Evidence
Controlled Experiment 2 — boundary authority hardening
Whether the blind spots Experiment 1 found can be closed mechanically, and what closing them does and does not change.
What this was
- Status
- Complete
- Repository
- kemiller2002/helix-note-application
- Branch
- experiment/boundary-authority-v1
- Baseline tag
- state-programming-boundary-authority-pre-v1
- Baseline commit
- c38a1b0
- Ending commit
- 97331562b718
- Executed by
- The same human-supervised session pattern as Experiment 1, acting as its own instrument.
Objective
Build one hardening mechanism for each blind spot Experiment 1 found, then run the same class of mutation again — a new ObservedValue case, ProcedureObserved — and measure what changed.
Hypothesis
Targeted hardening raises the proportion of required sites a machine can find. It was an open question whether it also reduces how many sites there are.
Method
Five mechanisms were designed against Experiment 1's specific blind spots: a SQL CHECK-constraint agreement test; a single-source wire discriminator; a closed-union exhaustive dispatch replacing string dispatch; a field-vocabulary agreement test; and a test fixture forced to a compile error when a case is missing. The hardened state was tagged, the trial mutation applied, and discovery classified as before. A deliberate out-of-scope mutation then re-created Experiment 1's silent-omission shape against the hardened architecture.
Recorded metrics
66.7% to 100%
Mechanical discovery rate after hardening
Self-reported
- Measures
- The share of required sites reported by a machine rather than found by looking, before and after five targeted hardening mechanisms.
- Collected by
- Same classification method as Experiment 1, applied to a new case on the hardened architecture.
- Does not show
- 100% is measured against the specific blind spots these five mechanisms were built to close. Experiment 3 showed the figure does not transfer to a tier the hardening does not cover, where it fell well below this.
Sourcekemiller2002/state-directed-engineering / research/evidence/EV-HN-2026-0004--experiment-2-boundary-authority.md — Key figures4.0, unchanged
Boundary files per semantic decision, after hardening
Instrumented
- Measures
- Boundary Change Amplification measured again on the hardened architecture, against the identical figure from the unhardened one.
- Collected by
- Counted from the committed diff, the same way as Experiment 1.
- Does not show
- This is the result that refutes a natural expectation. Hardening changed how required work was found and did not change how much of it there was. Do not cite hardening as reducing boundary work.
Sourcekemiller2002/state-directed-engineering / research/evidence/EV-HN-2026-0004--experiment-2-boundary-authority.md — Key figures0 of 4
Mechanical checks that caught a present-but-inert boundary arm
Instrumented
- Measures
- Of the compiler, the architecture check, the schema-agreement test and the field-vocabulary agreement test, how many detected a boundary arm that existed but did nothing.
- Collected by
- A deliberate out-of-scope mutation re-creating Experiment 1's failure shape against the hardened architecture, with all four checks run against it.
- Does not show
- Exhaustiveness checking guarantees the arm exists. It says nothing about whether the arm does anything. A live integration test detected it; nothing else did.
Sourcekemiller2002/state-directed-engineering / research/evidence/EV-HN-2026-0004--experiment-2-boundary-authority.md — Key figuresFindings
- Mechanical Discovery Rate rose from 66.7% to 100% for the targeted blind spots: every required site produced a machine-generated error.
- One hardening mechanism removed a required site outright rather than mechanising its discovery. Making the wire discriminator single-sourced left nothing at that site to find. This is the first direct evidence in the programme that hardening can reduce the work rather than only redirect how it is found.
- Five small, coherent mechanisms were sufficient. No large architectural change was needed.
What this refuted
- Boundary Change Amplification did not move. It was 4.0 boundary files and 7.0 mechanical edits per semantic decision both before and after hardening. Hardening changed how the required work was found; it did not change how much there was.
- Exhaustiveness checking was refuted as sufficient for behavioural correctness. A structurally-present but semantically-inert boundary arm passed the compiler, the architecture check, the schema-agreement test and the field-vocabulary agreement test, all clean. Only a live boundary-path integration test detected it.
Anomalies, failures and confounders
- This trial's mutation never required work at the API tier — a scope gap Experiment 3 deliberately extended into, and the dominant reason Experiment 3's numbers look worse than this one's.
- The completion tag hit the same HTTP 403 tag-push restriction as Experiment 1 and exists locally only.
Limitations
- A single trial and a single new case, executed by the same session that designed the hardening. An author measuring the effect of their own mechanism is not an independent test of it.
- 100% is a mechanical discovery rate against the specific blind spots these five mechanisms were built to close. It is not a general claim about all boundary work, and Experiment 3 showed it does not transfer to an unguarded tier.
- Nothing here establishes that the hardening was worth its construction cost; the experiment measured discovery, not cost.
Interpretation
This experiment succeeded at what it set out to do and, in the same run, refuted the thing
everyone expected it to confirm.
The success is real: five small mechanisms, none of them an architectural upheaval, took the
targeted blind spots from two thirds mechanically discoverable to all of them. One of the five
did something better still — making the wire discriminator single-sourced removed the site
rather than mechanising its discovery. That is the only instance in this whole evidence base
of construction optimisation rather than detection optimisation, and it is the one worth
imitating.
The refutation is that boundary amplification did not move at all. Four boundary files per
semantic decision before hardening, four after. This is the measurement that separates two
claims people habitually merge: making required work easier to find, and making less work
required. Only the first happened.
And then Phase 20, which is the reason this programme is careful. An out-of-scope mutation
deliberately re-created Experiment 1's failure shape against the newly hardened architecture.
Exhaustive dispatch guaranteed the arm existed. Nothing guaranteed the arm did anything, and
nothing mechanical noticed that it did not.
Read together: hardening works, for the thing hardening does. The gap it leaves is not a gap
in the hardening — it is a different class of failure that needs a different class of
mechanism.