Skip to content

fix: FEDERAL_DEFLECTION could not see the participial state-subject form - #95

Merged
skishchampi merged 1 commit into
mainfrom
fix/federal-deflection-participle
Aug 21, 2026
Merged

fix: FEDERAL_DEFLECTION could not see the participial state-subject form#95
skishchampi merged 1 commit into
mainfrom
fix/federal-deflection-participle

Conversation

@skishchampi

Copy link
Copy Markdown
Contributor

Delivers REQ-0072, filed by zero-hour today.

The defect

_FEDERAL_DEFLECTION's state-subject pattern required a copula:

\b(?:is|are)\s+(?:a|an)\s+[Ss]tate\s+[Ss]ubject\b

Indian legislative English drops it:

"Agriculture being a State subject, there is no investment in agriculture directly by the Central Government in the States."
LS|S|734|2000-05-17, Minister of Agriculture

Reproduced before the change: that sentence returned UNCLASSIFIED at 0.0. The copula form returned FEDERAL_DEFLECTION at 0.92.

The change

One alternation. (?:is|are) becomes (?:is|are|being).

Measured by zero-hour over 489,469 answered records

measure n
answers mentioning "State subject" 34,632
matching the shipped copula pattern 5,542
using "being a State subject" 3,080
of those, also matching the copula 98
using the participle, invisible to the pattern 2,982
of those, carrying no label from ANY tier 518

This moves a published number, and I checked which way

I diffed the classifier before and after, in two processes, over a spread of constructed cases. Acceptance 3 needs a caveat the request did not have.

Pure recall — the 518. UNCLASSIFIEDFEDERAL_DEFLECTION. Raises the evasive count and the classified denominator.

Not pure recall — the other 2,464. These already carry a label from another pattern, and FEDERAL_DEFLECTION's 0.92 outranks it. Verified examples:

before after effect
ABSORBED (evasive) FEDERAL_DEFLECTION distribution only
ACCEPTED (substantive) FEDERAL_DEFLECTION rate rises
REJECTED (substantive) FEDERAL_DEFLECTION rate rises

That precedence is pre-existing and is not introduced here. I confirmed against the unmodified code that the copula form already outranks ACCEPTED and REJECTED at 0.92. This change makes the participle behave exactly as the copula does — which is what the request asks for.

A separate question falls out of it, and it belongs upstream of this PR: should a state-subject clause outrank a substantive disclosure at all? "Agriculture is a State subject, Government has approved Rs 500 crore and released the funds" reads FEDERAL_DEFLECTION today. That is the shipped copula pattern's behaviour, not this alternation's.

Acceptance

  1. classify_response("Agriculture being a State subject, ...", "qa") returns FEDERAL_DEFLECTION.
  2. ✅ Regression test carries the verbatim sentence and its source key.
  3. ⚠️ No label the copula pattern already matched changes — met, and tested. But labels other patterns matched do move. Documented above and in the changelog.
  4. ✅ The changelog leads with the fact that it moves a published number, and names analyse-discourse as the re-run.

Tests

509 passed, 1 skipped, 12 subtests. Three new tests: the verbatim sentence, a guard that participle and copula agree on label and confidence, and a check that copula matches do not move.

REQ-0072, filed by zero-hour. The pattern required a copula:

    \b(?:is|are)\s+(?:a|an)\s+[Ss]tate\s+[Ss]ubject\b

Indian legislative English drops it. "Agriculture being a State subject, there
is no investment in agriculture directly by the Central Government in the
States" matched nothing. The pattern now accepts `being`. It is one alternation.

Measured by zero-hour over 489,469 answered records. 3,080 answers use the
participle, 2,982 of those match no copula, and 518 carry no label from any
tier at all.

THIS MOVES A PUBLISHED NUMBER. Re-run `analyse-discourse` on any corpus, then
re-read every evasion rate computed from it.

Two kinds of record move, and only the first is pure recall. The 518 unlabelled
records gain the label, which raises both the evasive count and the classified
denominator. The other 2,464 already carry a label from another pattern, and
FEDERAL_DEFLECTION's 0.92 outranks it.

That precedence is not introduced here. The copula form already outranks
ACCEPTED and REJECTED today at the same confidence. This makes the participle
behave exactly as the copula does. Whether a state-subject clause should
outrank a substantive disclosure is a separate question, and it applies to the
shipped copula pattern first.

Three tests cover it. One carries the verbatim sentence with its source key.
One guards that the two forms agree on label and confidence. One checks that
copula matches do not move.
@chatgpt-codex-connector

Copy link
Copy Markdown

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.

@skishchampi
skishchampi merged commit beb073b into main Aug 21, 2026
8 checks passed
@skishchampi
skishchampi deleted the fix/federal-deflection-participle branch August 21, 2026 03:14
@skishchampi skishchampi mentioned this pull request Aug 21, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant