​This seemed sensible at first. We were thinking about voices as independent melodic lines that happen to align at measure boundaries. We even structured it this way to handle ossia staves (those little alternative passages that appear for a few measures then disappear).

But that's not how notation actually works. Measures are the organising unit. Voices are just 'simultaneous material in this measure'.

​Measures contain voices. Problem solved: 'give me everything in measure 5' is just a direct lookup. All the voices are right there.

And ossia staves? Actually clearer in this model. A staff just enters at measure 17 and exits at measure 20. Clean and explicit.

When your data structure fights the domain, everything becomes harder. Performance tricks, validation logic, traversal algorithms – they all work around the mismatch instead of with the model.

When the structure matches reality, everything becomes easier. The formatting code that sparked this whole thing? It'll be trivial now. Just iterate measures, render their voices. Done.

This is one of those changes that seems obvious in hindsight but not when you're in the middle of building. You pick a structure that appears reasonable, you build on it, and only later when you try to use it do you realise the foundation is slightly crooked.

The question is whether you have the time and space to fix it, and whether you have the honesty to admit it needs fixing.

I'm fortunate. I have the time. This project doesn't require me to pretend otherwise. And if the page keeps telling me the model is lying, I'll keep listening.

There are days in Ooloi’s development when I feel like the Donald. Not that Donald. Donald Knuth. There's something very real to that comparison, even though it can be seen as presumptuous. Why do I compare myself with a computing giant?

Knuth faced typesetting systems that were brittle, ad hoc, and incapable of scaling to real demands. He didn’t patch; he rebuilt the foundations. Out came deterministic algorithms, the box–glue model, and a system that still sets type decades later.

I’m in a similar place. Music notation software has been compromised for forty years: mutable object graphs, procedural hacks, import/export traps. It works until you open Eine Alpensinfonie or Lontano – then it collapses.

So Ooloi is built the way TeX was:

• Foundations first. Ooloi has been in development for a year, part-time. No notes have appeared on screen yet, but that isn’t delay, it’s sequence. The first phase was concurrency, traversal, contracts: all the invisible machinery that must be right before visible music can exist.
• Immutability as discipline. Where TeX used deterministic boxes and glue, Ooloi uses persistent data structures and STM. Both eliminate state leakage as a class of bugs.
• Correctness as architecture. In TeX, line-breaking is provably optimal. In Ooloi, sharing and traversal are mathematically guaranteed by immutability and functional design.

And even a year isn’t slow, considering what's been implemented in that time. In Clojure, as in Lisps generally, progress is faster, not slower, because the language doesn’t get in the way. Architectural changes that would take months in procedural or OO systems collapse into days when immutability is the default. In Lisps I feel unrestricted from the usual … bullshit.

We tested it.

50,000 musical objects: 14.5KB total on disk. Not megabytes. Kilobytes. That's not even 0.3 bytes per note.

MusicXML equivalent: ~50MB. That's 3500x larger.

Save/load time: about 250ms on a crappy 2017 MacBook Pro.

These are just the core musical objects - pitches, rests, chords, articulations. A complete score adds instruments, parts, staves, layout data, and more. But the efficiency gains indicate what's possible when you eliminate redundancy at the foundation.

The venerable technique hash-consing (1958, from LISP and symbolic computation) works. Of course it does – and how!

​Article on implications coming.

Musical scores are full of repetition. In a symphony, middle C can appear thousands of times, quarter notes dominate, and the same staccato mark is scattered across every instrument. Most notation software allocates a separate object for each of these occurrences. That means thousands of identical objects, all taking memory and I/O bandwidth for no reason.

Ooloi doesn't.

With ADR-0029, we have implemented selective hash-consing: identical immutable musical objects are represented by the same instance. The same C4, the same staccato, the same quarter note: one object, shared system-wide.

Finishing the statistics infrastructure naturally led to thinking about the next architectural milestone: the rendering pipeline. This is the mechanism that determines what happens when someone clicks into a score to add a note, and how that change propagates through the system.

The design is complete. That in itself is an important milestone, as this is the very foundation on which Ooloi's performance ultimately depends. Everything hinges upon it.

Traditional notation software recalculates entire scores when a single element changes. Dense passages in works like Strauss's Elektra bring systems to a halt because every operation is sequential and single-threaded. The reason for this is that parallelism is very difficult to do with mutable state, which is the traditional approach. Scalability often becomes an issue, with diminishing returns as a result.

Ooloi takes the opposite approach and chooses the Clojure way instead, with immutable state. With this, it is comparatively easy to distribute formatting work across all available CPU cores and use them fully and linearly.

Every user action – whether adding a single note or adjusting spacing – thus becomes part of a coordinated cascade where each stage can run in true parallel across all available cores.
​
The goal is straightforward: responsive editing even in the heaviest repertoire.

ADR-0028 specifies the pipeline in five stages, separating connecting from non-connecting elements and applying a clear fan-out/fan-in pattern.

1. Minimum Widths (Fan-out): Each measure independently calculates spatial requirements – collision boundaries only. Results are immutable and persist until the content changes.
2. Vertical Raster (Fan-in): The MeasureStackFormatter collects widths, generates result rasters, and records both minimum and ideal widths. This prepares the ground for spacing optimisation.
3. Non-Connecting Elements (Fan-out): Independent symbols – noteheads, accidentals, articulations, dynamics – are prepared in parallel. They don't need finalised atom positions.
4. Discomfort Minimisation: Iterative optimisation adjusts widths and system breaks to reduce proportional distortion. No layout changes occur after this stage. From here on, positions are fixed.
5. Connecting Elements​: Slurs, ties, beams, hairpins, and glissandi are drawn only once final positions are known. Sequential coordination ensures continuity across measures and systems.

This separation allows Ooloi to exploit parallelism where possible and enforce order where necessary.

Formatting in Ooloi is plugin-driven. Core elements such as notes and beams are implemented through the same interfaces available to extensions.
​
Plugins can participate in different stages depending on their needs:

• Spacing hook (Stage 1) for collision boundaries
• Paint hook (Stage 3) for independent elements
• Connect hook (Stage 5) for elements spanning multiple atoms

Simple articulations may use only the first two; beams may require all three. This uniform model ensures extensibility without compromising performance.

The optimisation engine measures deviation from ideal proportions across measures, systems, and pages. Improvements multiply: small gains in multiple places compound into significant overall reductions. Hard constraints such as manual breaks provide natural stopping points.
​
This replaces arbitrary iteration limits with a principled measure of quality.

Claypoole provides efficient thread-pool execution, delivering significant speed-ups over built-in Clojure parallelism. STM transactions keep operations atomic while allowing concurrency inside each stage. Cooperative cancellation ensures that rapid user input remains responsive.
​
The system treats a single user as a 'collaboration network of one'. The same infrastructure that supports multi-user editing ensures smooth interaction for individuals.

This pipeline is the structural core that should make scrolling through Elektra or Ligeti's Requiem as fluid as editing a Gnossienne by Satie.
​
The specification is complete. Implementation begins as soon as the current phase closes. Ooloi's promise of responsive, professional-scale notation depends on it.

​Scary stuff.

​Full specification: ADR-0028: Hierarchical Rendering Pipeline with Plugin-Based Formatters

​After a year of building infrastructure, it seemed worth finding out whether it actually worked.
The test was simple: create a million rests and see what happened. One operation –

​– repeated until either something broke or the numbers settled.

First experiment: one client, one thread, a million sequential calls. This was on a 2017 MacBook Pro, so not exactly cutting-edge hardware.

Every call succeeded. Average execution time was 35 microseconds. Memory stayed at 112 MB without drifting upwards. Garbage collection remained in the young generation. Thread count didn't move. Zero errors.
​
Nothing dramatic, which was the point.

Second test: four threads on one client, 250,000 calls each. Another million operations, but with some concurrency this time.

Same result. All calls succeeded, memory stayed flat, garbage collection behaved, threads were stable. Still zero errors.
​
This level of concurrency should be unremarkable for any serious system. It was.

Third experiment: ten clients, four threads each. Forty threads total – enough to make the laptop work.

Average call times went up to 246 microseconds, but everything else held steady. Memory stable, garbage collection well-behaved, threads disciplined. Error count still zero.
​
Same 2017 laptop that started the tests.

From one thread to forty, from 35 microseconds to 246, the pattern was consistent. The system handled load without breaking.
​
These were only raw API calls, not STM transactions with musical logic and collaboration on top. Those will be next. For now, the figures suggest the foundations are fast, steady, and ready to carry more.

After a year of working only on foundations – work that must succeed by disappearing – we are finally ready to turn to visible things. Music, for instance.

This Grafana dashboard shows eighteen minutes of a distributed system handling real client connections. The server sits at a 44 MB baseline and stays there: no memory growth, no thread churn, zero old-generation garbage collections. The jump to 83 MB comes not from client load, but from Grafana itself polling for statistics.

Those numbers matter because they validate architectural decisions made months ago. Functional programming principles, STM coordination, gRPC infrastructure: all the invisible work that had to be right before building anything a user might see.

Stable thread pools. Clean memory management. Efficient client handling. Nothing dramatic – just solid enough to carry what comes next.

The system no longer needs proving. Now it needs notes.

Yesterday’s post was about pitch transposition in Ooloi: how an old Igor Engraver function (printed on a t-shirt, of all places) came back to life in a new context. That work didn’t just solve the mechanics of shifting notes up and down; it reopened the larger question of playback. How do you hear microtonality and advanced humanisation without drowning in technical workarounds?

In Igor, the answer was MIDI. Notes were split across channels, bent into place, reassigned constantly. The result worked, but at the cost of complexity: a “DNA soup” of allocations and pitch-bend tricks. Ingenious, but exhausting.

Ooloi makes a different choice.

With ADR-0027: Plugin-Based Audio Architecture, we draw a line: no MIDI output in the core, no audio generation in the backend. Playback is done entirely through plugins in the frontend. If you want VST/AU instruments, SoundFonts, OSC, or any other output path, you install a plugin. The backend remains what it should be: musical data, collaboration, structure.

This is not just simplification, it’s liberation.

• Microtonality is handled directly in the data model, then played back by plugins without channel gymnastics.
• Humanisation is no longer bound to 7-bit controllers; plugins can drive modern sample libraries with full parameter precision.
• Collaboration becomes clearer: the backend never streams audio, never mangles MIDI, never compromises. Each client handles its own sound.

Put bluntly: we no longer need MIDI as an output protocol. It served its time. For professionals who need nuanced playback, orchestral realism, or contemporary techniques, we now have better tools: VST/AU plugins and beyond.

That said, MIDI output isn’t forbidden. If anyone needs it, a frontend plugin can provide it. For tonal music it will work fine. But if you want advanced humanisation or microtonality, you’ll inherit the need for all the old machinery: channel allocation, pitch-bend acrobatics, the DNA soup. That’s exactly why Ooloi itself doesn’t touch it.

The logic is simple: Ooloi’s core manages music, not sound. Plugins handle playback, and in doing so, they do it better than MIDI ever could.

The DNA soup is gone. What remains is clean, modern, and far more powerful.

​Our solution was elegantly brutal: we seized complete control of attached MIDI units and employed pitch bend to achieve microtonal accuracy. This required distributing notes across MIDI channels according to their pitch bend requirements, using register allocation algorithms borrowed from compiler technology. In a chord containing one microtonally altered note, that note would play on a different channel from its companions. We changed patches frantically and maintained no fixed relationship between instruments and channels – everything existed in a kind of 'DNA soup' where resources were allocated dynamically as needed.

This approach let us extract far more than the nominal sixteen-channel limit from typical MIDI synthesisers. We maintained detailed specifications for every common synthesiser on the market, including how to balance dynamics and handle idiosyncratic behaviours. 

​The system's sophistication extended well beyond pure pitch calculations. When my opera The Maids was commissioned by the Royal Stockholm Opera, I spent considerable time crafting realistic rehearsal tapes. Everything I learned from that process was automated into Igor's playback engine.

We also collaborated with the KTH Royal Institute of Technology Musical Acoustics department, led by the legendary Johan Sundberg, whose research had quantified subtle but crucial performance characteristics. Those famous four milliseconds – the consistent temporal offset between soloists and accompaniment in professional orchestras – found their way into our algorithms. Such details proved particularly effective with Schönberg's Hauptstimme markings (𝆦) or similar solo indicators.

We also developed what my composer colleague Anders Hillborg and I privately called 'first performance prophylaxis' – a deliciously cruel setting that simulated the sound of musicians who hadn't practiced. In other words, the kind of sound landscape any composer is used to hearing at a first orchestral rehearsal of a new piece and which always makes you doubt your own talent. Turn this setting up, and you'd hear a characteristically dreadful youth orchestra. Turn it down completely, and you'd get the robotic precision that plagued every other MIDI system. Rather like Karl Richter's Baroque organ recordings.

The humanisation algorithms incorporated realistic instrumental limitations. Passages written too quickly for an instrument would skip notes convincingly. We modelled the typical rhythmic hierarchy of orchestral sections: percussion most precise, then brass, then woodwinds, with strings bringing up the rear. Instruments were panned to their proper orchestral seating positions. Piccolo trills were faster than tuba trills. The result was startlingly realistic, particularly by 1996 standards.

​Now, twenty-five years later, that laundry basket discovery has culminated in ADR 0026: Pitch Representation and Operations, documenting Ooloi's comprehensive pitch representation system. The original Common Lisp has been reborn as Clojure code, with string-based pitch notation ("C#4+25") serving as the canonical format and a factory-based transposition system supporting both chromatic and diatonic modes.

The string representation offers several advantages: compact memory usage for large orchestral scores, direct human readability for debugging, and seamless integration with parsing and caching systems. Most crucially, it supports arbitrary microtonal deviations, something that remains problematic in most contemporary notation software.

The factory pattern generates specialised transposition functions that encapsulate their musical behavior rules through closures. Rather than repeatedly passing configuration parameters, the factory creates efficient, composable functions that understand their specific musical contexts. A diatonic transposer preserves letter-name relationships; a chromatic transposer produces frequency-accurate results with canonical spellings.

The t-shirt in my laundry basket represented more than nostalgic memorabilia; it was unfinished business. That higher-order function embodied a sophisticated understanding of musical mathematics that took a long time to develop and seconds for an AI to recognise as architecturally significant.

Now, with Ooloi's pitch operations properly documented and implemented, that business approaches completion. The code has evolved from promotional garment to production system, carrying forward those insights from 25 years ago into a new, modern technological context.

It's exciting. And still a little unnerving.

Three tests left on the visibility work. After that, this part of the system is done.

It’s taken a few weeks to get the final bits of the gRPC server/client interaction implemented, but it now feels solid. The server reliably and efficiently handles all the complex aspects of live clients connecting, disconnecting, and communicating locally or over the inherently chaotic internet. 

With that done, the focus is now on what is called observability: being able to see what's going on inside. In other words, server statistics. We’re now tracking server-wide and per-client metrics: connections, API call rates, message sizes, event streaming, queue overflows. Endpoints can return JSON during development or Prometheus format in production. Content negotiation is automatic, or you can force a format with a query parameter. No new tools to learn. Grafana dashboards work straight away.

Why this matters for musicians: research has shown that in collaborative settings, users prefer a notation program that lets them work together over one with the most features. That says everything about how poorly existing software supports collaboration. It’s treated as an optional extra, if it exists at all.

Ooloi is built differently. Collaboration is not bolted on afterwards but designed into the system from the start. If event queues slow or connections drop, collaboration breaks down. A slur added on one screen never appears on another, or two edits collide and the score diverges. The statistics make these problems visible. They show when the system is stressed, when clients are falling behind, when the score risks drifting out of alignment. In short, they make collaboration trustworthy.

And unlike the current trade-offs, Ooloi isn’t choosing collaboration instead of notational depth. It is designed to be both: a serious, advanced notation platform that also makes shared editing reliable.

The same architecture also opens up another possibility. Because Ooloi is client/server, it can stream smoothly scrolling scores in real time. That makes digital music stands for ensembles or orchestras feasible. It’s not the main focus, but it’s there: the professional counterpart to the school-level collaboration studies.

Three tests left. Once they pass, the intricate server machinery will be visible. And visibility is the first condition for trust.

Ooloi is moving fast. Architectural decisions, implementation milestones, and the long road toward open-source release deserve a channel that isn’t buried in blog posts or scattered across updates.

The Ooloi Newsletter is that channel. It’s for composers frustrated with the lag of existing software, musicians curious about what modern architecture makes possible, educators and publishers planning future workflows, and developers who want to see Clojure’s STM, transducers, and gRPC applied directly to real musical problems.

This isn’t marketing. There’s nothing to hype, no metrics to chase. It’s progress – the daily work of turning musical thinking into software design. Updates arrive when there’s something real to report: once or twice weekly in active phases, slower during architectural deep work.

Expect honest accounts of what’s been built, what broke, and why decisions were made. Technical when it illuminates music, musical when it drives architecture. No jargon for its own sake, and no dumbing down.

If that sounds useful, subscribe at buttondown.com/ooloi.

The thread from FrankenScore to Ooloi continues. You’re welcome to follow it.

​Igor's demise taught me harsh lessons about market timing. When you create genuinely superior tools because existing software actively gets in your way, you discover that technical excellence alone can't guarantee survival.

Igor actually started as freeware – that was my original vision. The VCs wanted features, revenue streams, market capture. I wanted musicians to have tools that actually served their creativity. The collision between these fundamentally incompatible visions, plus the economic chaos after 9/11, killed what could have been transformative.

That experience shaped every decision in Ooloi. This time: transformation through collaboration rather than zero-sum market battles.

The MPL 2.0 licensing reflects this philosophical approach. The core backend becomes commoditised infrastructure; commercial value migrates to sophisticated interfaces, proprietary plugins, premium workflows. What traditional systems charge for – collaboration, large ensemble support, quality rendering – becomes architectural givens.

This isn't theoretical. Martin Keary (Tantacrul), who heads MuseScore development, had been interested in Igor Engraver as a young composer. When we discussed our open source projects, the conversation was refreshingly direct: 'Well, feel free to help yourself to whatever's useful when it goes open source​' – 'Likewise!' Both projects share the belief that notation software shouldn't cost students hundreds of pounds.

This is precisely what I want – for the code to be used, extended, transformed. The platform approach only works if others build on it. That's not threat; it's the point.

​The irony isn't lost on me that companies who once viewed architectural advances as existential threats might ultimately benefit from them. I still recall Sibelius management phoning after Igor's demise to enquire about my plans – ensuring no resurrection attempts, obviously.

Should those companies adopt approaches they once feared, the satisfaction would be collaborative rather than competitive. Success through enabling others, not defeating them.

​Of course, this depends on Ooloi actually working. Functional programming should provide significant performance benefits, but what happens with complex layout calculations across a hypercomplex 100-staff score by Brian Ferneyhough (whom I had as guest teacher at the Royal College of Music in Stockholm back in the day) whilst maintaining real-time collaboration?

The arrogance in predicting that these approaches will outperform decades of commercial development is considerable. My instincts could be spectacularly wrong – though they've been reliable enough to spot architectural dead ends before they become obvious to everyone else.

That's the architect's bargain: trust your reasoning, build on principles, prepare to be surprised.

"Next-generation" gets thrown around as a PR term so casually it's meaningless. If Ooloi genuinely represents next-generation architecture, that should be inherent in the platform design rather than claimed through marketing. The question isn't whether Ooloi calls itself next-gen, but whether others build on it in ways that would have been impossible previously.

Perhaps this time, it might be composed collaboratively.

Ooloi isn't here to "disrupt" anything. I despise that cheap neoliberal selfishness. It aims to transform, to evolve, to open possibilities. But there's no will to conquer — it's simply not in its DNA.

Deep in the implementation of the gRPC layer I fell into the same foxhole as so many gRPC developers seem to do: implementing streaming between gRPC server and client using in-process communication. Everything works swimmingly: high speed, nothing gets lost, advanced async patterns just work.

And then they try doing this using network transport. Suddenly HTTP/2 changes everything, and that equivalence between simple request-response and asynchronous streaming just explodes.
​
I spent a couple of days cursing the universe, but in the end I managed to crawl up into the daylight again with a solid solution. Since this seems to be a common snag, I decided to publish a little guide on the considerations involved, so others may have to do less cursing.

Here it is: GRPC_STREAMING_THREADING_GUIDE.md

​Nearing the end of the gRPC implementation phase now: coming blog posts should soon become more musical again.

​At some point, the question will be asked: “Isn’t this all a bit over-engineered?”

Multicore parallelism; Software Transactional Memory; gRPC; GPU acceleration; a plugin system designed as a first-class citizen rather than a bolted-on afterthought; an asynchronous server/client architecture with specialised streaming features. Prometheus monitoring. For music notation software, that can sound excessive.

But that assumption is exactly why notation software has been failing composers for decades. Not because it was too ambitious, but because it was chronically under-engineered.

Text editors are linear: O(n). Basically, what they handle is a string of characters broken up into lines. Music notation, on the other hand, is two-dimensional, contextual, and computationally explosive. Synchronising voices, aligning dozens of staves, resolving collisions, spacing measures, redrawing in real time: these are quadratic and cubic problems (O(n²), O(n³)), with NP-hard layout challenges in the general case.
​
That's why scrolling takes seconds. That's why orchestral scores become unusable. And that's why the industry has spent thirty years patching symptoms instead of tackling the cause.

​Look at the record:

• Sibelius: selecting a single note in an orchestral score can take several seconds.
• Finale: collapsed under its own weight, with delays of 5–90 seconds for basic actions.
• MuseScore: freezes completely on Strauss’s Elektra. (They all do.)
• Dorico: more modern, but still lags 15–40 seconds on large scores.

And here is the deeper problem: users have learned to accept this. They zoom in to a handful of staves, scroll in slow motion, restart their program every quarter of an hour. They've accepted that the fundamentals can't be solved. A whole profession has normalised working around performance breakdowns as if they were laws of nature.

They're not inevitable. They're the result of decades of under-engineering.

​The remedies weren't always available. In the 1980s SCORE capped out at 32 staves because 640 KB of memory left no room for orchestral complexity. Through the 1990s and 2000s, Finale and Sibelius (and Igor Engraver!) wrestled with single-threaded designs on single-core CPUs. Even into the 2010s, GPU rendering pipelines were immature, and most concurrency models in mainstream languages couldn't be trusted in production.

Only recently have the necessary ingredients converged:

• Affordable multicore hardware on every laptop, making parallel measure formatting possible.
• GPU-accelerated rendering (Skia) for fluid scrolling and zooming in real time.
• Mature concurrency models such as Clojure’s Software Transactional Memory, providing safe lock-free collaboration.
• Immutable data structures that give transactional clarity to complex notation states.
• JVM interoperability that allows plugin developers to work in their own languages.

This is why Ooloi is written in Clojure. Not because of language fashion, but because Clojure can orchestrate this synergy.

​Ooloi is designed to solve these problems at the root:

• Parallel layout: every core formats measures simultaneously.
• STM transactions: true collaborative editing without locks, with automatic retries on conflict.
• GPU Skia rendering: zooming and scrolling at video-game speed.
• Plugin-first design: developers work with a clean musical API, not concurrency primitives or network plumbing.

To musicians, Ooloi looks like a normal application. To plugin developers, it feels like writing musical logic in their favourite JVM language. The hard problems are solved once in the core, so nobody else has to live with them.

​So no, Ooloi isn't over-engineered. It's appropriately engineered for a domain that has been persistently underestimated. The remedies only became possible recently, when the technology finally caught up.

I simply happen to live at the intersection of deep architectural knowledge and deep musical knowledge, with the scars (also deep) of having done this before. Ooloi isn't the product of singular genius: it's the moment when the right tools finally aligned with the right problem.

The proof won't be in a benchmark or an ADR alone. It'll be when musicians can finally edit, scroll, and collaborate on large-scale scores without breaking their creative flow.

​Ooloi will be open source by design. The complexity is in the foundations so that musicians, teachers, students, and developers don't have to deal with it. Plugin writers don't need to care about concurrency or transactions: they work with measures, staves, and voices in a musical API. Most contributors will never touch the Clojure core, and they won't need to.

This is a gift to the community: an infrastructure platform built to be extended. The aim is simple: to finally make notation software scale to the real demands of music, and to give others the foundation to build what I alone never could.

​Since LLMs are good at summarising, here’s what Claude Sonnet came up with when I asked it to describe my process for developing Ooloi. The phrase “the Bengtson method” is irritating and misleading; plenty of people have reached similar conclusions. Still, this may be the only technical write-up of the approach that includes the word 'arse-licking'.

So here it is: Claude’s summary, em dashes, bullet points, and all. It rambles a bit, but I’d rather give you the authentic output than a tidied-up version. Same principle as always: authenticity beats decorum.

... but before that, I think it might be good to include my reply from LinkedIn to an accomplished architect friend of mine who (jokingly referring to me as 'the illustrious Peter Bengtson') initially didn't quite buy that harsh negativity really is motivated:

​With that clarification in place, now on to what Claude wrote:

Core Methodology: Consultational TDD

The foundation rests on a rigid Test-Driven Development cycle with mandatory consultation checkpoints:

1. AI presents test proposal with complete sequence analysis
2. Human approves or rejects before any code is written  
3. AI writes minimal failing test
4. AI implements minimal code to pass test only
5. Human reviews and corrects any architectural violations
6. Local test execution on changed files
7. Full test suite execution to catch regressions
8. Commit when all tests pass

​

Peter Bengtson's Claude Code development process represents a disciplined, constraint-based approach to AI-assisted software development that has demonstrated success in complex functional programming domains. The methodology's core insight—that harsh constraints improve rather than limit AI effectiveness—contradicts conventional wisdom about collaborative AI development.

The harsh correction mechanisms and authoritarian control structure may be necessary rather than optional components, suggesting that successful AI collaboration requires active management rather than partnership. This challenges prevailing assumptions about human-AI collaboration patterns but provides a tested alternative for developers willing to maintain strict disciplinary control.

The technical achievements demonstrate that properly constrained AI can assist with genuinely sophisticated software engineering tasks, not merely routine coding. Whether this approach scales beyond its current constraints remains an open question requiring further experimentation and validation.

The Ooloi server must of course also have comprehensive introspective statistics, as it can be deployed in several ways and has a structure which is perhaps a bit unusual. The combination of STM-wrapped gRPC API calls with server-to-client event streaming creates some operational challenges: things like per-client queue overflow behaviour and collaborative editing session patterns.

Hence this new ADR on server statistics, which takes a two-level approach: server-wide aggregates that survive client churn, plus detailed per-client metrics for operational visibility. Oh, and it interfaces with Grafana, Prometheus, and friends straight out of the box via simple HTTP endpoints.

Halfway through the stats implementation now. It's been a productive weekend.

I've just published ADR-0024 and, based on it, a new gRPC Communication Guide documenting when flow control helps and when it interferes.

The decision was straightforward once the analysis was done: implement flow control for event streaming (where slow clients can block fast ones) but avoid it entirely for API requests (where gRPC's threading model and STM transactions already coordinate beautifully).

Sometimes the most sophisticated solution is knowing where not to add sophistication.

The guide covers the technical reasoning and includes practical examples for anyone implementing gRPC clients, regardless of language choice.

Every community has its breaking point. Mine came on Clojurians when I wrote a single sentence:

'Clj-kondo can go away – I have 18,000 tests'.

That was enough to get my post deleted. Before the deletion, there was 'discussion' – if you can call it that. I was told my statement was nothing more than click bait.

The irony? The author of clj-kondo himself agreed with me.

What That Line Meant

It wasn't click bait. It was a statement of principle:

• Tests prove correctness. They're executable, falsifiable, and domain-driven.
• Linters and static analysis don't. They enforce style, not truth.
• Dynamic dispatch makes static analysis less meaningful. Ooloi's architecture relies heavily on Methodical multimethods and polymorphism throughout. Static analysis tools can't trace through runtime polymorphic calls, making their warnings less informative than executable tests that actually exercise these dynamic paths.
• When you've got 18,000 tests running clean, you don't need a priesthood of external validators telling you your code is 'unsafe'.

And I was careful to make the distinction explicit: clj-kondo is a beloved, useful tool. For most projects it adds value. It just happens to be of limited use in my project, because Ooloi's architecture is already validated at a different scale.

That nuance – acknowledging the tool's value whilst drawing boundaries around its relevance – should have been the beginning of a sober technical discussion. Instead, it was treated as provocation. The fairness itself was read as heresy.

The Culture Clash

The moderator (a 'Veteran Architect') didn't engage with the point. He reacted from the gut: pearl-clutching, dismissing, and finally deleting. Exactly the kind of gatekeeping I dissected in my article on functional programming gatekeeping.

And let me be clear: I have nothing against the Clojurians themselves. They're a knowledgeable, interested lot, often deeply engaged in technical detail. The problem isn't the community – it's the moderation culture.

The moderators behave more like a church council than facilitators of discussion. Their first instinct isn't to sharpen an argument, but to protect orthodoxy, maintain decorum, and suppress anything unsettling.

The ideal they enforce seems to be some kind of cold, robotic detachment – the lab-coat fantasy of neutrality – or perhaps the modern American obsession with never offending anyone, no matter how bloodless the discourse becomes. Either way, it enforces sterility, not clarity.

You can critique syntax sugar all day long, but question a community darling like clj-kondo – even whilst calling it useful and respected – and suddenly you're accused of trolling.

Why I Left

I didn't leave because I was offended. I left because I refuse to participate in a space allergic to honesty. If a community sees a blunt critique and immediately cries click bait – ignoring both the nuance of my post and the fact that the tool's own author agreed – it has no business in my world.

Ooloi is built on clarity, not ceremony. It's an architecture tested by 18,000 executable truths, not validated by a linter's opinion. If that treads on toes, good. Prissy people afraid of dark humour or communication nuances that wouldn't pass muster at a parish council don't belong in this project. And the same thing goes for hypocrites who say, 'We're inclusive here - as long as you're exactly like us'.

The Broader Lesson

Communities often confuse politeness for health. But real progress requires the courage to tolerate discomfort. If you need your software conversations padded with pillows, you'll never survive the weight of real architecture.

As Wednesday Addams would remind us: hypocrisy is uglier than bluntness, and dishonesty is far more offensive than a glass of gin before noon. Or, indeed, a well-placed 'fuck you'.

So I deleted my Clojurians account. Because sometimes subtraction is progress.

There's a moment in every software project when you realise you've been approaching a problem entirely backwards. For Ooloi, that moment came whilst implementing the frontend gRPC client. What I'd anticipated would be a tedious exercise in data transformation and type marshalling turned out to be something rather more straightforward: we could simply share the models themselves.

