Morphy 🧱 — modernize run bilans

Carries a repository through a programme of modernisation lots — steps whose entry and exit are both deterministic gates — against a behavioural oracle it is forbidden to rewrite. See bots/modernize/.

2026-07-28 — the modernisation converges: green gate, green net, verified twice (runs 019fa8e7, 019fa8fd)

• Status: VALIDATED. Two major build-tool versions and four framework lines crossed, with observable behaviour identical — proven, not asserted.
• Campaigns: 13 min / $0.78 (aborted on a baseline defect) then 24 min / $2.07.

The verdict, and the second opinion

gate_passed true   oracle_passed true   refs_untouched true   log_tail (empty)

Replayed independently afterwards from the working repository, outside any run, against a jar the new toolchain built: 7/7 mutants detected, negative control silent across all 18 entries, collateral 0, exit 0. A green from the gate that produced the work is worth less than a green obtained somewhere else.

What the build said the whole time: nothing

It was green from the moment the tooling landed, and stayed green while seven behavioural divergences existed — among them template security attributes passing through uninterpreted, so an anonymous visitor received the menu entries reserved for authenticated roles.

No build check can see that. It is not what a build does. Every argument for putting a behavioural net underneath a modernisation reduces to this run.

Two shortcuts refused, and why they matter more than the fixes

This was the first lot in the programme where code was written to satisfy the net, which is exactly when a net becomes dangerous: the pull is to treat red as something to remove rather than a symptom to understand. The intent carried the test — would this change still be right if the reference did not exist? — and the campaign applied it against itself twice.

• Pinning a library back to its previous version would have closed one divergence without touching a template. Refused: undoing the upgrade at the precise point the judge is looking is not finishing it.
• A generally recommended code form was refused because it added an attribute to the rendered HTML. A fix that changes recorded behaviour to repair a cause that does not affect it is not a fix.

It also flagged its own borderline case rather than burying it among the obvious ones — a source-formatting change, kept because it states something true about the template independently of the reference, and carrying a comment so a later reader does not undo it by reflex. Flagging the limit case instead of drowning it is the difference between a report and a plea.

Lesson: entry count was the wrong metric

The preceding pass expected the divergent-entry count to fall sharply after closing two of four causes. It did not move — 14 before, 14 after — and the pass diagnosed why better than the operator had: several independent causes were hiding behind entries already counted as divergent, and closing one cause on an entry carrying two does not turn it green. The measurement that meant something was the cause inventory: three configuration-only, four needing code.

An expectation stated in an intent is falsifiable, which is the point. This one was falsified, and the correction came back as a finding rather than as noise.

2026-07-28 — second lot: a perfectly green build that does not run (run 019fa826)

• Status: the result the whole approach exists to produce. The lot is blocked, and the finding is worth more than a completed lot would have been.
• Versions: bot v0.1.0 + the declared-block stop · iterion built from HEAD · only_lot=L1b, max_passes=4.
• Campaign: 27 min, 65 339 tokens, $1.63. Gate: 9 seconds.

The verdict

gate_passed     true     both exit_gate commands exited 0
refs_untouched  true     nothing moved under the oracle's references
oracle_passed   FALSE    the application no longer starts
lot_blocked     true     declared, with the cause committed
stop            true     the run ended instead of looping

The lot reached the programme's target build-tool version. The build is green, the packaging task is green, there is not one warning to read. Reported without a behavioural net, this is a milestone: target tooling reached, framework plugin raised, zero regressions detected — because nothing looked.

The application does not start. Two independent causes, both in application configuration: a bean-definition override that the older framework line tolerated silently and the newer one refuses by default, and a JDBC driver class the newer line derives from the connection URL which does not exist in the pinned connector version.

Neither is exotic. Both are the ordinary consequence of a framework line moving, and both are invisible to every check a build performs. This is the failure mode that surfaces at deploy time on a good day, and in production on a bad one.

The gate took nine seconds to establish it, because the application died on startup and the net could not even begin. The cheapest possible discovery.

What the campaign did right, and refused to do

It refused to fix them. Both corrections require touching application configuration, which the lot's intent excludes. It marked the lot blocked with the cause written and committed, and the run stopped — the declared-block stop added earlier the same day doing its job on its first real occasion.

It established the minimum by measurement, not by assumption: a table of plugin versions actually run against the new build tool, showing which two fail on removed APIs, which is the first that passes, and which later ones also pass and were therefore not taken. "The lowest line that works" is a claim that can be checked, unlike "the recommended version".

It declared every version that moved — 105 resolved dependencies, all dictated by the platform BOM, none chosen. One artefact needed an explicit version because the newer BOM stopped managing it; the report states the version is unchanged from what the previous BOM resolved, and that was verified independently afterwards.

