expected_changes.toml promises the harness will fail on a Ph. D. regression; it won't

tools/differential/expected_changes.toml:301-311 deliberately leaves trailing Ph. D. healing unclassified, and says why:

Adding a suppression rule for it would risk masking a real regression in this exact shape, so it is intentionally left unclassified: if it ever starts diffing, the harness must fail.

It does not. Reproduced — a three-name corpus containing John Smith, Jr. Ph. D.:

corpus: 296 names; intentional diffs: 103; unexplained: 0
## fix(comma-family) lone post-comma piece routes to suffix/title, not first (8)
  'John Smith, Jr. Ph. D.'

The divergence is real — 1.4.0 gives suffix "Ph. D., Jr.", 2.x gives "Jr. Ph. D." — and it is absorbed silently by a rule whose prose is about something else entirely.

Why

fix(comma-family) is declared as:

name_regex = ","
fields = ["first", "title", "suffix"]

Any suffix-only diff on a comma-bearing name is a subset of that, so the rule claims it. classify() (compare.py:40-48) returns the first matching rule, and compare.py:69 sorts on only two tiers:

rules.sort(key=lambda r: not isinstance(r.get("name_regex"), str))

— name_regex rules ahead of fields-only rules, stable within tier. Twelve of the thirteen rules are in the name_regex tier, so for almost every rule precedence is file order.

That is what makes this hard to fix locally. A narrower rule for the Ph. D. shape could only win by being written earlier in the file — which makes file position load-bearing again, the exact thing the sort's own comment says it exists to prevent ("sort is stable, so rules within a tier keep the order they were written").

Scope, not just this shape

The problem is structural. , and / are among the declared name_regex values, so any rule sitting above a narrower one can absorb its cases. Nothing detects it: a rule that swallows more than it describes produces unexplained: 0 and looks like a pass. The Ph. D. case surfaced only because a reviewer probed the file's own written promise.

Candidate directions

Give classify() a real specificity order — e.g. rank by name_regex length or by an explicit priority key — so a narrow rule beats a broad one regardless of position.
Report the runner-up. If two rules match, say so. Silent absorption is the failure mode; visibility may be enough without reordering anything.
Let a rule declare itself non-absorbing — a flag meaning "only claim a diff no other rule matches."
Assert the promise directly — a test that adds the Ph. D. shape to a scratch corpus and requires it to come back unexplained. Narrow, but it makes this comment honest without a harness redesign.

Interim

The divergence itself is now pinned by suffix_comma_split_phd_after_another_suffix in tests/v2/cases.py (classification fix(credential-pair-order)), and the toml comment records that the case table — not the harness — is what guards it. So nothing is unguarded today; what is wrong is the file's claim about its own behavior, and the general absorption hazard behind it.

Found during the #319 review (PR #327).

Re-verified against master at d910709 (2026-08-09): still valid, and the exposure is larger than filed.

Escalation: the shape is in the shipped corpus now

The original repro used a scratch three-name corpus. It no longer needs one — 'John Smith, Ph. D.' is in both corpus.jsonl and corpus_issues.jsonl today, and 'John Smith Ph. D.' is in corpus.jsonl.

Running master's classify() against master's ledger:

'John Smith, Jr. Ph. D.'  + a {suffix} diff  ->  fix(comma-family)      (the original repro)
'John Smith, Ph. D.'      + a {suffix} diff  ->  fix(comma-family)      (now shipped)
'John Smith Ph. D.'       + a {suffix} diff  ->  fix(suffix-routing)    (now shipped)

So a real regression on the exact shape the ledger comment protects would be absorbed by a rule whose prose is about lone post-comma pieces, and the run would exit 0. The promise at expected_since_1.4.0.toml:393 — "if it ever starts diffing, the harness must fail" — is false for names the harness actually parses on every run.

The mechanism is unchanged

_sorted_rules (compare.py:102) still keys on exactly one boolean — has a name_regex or not — and classify (compare.py:462) still returns the first match. Within the name_regex tier, file order still settles everything, which is what the issue identified.

The recent ledger guards do NOT cover this

#333 and #350 added guards to tests/v2/test_ledger_guards.py that bound what each rule may claim, including _CORPUS_CLAIMS. Worth stating plainly that none of them helps here, because the resemblance is misleading:

_CORPUS_CLAIMS records the number of corpus names each rule's regex reaches — 236 of 751 for fix(comma-family), and 'John Smith, Ph. D.' is already among them. A regression that makes an already-reached name start diffing changes that number by zero.

Those guards bound what a rule can claim. This issue is about what a rule does claim in a run, and nothing measures that.

Stale references in the description

Three things moved after filing; the analysis is unaffected:

tools/differential/expected_changes.toml was renamed to expected_since_1.4.0.toml (48819a7). The cited lines 301-311 are now ~385-401.
fields moved to the Role vocabulary, so fix(comma-family) now reads fields = ["given", "title", "suffix"], not ["first", ...].
compare.py:40-48 / :69 are now :462 / :102.

Half the shape did get a rule

The leading case is handled: fix(leading-credential) a split 'Ph. D.' before the name stays one unit now exists, deliberately anchored to the start, with a comment saying "widening this regex would mask a regression". It is the trailing case that remains unguarded by the harness.

The issue's "Interim" note still holds — suffix_comma_split_phd_after_another_suffix in tests/v2/cases.py pins the divergence, so nothing is unguarded today. What is wrong is the ledger's claim about its own behaviour, plus the general absorption hazard.

On the four candidate directions

The fourth ("assert the promise directly") is now considerably cheaper than when filed: tests/v2/test_ledger_guards.py already loads every ledger and the full corpus as module-level constants, so a test asserting that no rule claims the trailing-Ph. D. shape is a few lines and needs no harness redesign. It does not fix the structural hazard, but it would make this comment honest.

Replace a tautological guard with a recorded one

Say which way an over-wide exclusion fails, and drop "for a year"

Merge master (#370) into the ledger-exclusions branch

docs(rules): topic-review amendments from the ledger-arc session (pri…

Recognize the Japanese maiden marker 旧姓 (derek73#309)

Let classify() refuse a shape the ledger excludes

Assert in CI that no rule claims an excluded shape

Navigation Menu

Repository navigation

`expected_changes.toml` promises the harness will fail on a `Ph. D.` regression; it won't #328

Why

Scope, not just this shape

Candidate directions

Interim

Escalation: the shape is in the shipped corpus now

The mechanism is unchanged

The recent ledger guards do NOT cover this

Stale references in the description

Half the shape did get a rule

On the four candidate directions

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

expected_changes.toml promises the harness will fail on a Ph. D. regression; it won't #328

Description

Why

Scope, not just this shape

Candidate directions

Interim

Activity

derek73 commented on Aug 9, 2026

Escalation: the shape is in the shipped corpus now

The mechanism is unchanged

The recent ledger guards do NOT cover this

Stale references in the description

Half the shape did get a rule

On the four candidate directions

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions

`expected_changes.toml` promises the harness will fail on a `Ph. D.` regression; it won't #328