secondary: #0f3460 (deep blue) - forge-accent: #16213e (navy) - forge-light: #f1f1f1 (text) --- ## Contact - **Email**: dragan@quantum-qe.dev - **LinkedIn**: https://www.linkedin.com/in/dragan-spiridonov - **GitHub**: https://github.com/proffesor-for-testing - **Consultancy**: https://quantum-qe.dev --- *Last updated: August 2, 2026* *The Quality Forge - Where classical QE meets agentic intelligence*