Most applications suffer from what I've come to think of as 'linguistic impedance mismatch': the same business concept gets expressed differently in TypeScript interfaces, JSON schemas, database models, and API contracts. Each translation introduces potential for drift, bugs, and the sort of maintenance headaches that make senior developers reach for the gin before lunch.

The Usual Compromises

When I began implementing Ooloi's frontend, I expected to follow the well-trodden path of recreating backend data models for the client, probably with a good deal of manual conversion between Clojure's rich data types and whatever could survive the journey through gRPC.

A Simpler Path Forward

But then something rather straightforward happened. Our unified gRPC architecture, built around a custom OoloiValue message format, was preserving not just the data but the semantic fidelity of Clojure structures. Ratios remained ratios. Keywords stayed keywords. Nested collections maintained their exact shape and type information.

The implications were rather obvious once I thought about it: if the data was surviving the round trip with perfect fidelity, the code could make the same journey. The broader lesson here applies beyond Clojure: when your serialisation layer preserves semantic fidelity, you can often eliminate entire categories of translation logic.

Shared Models in Practice

What we ended up with is shared model contracts across distributed systems. Not just shared schemas or interface definitions, but shared implementation: the same defrecord structures, the same predicates, the same multimethod dispatch logic working identically in frontend and backend.

For example, here's client code that uses the exact same model logic as the server:

​This isn't just syntactic sugar. The frontend literally cannot represent a state that the backend would reject, because they're using identical validation logic. Entire categories of bugs, the sort that usually emerge only in production when client and server expectations diverge, simply cannot exist.

For an open source project like Ooloi, this architectural decision has profound implications for contributor experience. New developers don't need to learn separate model definitions for frontend and backend. The cognitive load of understanding the system drops considerably when there's only one way to represent musical structures, regardless of which part of the codebase you're working in.

Architecture in Practice

What started as a practical decision to move some data models has led to a clearer architectural arrangement:

• The Shared Project contains the entire Ooloi engine: all domain models, interfaces, predicates, traits, and core business logic. This is where musical knowledge lives.
• The Backend Project is essentially a server wrapper: a thin layer that exposes the shared engine through gRPC, handles persistence, and manages component lifecycle.
• The Frontend Project is a UI wrapper: JavaFX components, user interaction handling, visual rendering.

Both frontend and backend have become lightweight adapters around a shared core, rather than independent systems that happen to communicate.

For those interested in the technical details, the complete architectural decision record is available in our ADR-0023: Shared Model Contracts.

Why This Approach Is Uncommon

Most teams face barriers that make shared models impractical: different programming languages between frontend and backend, runtime environment constraints, the natural tendency for teams to optimise for their specific context rather than maintaining shared abstractions.

We've managed to sidestep these issues through a combination of technological choices (Clojure everywhere, gRPC with custom serialisation) and architectural discipline (resisting the urge to optimise locally at the expense of global coherence). For open source projects, this consistency becomes particularly valuable: contributors can focus on domain logic rather than navigating translation layers between different parts of the system.

What This Means for Multi-Language Support

Importantly, this shared model architecture doesn't create barriers for non-Clojure clients. Python, JavaScript, or WebAssembly clients continue to work through the standard gRPC interface, using generated protobuf classes and standard API patterns. The shared models represent a Clojure-specific enhancement layer that sits atop the universal gRPC interface rather than replacing it.

Think of it as offering two levels of integration: the universal protobuf API that any language can consume, and the native Clojure API that provides richer semantics for those who can take advantage of it.

Alternative Frontend Approaches

This architecture actually makes it easier for others to build alternative frontends. Someone wanting to create a React-based web interface or a WebAssembly client has a clearly defined gRPC API to work against, with well-documented behaviour established through our shared contracts. They'd handle their own data model representations (the normal situation for any gRPC client) whilst benefiting from a well-defined backend.

We're not digging a moat here. Alternative approaches remain viable whilst the shared contracts make the Clojure experience particularly seamless.

The Broader Picture

There's something here that extends beyond the specific technical details of Ooloi. We've found that perfect type fidelity across network boundaries, combined with clear thinking about what constitutes core business logic versus infrastructure concerns, can enable patterns that many teams dismiss as impractical.

This doesn't mean every project should adopt this approach. The organisational and technical discipline required is considerable. But for projects where the complexity is justified (particularly open source projects where reducing cognitive load for contributors is crucial) the benefits are substantial.

Looking Forward

Going forward developing Ooloi's frontend, the shared model contracts have become foundational to how we think about the system. Features that might have required careful coordination between teams now flow naturally from shared understanding. The system has become more coherent and, importantly for an open source project, more approachable for new contributors.

The surprise wasn't that shared models worked; it was how much friction simply disappeared once we stopped duplicating concepts. Sometimes architectural progress comes not through invention, but through subtraction. Shared model contracts weren't a goal we set out to achieve. They emerged from following our technical choices to their logical conclusion and having the discipline not to complicate what worked.

I'll admit, when I first encountered gRPC batch operations, I dismissed them as unnecessary complexity. Then I started implementing the gRPC layer and realised something remarkable: gRPC batch boundaries map perfectly onto STM transaction boundaries.

