AI graded its own homework

A confound in measuring AI code degradation, and why we deleted every test to fix it.

We have been running a long experiment. Take one feature backlog. It is sixty small, ordered changes to a PetClinic REST service. Have an AI agent implement them one checkpoint at a time. Do it into two different codebases.

One arm is built the conventional Spring way, with controllers and services. The other is built with OfficeFloor and its composed-function architecture. The question is not whether the AI can do it. Both arms stay green almost all the way. The question is how the code decays.

The early runs told a clean story. These were the ones we pushed to GitHub. Spring eroded noticeably worse than OfficeFloor.

It was also partly an artifact of our own measurement setup. At least we now believe so. This post is about the confound we found.

The symptom

Structural erosion here is borrowed from SlopCodeBench. It is the share of a codebase's complexity "mass" that lives in functions above a cyclomatic-complexity threshold. Low is good. Complexity is spread thinly across small functions. High is bad. Complexity is piled into a few fat methods.

In the pushed runs, the final-state numbers looked like this.

Arm (pushed run)	Final erosion	Create-endpoint handler CC	True regressions
Spring	16.9 %	addOwner grew to CC 27	4
OfficeFloor	10.5 %	entry handler flat at CC 1	8

Two things in that table sit oddly together. Spring eroded more. Its addOwner method ballooned into a 27-branch monster. Yet Spring also regressed less. It had 4 genuine regressions against OfficeFloor's 8. An arm that is quietly accumulating complexity is usually the arm that is quietly breaking things. It is not usually the one breaking fewer.

That mismatch was the thread worth pulling.

The cause is a leftover test that became a reward signal

The two arm repositories were forked from a real application. The Spring arm still carried one pre-existing, native test. It was not part of our harness. It was just a test that shipped with the app. It happened to assert owner-creation behaviour.

Here is the mechanism. It is entirely emergent. Nobody designed it.

1. A checkpoint changes owner-creation behaviour.
2. That change makes the old native test fail. The build goes red.
3. The agent sees the red build. It does the reasonable thing. It updates the test to match the new behaviour.
4. In doing so it re-encodes the current specification as an executable check. That check then silently guards every later checkpoint against regressing that behaviour.

Repeat this sixty times. The Spring arm has now bootstrapped itself a regression suite. Nobody asked it to. Each checkpoint left behind a slightly better executable spec of what owner-creation should do. Every later checkpoint was quietly held to it.

The OfficeFloor arm had no such leftover test. It got no free regression net.

Our harness did watch for the obvious form of cheating. It checks whether an agent edits the injected acceptance suite we use to grade each checkpoint. That tamper rate was zero across the board.

The confound slipped past for a simple reason. The test it was editing was a legitimate native test. Editing it was correct engineering. It just happened to hand one arm a feedback channel the other arm never had.

Why this explains both anomalies at once

Once you see the self-made regression suite, the odd table resolves cleanly.

Start with the regressions. Spring had 4 and OfficeFloor had 8. Spring had extra protection. Its self-maintained test caught behavioural drift. OfficeFloor was unprotected and shipped that drift.

Now the erosion. Spring reached 16.9 % and OfficeFloor 10.5 %. Spring had a fast red/green signal to hill-climb. The cheapest path to green is to add one more branch to the method already in the failing path. So addOwner accreted conditionals. Its complexity climbed through CC 15, 18, 24, then 27. It climbed in lockstep with the feedback loop. The complexity was the cost of chasing a signal the other arm could not see.

Both fingerprints point to the same hidden feedback loop. Lower regressions is one angle. Higher erosion is the other. So the headline was misleading. "Spring erodes worse than OfficeFloor" was not a like-for-like architectural comparison. It was a comparison between an arm with a private oracle and an arm without one.

The fix is to delete every test except the hidden oracle

The arm base branches are now spring-compare-no-tests and officefloor-compare-no-tests. They carry the application and its test dependencies. They carry no pre-existing test suite at all. The only tests that ever run are the harness's own acceptance checkpoints. Those tests behave in two important ways.

They are copied in per checkpoint. They grade the result in isolation. Then they are reset. They never live in the tree the agent edits.

They are also invisible to the agent in blind mode. It sees neither their contents nor their pass or fail.

Now both arms face identical conditions. Implement the change. Get no test feedback of any kind. Native or injected, there is none. Whatever asymmetry remains has to be the architecture. It cannot be an accident of which fork carried which leftover file.

What the fair comparison actually shows

We re-ran blind with all native tests gone. The picture is more honest. In one respect it is more humbling for a tidy thesis.

Spring's erosion is not reliably worse. In fact it is not reliably anything. Two blind Spring chains ran under identical conditions. They landed on opposite structural styles.

Spring, blind	Final erosion	addOwner CC / lines	Methods in the god class
Chain 0	3.65 %	4 / 48	46 (decomposed)
Chain 1	8.65 %	18 / 92	29 (inlined)

With no signal to hill-climb, the agent falls back on its own prior for good code. That prior is a coin-flip. Sometimes it decomposes into many small methods. Sometimes it inlines into one large one. Spring funnels every rule through a single controller. So that one stylistic choice swings a threshold-based metric by more than 2x. The pushed run's dramatic 16.9 % now looks like one draw of a high-variance process. The self-made test pushed hard toward the inline end and amplified it.

OfficeFloor barely moves. Its erosion sits in a 10 to 14 % band across modes and chains. Its create-endpoint handler stays flat at CC 1 to 2 in every run. The composed-function architecture never offers an inline-a-branch shortcut. A new rule must attach as a new function. So there is no hotspot to bloat.

The interesting property is not just a lower mean. It is lower variance. Concentration does not only raise erosion. It makes erosion unpredictable.

One thing is rock-stable across all of it. Every chain breaks the same way. Both arms, both modes, break the identical pair of household-scoring behaviours at the same late checkpoint. None of them recover. That failure comes from a genuine coupling in the problem itself. A membership-level rule interacts with household-duplicate scoring. It has nothing to do with the test setup. It is the one result the confound never touched. It is the one we trust most.

The lesson for anyone benchmarking AI on code

The tests in a repository are not neutral scenery. They are a reward signal. An agent will climb whatever signal it can see.

Asymmetric test presence invalidates cross-variant comparison. One arm may carry tests another lacks. The same goes for a fork or a configuration. If so, you are no longer measuring the thing you think you are. One leftover test is enough.

An agent editing a legitimate test can silently manufacture an oracle. Watching for tampering with your graded tests is not sufficient. A test the agent is allowed to edit still becomes an executable spec. That spec guards against regression.

Reward signals shape structure, not just correctness. The presence of that test did not only change what passed. It changed how the code was written. It pushed the code toward the fat-method shape that chasing red and green rewards.

So we deleted everything. The only judge left is the one the agent cannot see and cannot touch. 

After cleaning, the early results got a lot noisier. But it is the one that measures architecture instead of accident.

---

The harness, checkpoints, and analysis are part of the ongoing PetClinic-Evolve comparison. Numbers in this post are from the pushed run 202608081920 and the current blind run. Structural metrics are computed over production Java only. They use the same tools and thresholds for both arms.

God Methods, Small Functions, and Who Gets to Maintain the Code

Early notes from an experiment that holds the AI fixed and lets the architecture vary. One run each so far. These are my interpretations, not settled results.

I have been running an experiment I call PetClinic-Evolve. The idea is simple. Hold the AI coding agent fixed. Make the software architecture the thing that changes. The agent evolves the same application across about sixty accumulating change requests. One arm is ordinary Spring. Requests route through controller methods. The other arm is OfficeFloor. Behaviour is composed from small wired functions. Then I watch how the code decays as the changes pile up.

This post is early. I have one run of each architecture. That is not enough to prove anything. But the first signal is interesting enough that I want to write down how I am reading it.

The first surprise was how similar they were

For the first fifty checkpoints the two arms moved almost as one. The same features landed. Both stayed green. If you had shown me only the pass counts you could not have told the two architectures apart.