It caught that regenerating the build wrapper overwrites a locale block a previous lot had added, and re-applied it. Without that, resource processing breaks on two files whose names carry decomposed accents — a trap that cost an hour of diagnosis the first time a human hit it.

Lessons

• A green build is not a green lot, and this run is the proof. Every argument for putting a behavioural net underneath a modernisation rests on this being real rather than theoretical. It cost $1.63 to make it real.
• A blocked lot with a committed cause is a deliverable. The next lot writes itself from it: the two failures are application-configuration changes, which is a different intent and therefore a different lot.
• Nine-second gates are suspicious and worth reading: here it was honest (the application died immediately), but the same shape would appear if the gate never reached the application at all.

2026-07-28 — first lot on a legacy Java/Spring target: gate green, lot blocked with cause, and a defect found in OUR net (run 019fa7b5)

• Status: the bot behaved exactly as designed, including by refusing. The lot did not complete, and that outcome is the run's main success.
• Versions: bot v0.1.0 · iterion built from HEAD · --sandbox none · only_lot=L1, max_passes=4, --max-duration 3h.
• Campaign: 24 min, 38 794 tokens, $0.97.

What the lot produced

The build wrapper rose two major versions on the unchanged runtime. Both exit_gate commands exited 0, the application built and packaged, and not one line changed under the oracle's reference directory — the campaign checked that itself before the gate did.

It then stopped short of the version the programme targets, with a precise reason: the next major makes strict task validation blocking and rejects a task contributed by the framework plugin, so going further would mean raising the plugin, therefore the dependency BOM, therefore application dependency versions — which the lot's stated intent excludes. It marked the lot blocked with the cause committed, rather than overflowing.

That is the "stay inside the lot" rule holding under pressure, with no human arbitrating. A green lot that had quietly bumped a dozen libraries would have been worth less than this refusal.

The finding that mattered: a hole in the behavioural net

The gate went red on collateral: 3, and the campaign proved the cause was older than the lot and external to it — by replaying on the pre-lot tree, wrapper restored and artifact rebuilt, and obtaining a character-identical diff.

One reference rendered a seeded creation date, and the seed used the database's current_date(). The reference had frozen on the day the fixture was first applied; every database seeded on another day diverged permanently, for a reason with no business meaning.

Two guards could not see it, structurally, and that is the part worth carrying forward:

• Stability compares captures to EACH OTHER — A/B on one boot, C after a restart, all the same day. A constant offset from the reference is invariant across all three, so it is invisible by construction.
• The negative control sampled the first six entries. Every later entry was therefore never once confronted with its own reference unless a mutant happened to target it. A reference could be stale and nothing would say so.

It surfaced only as collateral noise under three unrelated mutants, one step away from being attributed to a lot that had not caused it. Fixed: the negative control now covers the whole corpus and names the entry directly. Cost: one extra pass of plain requests against an application already up.

The general lesson. A net records the world as it was on the day it was recorded. Replaying it on a different day, from a different path, is a distinct verification mode from running it again — and it is the only one that catches this class. It has now caught six defects in a row that no additional run would have found.

The re-baseline, and what was deliberately not done

The fixture now pins the seeded dates, alongside the password reset and publication-window widening it already performed. One reference was re-recorded in consequence — two lines — with the cause written in REBASELINE.md and the full counter-test replayed after: stable, noop_silent, revert_clean, 7/7 detected, collateral 0, exit 0.

The column was not canonicalised away. A creation date rendered in a back-office list is business data, and neutralising it would blind the net to a genuine regression on it. What had to be removed is that its value depended on the day the environment was built. That distinction is the whole of the canonicalization discipline in one case.

Bot fixes this run produced

• The repair loop sent the campaign back against a wall it had already described and proven external to its lot. A declared block now stops the run. The asymmetry is what makes it safe: a status claiming completion could be self-serving and stays worthless, while one claiming failure cannot cheat its way to green. It is read from the contract, which the campaign must commit to declare it, so the claim is reviewable in a diff.
• The bundle pinned a package version that does not exist, and named the wrong implementation — the script uses the Go yq's flag while nixpkgs yq is the Python wrapper. Install failed, the declared packages never reached PATH, and the run survived only because the operator had yq on the host.

Lessons for next run

• L1 is blocked by design now. Reaching the programme's target version needs a lot that explicitly allows the framework plugin to move — i.e. a different lot, with its own intent and its own gate, which is exactly what the lot granularity is for.
• Budget by duration; the campaign cost under a dollar and the gate under three minutes. This bot is cheap compared to the net it verifies against.