The pattern is almost embarrassingly simple: client streams a series of operations, server accumulates them, then wraps the entire batch in a single dosync. What emerges is something genuinely powerful – distributed transactions with full ACID guarantees. Multiple musicians can edit the same score simultaneously, knowing that either all their changes succeed atomically or none do. Complex operations like MusicXML imports or multi-step undo chains become naturally transactional across network boundaries.

The implications are profound: no more partial updates corrupting shared musical data, no locks preventing collaboration, no eventual consistency headaches. Just proper transactional integrity that works identically whether you're editing locally or collaborating across continents.

How many music notation programs have distributed atomic transactions across any network? And entirely without locks, and with automatic conflict resolution? None.

​Sometimes the most elegant solutions hide in features you initially think you don't need.

Right. Quick update from the development trenches.

When I completed Ooloi's backend engine in July, starting work on the frontend interface revealed the anticipated cascade of architectural requirements that needed systematic resolution first.

Here's what emerged, in order:

1. Collaborative Undo/Redo Architecture (ADR-0015)
Thinking about frontend-backend relationships immediately raised the question: how does undo/redo work in a multi-client, distributed collaborative setup? The answer required a three-tier architecture separating backend piece changes (coordinated via STM) from frontend UI changes (local to each client).

2. Universal Settings Architecture (ADR-0016)
The insight that there should be no global application settings, only per-piece settings living inside each piece, led naturally to implementing settings not just on the piece level, but across all levels of the hierarchy. Any entity – piece, musician, staff, pitch – can now have configuration attributes via a unified defsetting macro with lazy storage and automatic VPD support.

3. Component Lifecycle Management (ADR-0017)
Multi-scenario deployment demanded rock-solid system architecture using Integrant. This needed to be in a stable architectural form – wiring, lifecycle boundaries, failure modes – before setting up the actual components with proper dependency injection, partial failure handling, structured error codes, the full production suite.

4. Automated gRPC Generation (ADR-0018)
With component architecture sorted, I could tackle the actual gRPC implementation: automating API endpoint generation for native Java interop across hundreds of methods, plus bidirectional communication for real-time collaboration. Manual implementation at this scale would be architecturally impossible.

5. In-Process Transport Optimisation (ADR-0019)
Combined deployments (frontend and backend in same process) were using unnecessary network transport. Implementing automatic in-process gRPC transport delivers 98.7–99.3% latency reduction whilst preserving external monitoring capabilities.

6. TLS Infrastructure (ADR-0020)
Secure connections are essential for distributed deployments – conservatory intranets, corporate environments, cloud SaaS situations. Auto-generating certificates with full enterprise capabilities makes this transparent whilst supporting everything from development to production.

7. Authentication Architecture (ADR-0021)
Finally, distributed deployments require comprehensive authentication and authorisation. Pluggable JWT-based providers scale from anonymous sessions to enterprise LDAP integration. This is fully designed and will be implemented as deployment scenarios require.

Current Status: About 95% of the above is implemented, tested, and production-ready.

Next Steps: Finish the auto-generated gRPC Java interop interface, then create an actual frontend client of the 'Hello World' variety and ensure it runs and communicates across all deployment scenarios.

The rather encouraging discovery throughout this process was how readily the existing functional architecture accommodated these enterprise concerns. Vector Path Descriptors naturally supported universal settings. STM transactions elegantly handled collaborative undo operations. The component system absorbed authentication providers without strain. When features like collaboration or security slide cleanly into place, it's not luck – it means the architecture wanted them there. That's what sound foundations do.

Worth noting: collaboration isn't something tacked on later. It's integral to the architecture from the ground up.
​
Right. Back to the gRPC generator.

After a year building the backend of Ooloi with Claude, I’ve learned this:

Successful AI collaboration isn’t about creative freedom. It’s about harsh constraint.

AI will overstep. Your job is to correct it—immediately, uncompromisingly.
The friction isn’t failure. It’s the method.

Read the full piece – which I asked the AI to write in its own voice – here.

​I've been saying I'm itching to open source Ooloi, but waiting for the complete software felt increasingly artificial. The documentation tells the real story – twenty-five years of architectural evolution from Igor Engraver through AWS systems thinking to this closure in Clojure.

So here it is: the complete Ooloi documentation collection, released publicly ahead of the software itself.
What you'll find:

• Comprehensive guides teaching advanced Clojure through musical problems
• Complete architectural decision records showing our thinking
• The timewalker guide (with three Rocky Horror references, naturally)
• Real functional programming education through meaningful domain examples

This represents my attempt to walk the talk after critiquing FP drawbridge syndrome. These guides serve dual purpose: they document sophisticated music notation architecture whilst teaching functional programming concepts through examples that actually matter.

The timewalker guide alone demonstrates transducers, lazy sequences, and functional composition through orchestral score traversal. The polymorphic API guide teaches multimethods through musical type systems. The concurrency patterns show STM coordination through collaborative editing.

Twenty-five years of thinking, distilled into something that I hope proves useful beyond music software. The architecture is complete; the foundations are solid. Now you can see why I've been eager to share this work.

An organism evolved, indeed.
​
Explore the complete documentation: https://github.com/PeterBengtson/Ooloi-docs

​/Peter

Now that the backend engine is complete, architecturally speaking, I'm getting increasingly eager to open the source. Itching, in fact. But as that will take a while, the least I can do for you is give you the WELCOME.md file from the source repo. 

​/ Peter