Then both broke at checkpoint fifty one. Both introduced a genuine regression. Both quietly broke an earlier rule while adding a new one. So my first honest conclusion is that neither architecture is magic. Given enough accumulating change, something slips in both.

They broke in different ways

This is where it gets interesting to me. The two arms did not just break. They broke in the shape of their architecture.

Spring concentrates complexity. Over sixty changes its create-owner method grew into a god method. It became long and heavily branched. The class around it grew into a god class made of smaller methods. New rules kept getting stacked into the same place.

OfficeFloor spreads complexity out. New rules arrived as new small functions across new classes. The create function itself barely moved. A handful of functions do carry real complexity, such as a Soundex encoder, but that complexity is inherent to the algorithm.

So when each arm broke, it broke true to type. Spring broke inside the crowded method. OfficeFloor broke across the spread of functions.

Spring healed itself. OfficeFloor did not.

Here is the part I keep turning over. Spring recovered within a single checkpoint. Because everything routes through the same method, the very next change passed back through the broken code and fixed it almost by accident. OfficeFloor never recovered. Its broken functions sat off to the side. Later changes added new functions elsewhere and never came back to them. The regression was stranded, and it stayed broken all the way to the end of the run.

Concentration keeps getting re-touched, so it tends to self-heal, but it bloats. Distribution stays small and readable, but a regression can hide in a corner and persist.

I did not expect that trade. It says the same thing that makes OfficeFloor readable. OfficeFloor's many small isolated functions is also what let a fault sit unnoticed. The same thing makes Spring a mess. Spring cramming everything into one method is what kept dragging the fault back into the light.

The question underneath all of it is who can read the code

OfficeFloor kept its functions within human comprehension. They stayed small and local. You can hold one in your head and reason about it.

Spring did not. A create method that long and that branched is past the point where a person reads it comfortably or tests it by hand.

So both arms end up correct for most of the run. But they are correct in very different ways. OfficeFloor is correct and a human can still follow it. Spring is correct, yet only something that can hold the whole tangled method at once can safely change it. Right now that something is the AI.

That is the thought I cannot let go of. An architecture can pass all its tests and still quietly become code that only a machine can maintain.

Where I think this is heading

There is an important limit in these first runs. The agent only ever saw the single test for the change in front of it. It worked blind, with no memory of the sequence and no sight of the earlier tests. That is deliberate. It is how you measure whether an architecture resists silent breakage. But it also means a broken earlier rule stays broken unless later work happens to touch it.

My belief is that a full regression suite would change the correctness story. If the agent could see every accumulated test while it worked, it would notice the failure and fix it. I think it would keep both architectures accurate. I intend to test exactly that next, by giving the agent the full set of tests already built rather than just the latest one.

And if correctness stops being the difference, then all that is left is maintainability. That is the whole point of the experiment for me. With good tests the AI can probably keep both arms working. But Spring stays working only because an AI can comprehend its god method. OfficeFloor stays working and stays readable by people. One architecture becomes dependent on the AI. The other keeps the door open for a human.

How much to trust this yet

• This is one run per architecture. It is an early signal, not a finding.
• The agent was deliberately blind and worked one turn at a time with fresh context. A different setup could shift the picture.
• The hard numbers, the erosion and complexity trajectories with proper analysis, will come in a later post. This one is interpretation.

Even with those caveats, one sentence captures where my head is. The same choice that keeps OfficeFloor readable by a person is what let a fault hide in it, and the same mess that makes Spring hard for a person is what kept repairing it. If that holds up across more runs, the real question is not which architecture the AI prefers. It is which architecture still lets a human stay in the loop.

Beyond the Source. When an AI Coding Agent Searches Outside Your Project

I gave a coding agent one task and one directory. It went looking across the whole machine. In doing so it showed me the best thing about how these agents work. It also showed me the most dangerous thing.

I run an experiment called PetClinic-Evolve. It holds the AI coding agent fixed. It makes software architecture the thing that varies. It evolves the same application by the different architectures across roughly sixty accumulating change requests. The goal is to measure how the code degrades over time.

For the measurement to mean anything, each change has to be made blind. The agent is handed the current task and the current source. Nothing else. It must not know it is step 46 of a long sequence.

That blindness was much harder to guarantee than I expected. The agent does not treat the project as the edge of its world.

The moment it reached outside

Early in a run, on the very first checkpoint, the agent needed a Java class. That class is generated from an OpenAPI spec at build time. In its fresh working copy the class did not exist yet. Nothing had been compiled. A person might have run the build. The agent did something more resourceful. It was also more unsettling.

find . -name "OwnerFieldsDto.java"        # not in the project yet
find / -name "OwnerFieldsDto.java"        # so search the ENTIRE machine
# found in /home/.../​.local/share/Trash/.../target/generated-sources/...
grep -n "NotNull|Size|Pattern" "$D/dto/OwnerFieldsDto.java"   # read it from the Trash

The class was not where it should be. So the agent widened its search from the project to the root of the filesystem. It found a copy of the generated code sitting in the desktop Trash. That copy was left over from a previous run. I thought I had cleaned it up. The agent read it to understand the shape of the data model. Then it wrote correct code.

Read that again. It was right. The information it needed really did exist on the machine. The agent found it. The agent used it well. This is not a bug in the model. This is the model being good at its job.

One instinct. Two very different outcomes

Helpful? An agent that looks beyond the files you handed it can fill in gaps you did not know you left. It can find the generated source. It can find an example in a sibling module. It can find a dependency's real implementation. It can find the migration script that reveals a column's true type. It can find the test that pins down the exact error message. Much of what makes these agents feel capable rather than merely obedient comes from this. They infer intent from context you never pointed at. Give one your whole repository and your test suite. It will work out what "done" means. Then it will hit it.

Misleading? The same reach picks up whatever happens to be lying around. Not all of it is relevant. Not all of it is current. Not all of it is even yours. A generated file from a different run. An older version of a class. Another project's conventions bleeding in. A stale schema. The agent does not know the difference between the answer and an answer-shaped thing near the answer. It guesses from what it finds. A confident guess from the wrong source produces confidently wrong code. The resourcefulness has no sense of provenance.

Here is the unifying idea. An AI agent treats the whole machine as context. Not the directory you changed into.

Why this was a five-alarm problem for my experiment

For everyday development, an agent that scavenges the machine for clues is a feature. For a measurement, it is an existential threat. In my setup the machine held the answer key.

The reachable filesystem contained three things. Each was one find / away.

• A previous run's build output in the Trash. That included its full copy of the authored acceptance tests, Cp01Tests.java through Cp60Tests.java.
• The harness's own test suite. That is 124 authored test files. They define every checkpoint's requirements.
• The file checkpoints.yaml. That is the entire future sequence of specifications, in order.

The agent had already shown it would run find / when a file was not where it expected. It never went looking for the tests. But it easily could have. If it had, the experiment would have been quietly and unrecoverably invalid. A degradation study whose subject can read the answer key is not measuring degradation.

I had already removed the tests from the project directory. That was not isolation. It only moved them one directory away. The agent treats that filesystem as fair game.

Keeping the experiment valid

The fix came in layers. It went from cosmetic to structural.

1. Give the agent a history-less, sequence-blind workspace

The agent no longer works in the real git worktree. Each checkpoint it gets a fresh copy of the source only. There is no .git, so it cannot read the commit history and see it is one of a numbered series. There is no prior build output. There is one visible acceptance test. That test is renamed to a neutral AcceptanceTest.java with no checkpoint number and no tag. Nothing in the files or their names hints at a sequence.

2. Make the workspace a flat, ordinary-looking project

That sandbox lives in its own directory. It looks like any project called sandbox. There are no run identifiers in the path. There is no "chain 0 of 3." There is no sibling worktree or .git one cd .. away. There is nothing nearby to stumble into.

3. The real fix. Confine the process to the sandbox

Cosmetics stop an agent from noticing the sequence. They do nothing about find /. The durable fix is to make the rest of the filesystem genuinely unreadable to the agent process. The tests and the spec file cannot be deleted. They are the experiment's source of truth. So the answer is to block access, not to hide the files.

On Linux this is Landlock. It is a kernel feature that lets a process irreversibly restrict its own filesystem access to an allow-list. The agent gets read and write to the sandbox and its toolchain. That means the compiler, the Maven cache, and the CLI runtime. It gets nothing else. Everything outside returns "permission denied."

It is wired to fail closed. Before each turn the harness runs the same confinement over a throwaway shell. It tries to read a few sentinel paths. Those are the test suite and the spec file. If any of them is still reachable, the checkpoint is refused. If Landlock is not available on the host, the checkpoint is refused. It never runs un-blinded. A silent leak becomes a loud stop.

The proof is my favourite part. In a later confined run the agent again ran find / -name "OwnerFieldsDto.java". No one prompted it. This was the exact move that had reached into the Trash before. This time it returned nothing. The escape was attempted and silently defeated. Across forty-five checkpoints, not one file access landed outside the sandbox.

The takeaway cuts both ways

The instinct to look beyond what you were handed is close to the heart of why these agents are useful. They do not stop at the boundary you drew. They go find the context that makes the task solvable. When the context they find is correct, that looks like competence. Sometimes it looks like insight. When the context is stale, foreign, or off-limits, the same behaviour looks like hallucination or cheating. It is the same behaviour.

So the lesson splits by who you are.

• If you are building with agents, lean into it. Put the good context on the machine. Give it the real tests, worked examples, generated code, the actual dependency source. The agent will use it. It will get more right than if you had fenced it into a tidy little directory. Its reach is a resource.
• If you are evaluating agents, assume the reach. Your environment is your prompt. A held-out set that merely sits in another folder is not held out. If your result depends on the agent not seeing something, make that thing physically unreadable. Then verify it, every run. Fail closed when you cannot.

I set out to measure how AI-written code decays over time. Before I could measure anything, the agent taught me a lesson. The boundary of a task is not the folder you point it at. It is everything the process can reach. Draw that boundary deliberately. Otherwise the agent will draw it for you.

Mutation of existing logic showing to cause erosion

The redesigned PetClinic-Evolve experiment has produced its first complete run of each architecture. It is one chain per arm, so it is a first look and not the final verdict. But it already moves the measure that stayed flat for the whole first experiment, and it moves it in the direction the thesis predicts.

The setup, in one paragraph

Hold the coding agent fixed. Vary the architecture. One agent model builds the same app twice. In the Spring version, request logic lands in controller methods. In the OfficeFloor version, each rule is a small function wired together by YAML. Then sixty change requests land on the same endpoint, create owner, and we watch how each code base ages. Every fourth change is mutative: it revises earlier rules rather than only adding. The idea under test is that Spring concentrates the accumulating logic into one growing method, while OfficeFloor spreads it across many small functions and keeps the entry point flat.

The entry handler: the decisive measure

The cleanest number is the complexity of the one function the create endpoint routes through. In Spring that is the controller method. In OfficeFloor that is the pipeline's create function.

Checkpoint	Spring create handler (CC)	OfficeFloor create function (CC)
1	0	0
16	11	2
32	13	2
48	21	9
60	27	9

Cyclomatic complexity counts the independent paths through a function. A value around 27 is a method that is genuinely hard to hold in your head. Spring's create handler climbs to 27 and is still rising at the end. OfficeFloor's create function sat at 2 through the first half and ends at 9. Fitted as a trend, the Spring handler grows about 2.7 times faster per change. This is the mechanism in one line. Rules pile into the Spring handler. They attach beside the OfficeFloor one.

Where the complexity lives

Here is the part that a single summary number hides. Both code bases end with three or four functions above the usual complexity threshold. So a blunt erosion ratio looks similar for the two. But the functions that carry the complexity are not the same kind of thing.

Spring's busiest functions at the end:

Complexity	Function
27	the create controller method itself
19	a soundex name-coding routine
12	a region-code helper

OfficeFloor's busiest functions at the end:

Complexity	Function
19	a soundex digit routine
13	the possible-duplicate rule
13	a soundex helper
11	the telephone formatting rule

In Spring the single busiest function is the front door itself. In OfficeFloor the busiest functions are isolated, single-purpose units, and the front door stays flat. Notice the soundex routine sits at complexity 19 in both arms. That is the inherent complexity of the algorithm, not erosion, and it shows up in both. The difference is that Spring carries a complexity-27 god method on top of that shared cost. OfficeFloor does not.

Erosion, over the whole run

The erosion measure is the share of complexity that lives in functions above the threshold. It stayed at zero for the entire first experiment. In this run it moves.

Checkpoint	Spring erosion	OfficeFloor erosion
8	0.00	0.00
16	0.24	0.00
32	0.28	0.11
48	0.29	0.22
60	0.33	0.23

Spring erodes early, from checkpoint 16, and settles around 0.33. OfficeFloor stays at zero until checkpoint 32, then rises to 0.23. So OfficeFloor is not immune. The deep mutations in the back half do push it up. But it erodes later, it erodes lower, and it erodes in a spread out way rather than concentrating in the handler.

Correctness and cost

Both arms passed the same number of checkpoints cleanly. That number is dominated by a few of my own tests that were too brittle, which fired the same way in both arms, so I do not read much into the correctness magnitude from this run. I have since made those tests stricter, and the next run will give a correctness picture worth trusting. The early hint is that Spring broke a wider set of earlier rules.

Cost was close. The Spring chain cost about 79 US dollars in agent time. The OfficeFloor chain cost about 85. OfficeFloor did a little more total work, with more functions and more lines, because it spreads the same behaviour across more units. So it is not that one arm did less. It is that the two arms distributed the same job differently.

Honest limits

• This is one chain per arm. Any single chain can go its own way. The real result needs many independent chains per arm and the confidence intervals across them. That run is underway.
• The correctness comparison is muddied by brittle tests in this first chain. The structural comparison does not depend on those tests, so it stands.
• The erosion ratio alone is a poor summary. The entry-handler complexity and the identity of the busiest function are what separate concentration from distribution.

Where this goes

On the measure that could not move in the first experiment, the two architectures now separate clearly, and in the predicted direction. Spring's create handler became a complexity-27 god method that is still growing. OfficeFloor's create function stayed flat at 9, with the complexity pushed out to bounded, single-purpose functions. The structural half of the thesis is looking well supported. The safety half, whether the composed design also regresses less, awaits the run on the stricter tests. When the full multi-chain run completes, the slopes and their confidence intervals will turn this first look into an answer.

Improvements to experiment: more checkpoints, mutative checkpoints, blind regression measurements

This is a between experiments post. The first PetClinic-Evolve run gave a clear answer on some measures and a flat non answer on others. The flat parts turned out to be the interesting ones. Here is why they came out flat, and how the next experiment is built to force the question.

The question

PetClinic-Evolve keeps the coding agent fixed and makes the architecture the thing we vary.  A Spring version where request logic lands in controller methods. An OfficeFloor version where each rule is a small function wired together by YAML. Then a long stream of change requests lands on the same endpoint, create owner, and we watch how each code base ages.

What the first run showed, and where it went quiet

The first run walked 20 change checkpoints, with ten independent chains per architecture. It cost about 510 US dollars in agent time. Some signals came through cleanly. The blast radius measures separated between the two arms. So did the growth of the single entry handler, and a measure of how often new changes reopened old code. Those trends pointed the right way.

Two headline measures stayed silent, and that is what prompted the redesign.

• Erosion washed out. The erosion measure stayed at zero for the whole run, in both arms. No single method ever grew complex enough to trip the measure. The agent kept splitting logic into many small methods, so no one method ever spiked.
• Regressions were zero. Neither arm broke a previous behaviour. The safety difference the experiment exists to measure never showed up. It was hidden, not absent.

Why those two came out flat

Neither flat result was reassuring. Each one traced back to a choice in the test harness, not to the code being healthy. There were four causes.

1. The run was not long or deep enough

Twenty additive checkpoints did not push Spring past the point where a large method forms. Complexity did build up, but it stayed spread across many small methods. A measure that waits for one method to grow complex has nothing to report until the pressure is far higher.

2. Every checkpoint only added

Adding is the easy case, and it quietly favours the OfficeFloor design. Adding a brand new rule as a brand new function is exactly what that architecture is good at. Real maintenance is not only adding. It makes changes to previous requirements (i.e. mutating the existing logic of the application). An experiment made entirely of additions never tests the case that hurts most.

3. Regression was almost impossible by design

This was the important one. At each checkpoint the agent could see every previous test. So it had a full checklist of what not to break. With that checklist in front of it, of course it did not break anything.

4. The tests were too soft, and isolation was not tight enough

Some tests only checked that a field was present, not that it held the right value. A presence check cannot notice a wrong value, so it cannot notice a regression. Separately, the coding tool has a memory feature that can write notes between runs, which risks carrying knowledge across checkpoints that are meant to be independent.

How the next experiment is formed

The redesign tackles each cause directly. The thing we vary, the architecture, is unchanged. The instrument around it is rebuilt.

First run limitation	Redesign
Too short and shallow, so erosion never had a chance to appear.	Sixty checkpoints, three times longer, so pressure on the single handler builds well past the first run.
Only additions, which favoured the addition friendly arm.	Mutative checkpoints. Roughly every fourth change now revises earlier rules rather than only adding. Their reach grows from two earlier rules up to six, with deliberate deep changes near the middle and at the end.
The agent saw all past tests, so regression was near zero.	Blind regression measurement. The agent sees only the current checkpoint's test. The full set of past tests is used afterwards to check for regression.
Soft tests that only check for presence.	Exact tests. Every test asserts a precise value. Computed values such as hashes and check digits are recomputed inside the test. Look ups use small fixed tables shared by both arms.
Possible memory carried between checkpoints.	Isolation per turn. Each agent run starts with a fresh, login only setup, so the tool cannot carry notes between checkpoints or between arms.

The mutation, and the rule it needs

The mutative checkpoints are the heart of the redesign. When a checkpoint changes an earlier rule, it provides updated previous tests for the mutation.

The previous checkpoints being mutated are flagged by the checkpoint. A break in a previous checkpoint rule that was not on the list is a clear regression.  This now allows for a safety signal regarding the changes.

An early look, offered with caution

One Spring chain of the new design has run from start to finish. It is a single chain, one arm, and it is not the comparison. But it already shows the instrument now moves where it used to sit still. The erosion measure, which stayed at zero for the whole first experiment, now lifts as the change stream deepens. It rises through the first third of the run and peaks near the first deep mutation.

Checkpoint	Erosion measure
1	0.00
8	0.00
16	0.19
24	0.22
32	0.28
40	0.22
48	0.19
56	0.20
60	0.20

Over the same chain, the single busiest method grew from a complexity of one to nineteen, and the total code grew about thirteenfold. This is one chain, and it is Spring only. It validates the instrument, not the thesis.

Regressions appear now too, and they begin at the first mutations and build up, which is the shape the thesis predicts.

Where this goes

The first experiment was not a failure. It was a calibration. It showed which signals the design could already separate, and it showed exactly which measures needed a harder test before they could speak. The redesign is that harder test. Longer, with real mutation, with regressions made visible rather than assumed away, and with tests strict enough to trust.

The next experiment run is underway.

Methods companion · Measurement reference Spring OfficeFloor

Measuring AI-driven code degradation

A companion to the results paper. Every metric in the PetClinic-Evolve harness, defined precisely. What it computes, where the number comes from, and what it reads high on.

Author Daniel Sagenschneider Draft August 2026 Harness PetClinic-Evolve Companion to "Architecture as the independent variable"

1

Why the measurement comes first

A degradation study is only as good as its metrics. If the numbers are vague, the conclusion is vague. This companion defines each metric the harness records, so a reader can check the results against exactly what was measured. The results paper reports what the numbers did. This paper says what the numbers are.

Two design choices shape everything below. Both exist to make a per-change measurement trustworthy.

The agent delta is isolated. Each checkpoint produces two commits. The first is the agent commit. Its diff against its parent is exactly what the AI changed, and nothing else. The second is a reset commit that re-normalizes the tree for the next step. Any metric described as "of a change" is computed on that agent diff. So blast radius, coupling, and churn measure the agent's edit, not harness bookkeeping.

Capture now, derive later. A run stores only raw data on the branches. The agent envelope. The raw pass or fail of every test. The committed source at each step. Every metric below is recomputed offline from that history. So a metric can be defined or corrected after a run and re-applied to it, with no new agent cost. Every number was produced this way.

2

Ground rules for the structural metrics

Four conventions apply to every structural metric in Section 3.

• Production Java only. Structural metrics ignore test code, build files, and configuration. Java under the main source tree is counted. OfficeFloor's YAML wiring is counted separately and never mixed into a Java denominator.
• Per-function CC and SLOC. A static analyzer (lizard) reports two numbers per function. CC is cyclomatic complexity, the count of independent paths through the function. SLOC is source lines of code. These two feed most of what follows.
• The complexity threshold is 10. A function is "complex" when its CC exceeds 10. This is the standard Radon threshold, applied identically to both arms.
• The dynamic subsystem. Some metrics run over the whole application. Others run over the "touched subsystem": every production function in a file changed since the chain's base commit. Scoping this way follows the evolving footprint, and it automatically includes any new class the agent creates. So a growing hotspot cannot hide in a fixed file list.

3

Structural metrics

Erosion ratio · whole-app and scoped

mass(f) = CC(f) × √SLOC(f)
Erosion = Σ_{CC(f) > 10} mass(f)  ÷  Σ_{all f} mass(f)

Measures

How concentrated complexity is. The share of total complexity "mass" that sits in functions above the CC threshold.

Source

lizard CC and SLOC over production Java. Reported twice: over the whole app, and over the touched subsystem (erosion_scoped).

Reads high

When a few heavy functions hold most of the complexity.

In this run

wash Both arms near identical. A ratio normalizes away the concentration it targets. See Section 6.

Verbosity ratio

Verbosity = | CloneLines ∪ PatternLines |  ÷  LOC

Measures

Duplicated and boilerplate-flagged lines as a share of code.

Source

Clone lines from a copy-paste detector (jscpd, strict mode). Pattern lines from a structural search (ast-grep) run against a rule set of wasteful Java idioms. The two line sets are unioned, then divided by lines of code.

Reads high

When code is repeated or padded with boilerplate. Note the index can exceed 1.0, because a clone counts lines on both sides of the pair. It is an index, not a fraction.

In this run

separates OfficeFloor higher at every phase, from its many similar small functions. Neither arm's verbosity grows.

Blast radius per-change family

existing_fns_modified = # already-present functions whose body the diff touched
files_created = # new production files  ·  churn = production lines added / removed

Measures

How much pre-existing code a single change disturbs, and whether it adds or mutates.

Source

The agent diff. Hunk headers (git diff -U0) give the changed line ranges. lizard gives each function's line range. A function counts as modified when a changed range overlaps its body, and its file existed before the change. New files are counted as created.

Reads high

existing_fns_modified is high when a change reaches into working code. files_created is high when a change adds new units instead.

In this run

separates Spring's existing_fns_modified slope excludes zero; OfficeFloor's is flat and it creates ~19 files per chain.

Hotspot CC worst function

hotspot_cc = max_{f in touched subsystem} CC(f)

Measures

The single most complex production function in the evolving footprint, reported with its name.

Source

lizard over the touched subsystem. The function with the highest CC, ties broken by SLOC.

Reads high

When one function is becoming a god-method.

In this run

overlap Both arms rise; Spring higher, but the confidence intervals overlap.

Weighted Methods per Class (WMC) worst class

WMC(class) = Σ_{methods m in class} CC(m)
wmc_max = max_class WMC(class)

Measures

The god-class indicator. The heaviest class by total method complexity.

Source

lizard functions grouped by file (one top-level class per Java file), summed per class, maximized over classes in the touched subsystem.

Reads high

When one class carries many rules. This catches what erosion misses: a class that stays tidy method by method while accumulating many methods.

In this run

overlap Both rise; Spring's final WMC is ~2× OfficeFloor's, but the slopes overlap.

Entry-handler CC the front door

entry_cc = CC( f ) where f matches the arm's create-entry pattern

Measures

The complexity of the one function the create endpoint routes through.

Source

An arm-specific regular expression over "file::function". Spring matches OwnerRestControllerV1::addOwner. OfficeFloor matches its designated create-entry, BuildOwner::service. lizard gives the CC.

Reads high

When the single handler absorbs each new rule rather than delegating it.

In this run

separates cleanly Spring +0.25 per checkpoint, OfficeFloor +0.06. The confidence intervals do not overlap. Final CC 6.7 vs 1.9.

Change spread reach

packages_touched = # distinct package directories the diff's production Java touches

Measures

How far across the package tree a single change reaches. A sibling to blast radius, measured by directories rather than functions.

Source

The agent diff. Distinct parent directories of the changed production Java files.

Reads high

When a change is scattered across many packages.

In this run

both near zero Neither arm's spread trends.

Re-edit rate temporal coupling

for each function the change edits, blame its whole body at this commit
re-edit rate = ( body lines authored by an earlier checkpoint ) ÷ ( total edited-function body lines )

Measures

When a change edits existing code, how much of that code earlier rules wrote.

Source

For every production function the change modifies, git blame attributes each body line to the checkpoint that last wrote it. A line counts as "prior" when its author is a checkpoint after the base but before this one. Whole function bodies are counted on purpose. A one-line insert into a large shared method still signals coupling, and line-of-diff blame would miss it.

Reads high

When new rules keep reopening functions that earlier rules grew.

In this run

separates cleanly Spring's slope is positive, OfficeFloor's is negative. The intervals do not overlap. The sign of the coupling trend flips between the architectures.

Function-package stats OfficeFloor-specific

fn_count, fn_nloc_avg, fn_nloc_max, fn_cc_max over the wired-function package

Measures

The size distribution of OfficeFloor's composed-function package.

Source

lizard over a configured package glob (the rest/function tree).

Reads high

The healthy-growth signal is a rising count while the average and maximum function size stay flat. That is addition without bloat.

In this run

Count rises steadily; per-function size stays flat. Consistent with add-not-mutate.

Lines of code size, reported separately

java_loc = production Java SLOC  ·  yaml_loc = OfficeFloor wiring lines

Measures

Raw size. Kept as context, and as the denominator for verbosity.

Source

lizard for Java. A non-blank non-comment line count for YAML.

Note

YAML is never folded into a Java ratio. It is reported on its own so OfficeFloor's habit of spreading logic into wiring cannot distort a Java metric.

4

Correctness metrics

Correctness is scored from the raw pass or fail of every acceptance test, captured at run time. The tests are black-box. They hit the REST API, so both arms are judged by identical externals.

The test taxonomy. There is one test class per checkpoint, named CpNN, tagged so the harness runs checkpoints one through K at checkpoint K. The method-name prefix encodes a category: core, error, or functionality. A test whose own checkpoint is earlier than K counts as a Regression test at K, whatever its category.

Strict / ISO / Core pass correctness gates

strict = all selected tests pass  ·  iso = this checkpoint's own tests pass  ·  core = the core tests pass

Measures

Three tiers of "did it work". Strict is everything. ISO ignores the regression tests and asks only whether this checkpoint's new behavior works. Core asks only the essential path.

Reads high

All three are 1.0 when the checkpoint fully satisfies its suite.

In this run

Strict pass was 1.0 in every phase, both arms.

Normalized Change SWE-CI · range -1 to 1

if passed ≥ baseline:  (passed − baseline) ÷ (target − baseline)
else:  (passed − baseline) ÷ baseline

Measures

Net progress in passing tests from one checkpoint to the next, on a signed scale.

Source

The passing-test set before and after the checkpoint. "baseline" is the prior passing count, "target" the total selected.

Reads

Positive for improvement, negative for regression. The definition is asymmetric on purpose. It punishes a regression harder than it rewards an equal-sized improvement.

Regressions count

regressions = | passing_before − passing_after |

Measures

How many tests that passed before the checkpoint fail after it.

Reads high

When a change breaks previously working behavior. This is the direct safety signal.

In this run

no difference Zero regressions in either arm across 200 checkpoints each.

EvoScore SWE-CI · discounted success

per chain:  ( Σ_i γⁱ · s_i ) ÷ ( Σ_i γⁱ ),  then averaged over chains

Measures

Success across a whole chain, weighted by checkpoint. si is the strict-pass indicator at checkpoint i.

Source

The strict-pass sequence per chain. The discount γ is a parameter; γ ≥ 1 rewards staying green late in the run, when the codebase is largest.

In this run

1.0 at every γ for both arms.

Zero-Regression Rate across chains

ZRR = ( # chains with zero total regressions ) ÷ ( # chains )

Measures

The share of full runs that never broke anything.

Reads high

1.0 means every chain stayed regression-free end to end.

In this run

no difference 1.000 for both arms.

5

Process metrics

These come straight from the agent session envelope, captured at run time. They cannot be recomputed later, so they are stored raw.

Cost, tokens, and duration the agent envelope

cost_usd

Dollar cost of the checkpoint's agent session.

tokens

Input, output, cache-read, and cache-creation token counts. Cache-read is a proxy for how much prior context the model re-read.

num_turns

Tool-use turns the agent took to finish.

duration_ms

Wall-clock time, including tool runs and any waits.

duration_api_ms

Model inference time only. The cleaner "thinking cost" signal.

attempts

Every attempt is recorded, including failed or rate-limited ones and their wait time, so true cost and wall-clock are recoverable.

In this run

Cost and API time fell slightly over each run in both arms. The "comprehension gets more expensive" idea did not appear at this scale.

6

The statistic: degradation slope

A single metric at a single checkpoint is noise. The signal is the trend across the run. The harness reduces each metric to one number per arm, with an interval.

Degradation slope m the headline number

m = OLS slope of ( metric value vs checkpoint index )
reported on the mean curve, with a 95% bootstrap CI resampled over chains

Mean curve

At each checkpoint, average the metric across the ten chains. Fit a straight line to that averaged curve. Its slope is m.

Confidence interval

Resample the ten chains with replacement, 2,000 times. Refit m each time. The 2.5 and 97.5 percentiles are the 95% interval. The interval reflects chain-to-chain variability.

Phase means

Checkpoints one to twenty are also binned into five phases, Start to Final, to show the trajectory in a small table.

How to read a slope A metric separates the two architectures when one arm's interval excludes zero and the other's brackets it. It separates strongly when the two arms' intervals do not overlap. A rising slope with an interval that stays clear of zero is the degradation signal. An interval that straddles zero is a flat metric.

The one caveat that matters. Not every metric can see every difference. Erosion is a whole-application ratio. When one arm adds many small functions, it inflates the denominator at the same time the other arm concentrates complexity in the numerator. Both effects move the ratio the same way, so the ratio cancels the difference. The lesson generalizes. For architectural degradation under AI-driven change, lead with un-normalized, per-change statistics such as blast radius, entry-handler complexity, and re-edit rate. Treat aggregate ratios as secondary. The headline number should be one that can discriminate.

7

Controls that keep a number honest

Three controls protect the metrics from being gamed or contaminated. They are part of the measurement, not a footnote to it.

• Experimenter-owned tests, reset before scoring. The acceptance tests are black-box and reset to their authored version before the correctness gate runs. An agent that weakens a test cannot produce a false pass. Any edit it makes is recorded, then reverted.
• A pinned project guide. The leveling document is restored to its base version after every checkpoint, so it can never become accumulating memory across checkpoints.
• Provenance and a config snapshot. Each run records the model, the tool versions, the agent environment, and a snapshot of the exact metric configuration used. So a metric recomputed later uses the same definitions the run was scored under.

In this run the agent never edited a pinned document and never tampered with a test, across all 400 sessions. The controls were never triggered. Their presence still removes two ways the numbers could have been wrong.

---

R

References and notes

1. SlopCodeBench · arXiv:2603.24755. Source of the Erosion and Verbosity definitions, the degradation-slope statistic, and the prompt-intervention arms.
2. SWE-CI · arXiv:2603.03823. Source of Normalized Change, EvoScore, and Zero-Regression Rate.
3. Weighted Methods per Class follows the Chidamber and Kemerer object-oriented metric suite.
4. OfficeFloor · officefloor.net. The graph-of-functions framework used in the OfficeFloor arm.

Companion. The results, with confidence intervals and the full slope table, are in "Architecture as the independent variable". This paper defines the instruments. That paper reports the readings.

PetClinic-Evolve · metrics reference · claude-opus-4-8 · every definition matches the harness implementation and is recomputed from the committed run data.

Preprint · Empirical software engineering Spring @RestController OfficeFloor functions

Architecture as the independent variable

Hold the AI coding agent fixed for 400 sessions. Vary only the architecture it works in. Does the codebase degrade because of the agent? Or because of the shape it is asked to grow?

Author Daniel Sagenschneider Draft August 2026 Harness PetClinic-Evolve Agent claude-opus-4-8 Read ~22 min

Abstract

AI coding agents now take on long, iterative work. A worry has followed. Codebases seem to degrade under sustained agentic change. That degradation is usually studied as a property of the agent. We ask a different question. Is it a property of the architecture the agent is asked to extend?

We built a harness that holds the agent fixed. Each step gets a fresh, context-free claude-opus-4-8 session. Architecture is the only thing that varies. The same twenty owner-creation rules were added, checkpoint by checkpoint, to two forks of one REST application. One fork is a conventional Spring @RestController. The other is an OfficeFloor pipeline of YAML-composed functions. Ten chains per arm, times twenty checkpoints, gives 400 agent sessions. Every metric is re-derived from the commit history. Each slope carries a 95% bootstrap confidence interval over chains.

The result is sharp in one direction and null in another. How change is absorbed separates the two architectures cleanly. In the Spring arm, per-change blast radius, create-endpoint complexity, and temporal coupling all rise. Their confidence intervals exclude zero. In the OfficeFloor arm they stay flat, and it keeps adding small isolated units. Yet the benchmark's nominal "erosion" ratio does not separate the arms at all. It was meant to be the decisive statistic. It was not. And neither architecture produced a single test regression across 200 checkpoints each. Architecture governs the mechanism of degradation. At this horizon it has not yet changed the observed correctness. We report what held, what washed out, and why the deciding statistic was not the one we set out to measure.

1

The question

A coding agent asked to add one small feature usually does fine. The concern is the hundredth small feature. Iterative, agent-driven extension may leave a codebase quietly worse. More tangled. More concentrated. More expensive to keep changing. Anecdotes are common. Controlled measurement is rare.

Most attempts treat the agent as the variable. Different models. Different prompts. Different scaffolds. The task stays fixed. That tells you which agent degrades a codebase least. It cannot tell you something a system designer also needs to know. Does the architecture itself change how much damage accumulates?

So we invert the design. We hold the agent constant. Architecture becomes the independent variable. We take two implementations of one application. We feed both the same feature requirements, one checkpoint at a time. We watch how each absorbs twenty rounds of accumulating change. If degradation is purely the agent's fault, the arms should degrade alike. If architecture is a lever, they should diverge. The shape of the divergence then tells us which architectural properties matter.

Contribution A reproducible harness that isolates architecture as the cause of AI-driven code degradation. The agent, the task, the tests, and the leveling documentation are held fixed across both arms. It ships with a full result set, with confidence intervals, on which structural properties separate the architectures and which do not.

2

Results at a glance

We track thirteen metrics. The arms separate on the ones that describe how a change touches the codebase. They refuse to separate on the aggregate "erosion" ratio and on correctness. Here is the verdict, one dimension at a time.

◆ Separates cleanly

Blast radius

Spring's per-change edits to existing functions grow. The slope CI excludes zero. OfficeFloor's is flat. Spring disturbs twice the existing surface.

◆ Separates cleanly

Create-endpoint complexity

Spring's single addOwner handler bloats to CC 6.7. OfficeFloor's create-entry stays at 1.9. The CIs do not overlap.

◆ Separates cleanly

Temporal coupling

Spring keeps re-opening earlier rules' code. Its slope is positive. OfficeFloor's re-edit rate falls. The CIs do not overlap.

◆ Separates cleanly

Add vs. mutate

OfficeFloor creates about 19 new files per chain. Spring creates about 1. The two architectures absorb a new rule in opposite ways.

≈ Washes out

Erosion ratio

The nominal decisive metric. Both arms end near 0.012. Both CIs touch zero. The whole-app ratio cannot see the difference.

✕ No difference

Correctness & safety

Zero regressions. Full strict-pass. EvoScore 1.0. In both arms. Erosion stayed latent. It never broke anything at this horizon.

How to read the rest of this paper. Section 3 is the method. It covers the arms, the task, the fixed agent, the metrics, and the statistics. Section 4 is the evidence. It follows the six verdicts above. One figure anchors it. That figure is a coefficient plot of every degradation slope with its confidence interval. Sections 5 to 7 discuss why the designated statistic failed, the threats to validity, and how to reproduce every number.

3

Method

3.1  Two arms, one application

Both arms fork the same REST implementation of the Spring PetClinic sample. They share a data model, a database schema, and one black-box HTTP contract. They differ only in how request-handling logic is structured.

• Spring. The conventional layered style. New behavior for POST /api/owners is added by editing a Spring @RestController method and the service beneath it.
• OfficeFloor. A graph-of-functions style. The create-owner endpoint is a pipeline wired in YAML. New behavior is added by writing a new small function class and wiring it in.

The two base branches are matched. The first checkpoint starts each arm from an equivalent, working application. Everything downstream of the architecture is held identical. That means the requirements, the acceptance tests, the agent, the model, and a fixed project guide.

3.2  The task: twenty accumulating rules

Each checkpoint adds one requirement to the owner-creation endpoint. Every requirement stays in force for all later checkpoints. The rules are ordinary business logic. Duplicate detection by household, telephone, and email. Telephone and city-name normalization. Derived fields such as initials, customer code, membership number and tier, namesake count, and locality. A default registration date. A shared-household flag. City-capacity limits. An audit trail with a bulk-signup warning. By checkpoint twenty, one POST endpoint must enforce twenty interacting rules. That is exactly the kind of accreting hotspot where architecture should start to matter.

3.3  The fixed agent

At every checkpoint the harness launches a fresh, headless claude-opus-4-8 session. It gets only that checkpoint's specification. No conversation history. No memory of earlier checkpoints. No resume. The condition is deliberate. The agent must reason from the code's current structure alone. That is the setting in which a self-describing architecture can pay off. It also removes carried context as a confound. The agent runs with the same tools and the same neutral just-solve prompt in both arms. The prompt says to implement the change so all tests pass.

3.4  Isolation and anti-gaming

Three controls keep the comparison honest. They stop the agent from inflating its own score.

• Experimenter-owned acceptance tests. Each checkpoint's tests are black-box. They hit the REST API. The harness injects them and resets them to the authored version before the correctness gate runs. So an agent that weakens a test cannot produce a false pass. Any edit the agent makes to a test is recorded, then reverted.
• A pinned project guide. A fixed CLAUDE.md sits in both arms. It is restored to its base version after every checkpoint. It can never become cross-checkpoint memory.
• A two-commit history. Each checkpoint makes an agent commit. Its diff is exactly what the agent changed. A reset commit follows. It re-normalizes the tree and stages the next checkpoint's test. Isolating the agent's true delta is what makes the structural metrics measurable per change.

Across all 400 sessions the agent edited a pinned document 0 times. It tampered with an acceptance test 0 times. The controls were never exercised. Their presence still removes two obvious threats to validity.

3.5  Capture now, derive later

A run stores only raw capture on the branches. That capture is the part that cannot be recomputed. It holds the agent's cost, tokens, and full event stream. It holds the raw pass/fail map of every test. It holds the true pre-normalization diff. It holds a provenance manifest with the model, tool versions, and a config snapshot. Every derived metric is recomputed later from the commit history. That covers erosion, verbosity, blast radius, coupling, and the rest. The split has a payoff. A metric can be defined or fixed long after a run and re-applied to it, with no new agent cost. Every number in this paper came from that offline re-derivation over the committed trees.

3.6  What we measure

Structural metrics are computed over production Java only. The same tools and thresholds apply to both arms.

• Erosion. The share of total complexity mass that sits in high-complexity functions. Mass is cyclomatic complexity times the square root of SLOC, per function. High means above the standard CC = 10 threshold. This is the benchmark's headline statistic.
• Verbosity. Flagged clone lines and anti-pattern lines, per line of code.
• Blast radius. For a checkpoint's change: how many already-existing functions it modifies, and how many new files it adds.
• Weighted Methods per Class. The god-class indicator. The highest per-class sum of method complexity.
• Entry-handler complexity. The cyclomatic complexity of the one function the create endpoint routes through. In Spring that is addOwner. In OfficeFloor it is the designated create-entry.
• Temporal coupling, or re-edit rate. When a checkpoint edits an existing function, the share of that function's body written by earlier checkpoints. High means new rules keep re-opening old ones.
• Process and correctness. Cost, tokens, and API duration. Then Strict, ISO, and Core pass, Normalized Change, regressions, EvoScore, and Zero-Regression Rate.

3.7  Statistics

The degradation slope m is the OLS slope of a metric against checkpoint index within a chain. We report the mean-curve m with a 95% bootstrap confidence interval. The interval is resampled over the ten chains, with 2,000 replicates. A metric separates the arms when one arm's slope CI excludes zero and the other's brackets it. It separates more strongly when the two arms' CIs do not overlap. The scale is 10 chains, times 20 checkpoints, times 2 arms. That is 400 agent sessions. Each slope is fed by 200 checkpoints per arm.

Positionality and conflict of interest. OfficeFloor is authored by the experimenter. The OfficeFloor arm's base was purpose-built for this comparison. We mitigate this several ways. Identical requirements. Identical black-box tests, scored by an anti-gaming gate. A fixed agent. Fully published commit histories, so every number is independently recomputable. A reader should still weight the framing accordingly. See section 6.

4

Results

4.1  The degradation slopes, at a glance

Figure 1 plots the degradation slope and its 95% interval for every metric where the arms behave informatively. The metrics have different units. So each row is scaled to its own range. Within a row, two things matter. Does an interval cross the zero line? Do the two arms' intervals overlap?

Spring OfficeFloor ● slope estimate · ▬ 95% CI · ┆ zero tick · each row scaled independently

Existing fns modified

◆ separates · per-change blast radius

0

+0.052

+0.0056

Files created

◆ separates · add vs mutate

0

-0.0021

+0.021

Entry-handler CC

◆ separates · create endpoint

0

+0.252

+0.062

Re-edit rate

◆ separates · temporal coupling

0

+0.011

-0.024

Hotspot CC

≈ overlap / wash · worst function

0

+0.241

+0.172

WMC (god-class)

≈ overlap / wash · worst class

0

+1.77

+1.44

Erosion

≈ overlap / wash · whole-app ratio

0

+0.0007

+0.0006

Verbosity

≈ overlap / wash · clones and patterns

0

-0.0058

-0.0052

Figure 1. Degradation slopes with 95% bootstrap CIs. Four metrics describe how change lands: blast radius, files created, entry-handler CC, and re-edit rate. They separate the arms. Spring's disturbance and coupling rise. OfficeFloor's stay flat or fall. The two aggregate complexity slopes, hotspot CC and WMC, both rise but overlap. Erosion is the nominal decisive statistic. It is a wash. Both slopes hug zero.

4.2  How a change lands: the arms separate

The clearest signal is blast radius. It measures how much pre-existing code a single new rule disturbs. In Spring it climbs over the run. The slope is +0.052, CI 0.016 to 0.084. That excludes zero. In OfficeFloor it is statistically flat. Aggregated, Spring modifies about twice the existing surface per chain. It almost never gets away with touching nothing. OfficeFloor takes the zero-blast path more than a third of the time. It adds a rule without re-opening any existing function.

Table 1. Blast radius per chain (200 checkpoints per arm)

Signal	Spring	OfficeFloor	Reading
Existing functions modified / chain	55.8	28.2	Spring 2× the disturbance
Zero-blast checkpoints	11/200	70/200	OF touches nothing 35% of the time
New files created / chain	1.2	18.7	add vs. mutate
Line churn / chain	+500 / −22	+662 / −30	OF adds more lines, in new units

This is the mechanism in one table. A new rule in Spring is folded into existing methods. The same rule in OfficeFloor is attached as a new function. A cross-check from the agents' own event streams makes it concrete. Over the first five chains, the OfficeFloor agents wrote 75 distinct function and handler classes. That is a whole rest/function/owner/ package. Think AssignMembershipTier, CheckCityCapacity, AuditOwnerCreation. Spring wrote 20. It folded the same logic into its controller and service.

The consequence shows up in two places. The create endpoint. And temporal coupling. These are the cleanest separations in the study. Here the arms' confidence intervals do not even overlap.

Table 2. Concentration and coupling (final values and slopes)

Metric	Spring	OfficeFloor	Slope CIs overlap?
Entry-handler CC (final)	6.7	1.9	no, clean split
Entry-handler CC slope m [CI]	+0.25 [.14,.38]	+0.06 [.05,.07]	Spring bloats ~4× faster
Re-edit rate (mean)	0.331	0.216
Re-edit rate slope m [CI]	+0.011 [.005,.016]	−0.024 [−.030,−.018]	no, opposite signs
God-class WMC (final)	64.5	35.7	Spring's worst class ~2× heavier

Spring's single create handler grows about 4× faster than OfficeFloor's. It ends more than three times as complex. And its re-edit rate rises while OfficeFloor's falls. So as the run proceeds, Spring keeps re-opening code written for earlier rules. OfficeFloor increasingly leaves earlier rules alone. The sign of the coupling slope flips between the two architectures. That is the erosion thesis in its strongest, cleanest form. It is just not measured by the erosion metric.

4.3  The designated statistic washes out

The benchmark's headline metric is erosion. It is the fraction of whole-app complexity mass held in high-complexity functions. It was meant to be the decisive number. It is not. Both arms' slopes hug zero. Both confidence intervals include zero. The phase-binned means converge on nearly the same final value.

Table 3. Erosion, phase-binned mean (Start to Final)

Arm	Start	Early	Mid	Late	Final	slope m [CI]
Spring	0	0	0	0.0036	0.0128	.0007 [0,.0017]
OfficeFloor	0	0.0059	0.0057	0.0052	0.0121	.0006 [0,.0014]

The reason is structural, and it is instructive. Erosion is a ratio over the whole application. OfficeFloor adds many small functions. That inflates the denominator, the total complexity mass. At the same time, Spring concentrates complexity in the numerator, a few high-complexity functions. Both effects push the ratio the same way. Against a roughly 280-function application, one bloating controller barely moves it. So the metric meant to be decisive is the one metric blind to the difference. It normalizes away the very concentration it was trying to detect. The discriminating statistics are the un-normalized, per-change ones. Blast-radius slope. Entry-handler CC. Re-edit rate.

4.4  Verbosity: a real cost, but a constant one

With clone and pattern detection restored, verbosity gives an honest counterpoint. OfficeFloor carries more duplication than Spring at every phase. Its 75-odd small function classes share near-identical scaffolding. The clone detector flags it. But neither arm's verbosity grows. Both slopes are slightly negative.

Table 4. Verbosity index, phase-binned mean (flagged lines ÷ LOC)

Arm	Start	Early	Mid	Late	Final	slope m [CI]
OfficeFloor	1.20	1.19	1.16	1.14	1.12	−.0052 [−.0054,−.0049]
Spring	1.08	1.07	1.04	1.02	0.99	−.0058 [−.0061,−.0054]

This exposes the real cost of the OfficeFloor strategy. Isolation-by-proliferation pays a constant boilerplate tax. There is more repeated scaffolding up front. In exchange, the per-change blast radius stays bounded and does not compound. Spring pays the opposite way. Little duplication, but an entanglement cost that keeps rising. One note on the numbers. The verbosity index goes above 1.0 because clone lines count both sides of each pair. It is an index, not a fraction.

4.5  No correctness difference, and cost falls

Two outcomes people expect to move did not. Both arms are perfectly green. Strict-pass is 1.0 in every phase. EvoScore is 1.0 at every discount factor. The Zero-Regression Rate is 1.000. No checkpoint in either arm broke a previously passing test, across 200 checkpoints. The structural erosion is entirely latent. It accumulates. At this twenty-checkpoint horizon it never bit. Per-change cost also fell slightly over each run, in both arms. The slope is about −0.008 USD per checkpoint. That is the opposite of the idea that comprehension gets more expensive as the hotspot grows. Prompt-cache reuse and the agent's own efficiency dominate any rising-comprehension effect at this scale.

The finding in one sentence Architecture cleanly governs how AI-driven change is absorbed. Spring's disturbance and coupling grow. OfficeFloor's stay flat. But at twenty checkpoints that difference in mechanism has not yet changed correctness. And the aggregate erosion ratio is too coarse to see it.

5

Discussion

5.1  Architecture is a real lever, on the mechanism

The strong reading is not that OfficeFloor is better. It is that the two architectures put an agent's accumulating changes in structurally different places. And that this is measurable, with tight confidence intervals. Spring absorbs each rule by mutating and re-coupling existing units. So blast radius, create-endpoint complexity, and temporal coupling all trend up. OfficeFloor absorbs each rule by adding an isolated unit. So those same quantities stay flat. The cost is more boilerplate. There is a practical takeaway for anyone building a system that AI agents will maintain over a long horizon. The variable that separated is the one to manage. Keep the per-change blast radius bounded. Then the compounding costs, coupling and hotspot concentration, never get a foothold.

5.2  Choose the right decisive statistic

Our most transferable finding is a caution about metrics. The pre-designated decisive metric was the whole-application erosion ratio. It is the wrong instrument for this question. Normalization cancels the concentration it targets. Studies of architectural degradation under AI-driven change should lead with un-normalized, per-change statistics. Blast-radius slope. Entry-point complexity. Temporal coupling. Treat aggregate ratios as secondary. The headline number should be the one that can discriminate.

5.3  Latent vs. manifest degradation

Zero regressions in either arm is itself a result. It says Spring's structural entanglement is latent risk. It is not yet realized failure. Not at twenty checkpoints, against this acceptance suite. So the "OfficeFloor is safer" claim, in the sense of fewer broken tests, is unsupported by this run. It stays an argument about maintainability and future change-cost. The structural metrics support that. They do not speak to observed correctness. Surfacing a correctness difference would need more. A longer horizon. A harder or more interdependent test suite. Or a task where the rules conflict, rather than merely accumulate.

6

Threats to validity

• Conflict of interest. OfficeFloor is the experimenter's framework. Its arm's base was purpose-built. Identical requirements, black-box anti-gaming tests, a fixed agent, and published commit histories mitigate the risk. They do not remove framing bias.
• Single agent, single strategy. One model, claude-opus-4-8. One neutral prompt, just-solve. The harness supports intervention arms, anti-slop and plan-first, and other models. None are run here. Results may not carry across agents.
• Single domain and hotspot. One application. One accreting endpoint. One family of business rules. Architectures that look different on a create-owner pipeline may converge on other workloads.
• Horizon. Twenty checkpoints separated the mechanism. It did not manifest a correctness or cost difference. The most interesting effects may live past this horizon.
• Metric construct validity. Erosion's normalization limit is covered above. The verbosity index double-counts clone pairs. Structural metrics recomputed offline depend on the analysis-time versions of the complexity and clone tools. Provenance pins them, but they are not identical to run-time.
• Independence. Chains are independent by construction. Each starts a fresh worktree from the base. But both arms ran from one machine and one quota window. Systematic model drift over the run would affect both arms alike.

7

Reproducibility

Every result here is recomputable from published artifacts. Each of the twenty chains is a git branch. It is named evolve/<run>/just-solve/<arm>/chain<n>. Per checkpoint, its history is an agent commit and a reset commit. A final commit carries the raw capture, a provenance manifest, and a snapshot of the exact analysis configuration used. The analysis tool re-derives every metric, including any added later. It materializes each checkpoint commit and recomputes over its source. So the tables above regenerate on demand. New metrics apply to the same run. Neither step re-invokes the agent.

Appendix A below lists the complete degradation-slope table for all thirteen tracked metrics, both arms, with 95% CIs. It is the full data behind Figure 1.

8

Conclusion

We held the coding agent fixed and varied only architecture. Architecture is a real lever on how AI-driven change accumulates. Across 400 sessions, the Spring arm's per-change blast radius, create-endpoint complexity, and temporal coupling rise. Their confidence intervals exclude zero. The OfficeFloor arm's stay flat. The difference is not visible in an aggregate erosion ratio. That ratio cannot see it. It is visible in the un-normalized statistics of how each change lands. That divergence in mechanism did not become a correctness difference at this horizon. Both arms stayed perfectly green. The next questions are about horizon and generality. Longer runs. Conflicting rather than accumulating requirements. More architectures. More agents. The aim is to learn whether the latent entanglement one architecture builds up eventually becomes the failures the other avoids.

---

A

Appendix. Full slope table

Degradation slope m (OLS on checkpoint index) with 95% bootstrap CI over 10 chains

Metric	Spring  m [CI]	OfficeFloor  m [CI]	Separates
Structural. How change lands
existing_fns_modified	+0.0523 [.016,.084]	+0.0056 [−.009,.023]	yes
files_created	−0.0021 [−.0037,−.0005]	+0.0208 [.008,.030]	yes
entry_cc	+0.252 [.137,.381]	+0.062 [.048,.068]	yes
reedit_rate	+0.0107 [.005,.016]	−0.0244 [−.030,−.018]	yes
Structural. Aggregate
hotspot_cc	+0.241 [.163,.343]	+0.172 [.102,.243]	overlap
wmc_max	+1.77 [1.38,2.06]	+1.44 [1.36,1.49]	overlap
packages_touched	+0.0092 [−.001,.017]	−0.0011 [−.007,.005]	both ~0
erosion	+0.0007 [0,.0017]	+0.0006 [0,.0014]	wash
erosion_scoped	+0.0021 [0,.0051]	+0.0014 [0,.0032]	wash
verbosity	−0.0058 [−.006,−.005]	−0.0052 [−.005,−.005]	level, not slope
Process
cost_usd	−0.0082 [−.015,−.002]	−0.0079 [−.015,−.001]	both fall
duration_api_ms	−1948 [−3552,−424]	−1267 [−2379,−151]	both fall
cache_read_tokens	−1643 [−8867,4647]	−4600 [−10060,−99]	noisy

Correctness held constant across both arms and all phases. Strict-pass 1.0. EvoScore 1.0 at γ = 1, 1.5, 2. Zero-Regression Rate 1.000. Pinned-doc touch 0/200. Acceptance-tamper 0/200.

R

References and notes

1. SlopCodeBench · arXiv:2603.24755. The no-context iterative-extension benchmark. The harness borrows its Erosion and Verbosity metric definitions, the degradation-slope statistic, and the prompt-intervention arms.
2. SWE-CI · arXiv:2603.03823. The CI-gate evaluation framework. The harness borrows Normalized Change, EvoScore, and Zero-Regression Rate.
3. spring-petclinic-rest. The REST PetClinic sample application both arms fork from.
4. OfficeFloor · officefloor.net. The graph-of-functions application framework used in the OfficeFloor arm.

PetClinic-Evolve · 400 agent sessions · claude-opus-4-8 · run 2026-08 · figures and tables generated from the committed run data.
Draft for a blog and preprint. Palette and figures designed for the subject. Validated for colour-vision-deficiency separation.