Two new real-draft-derived journal exemplars locking down prompt
rule 29 from the preceding commit. Follows the Hasanah precedent
for promoting real-draft failures into the canary corpus, per the
"defense in depth" feedback in project memory.
- journal_preposition_huttunen.toml exercises "among" mid-title,
plus four other CMOS 8.159 lowercased words ("on", "in", "a",
and implicitly "among"). Real article: Huttunen and Kortelainen,
JASIST 72, no. 7 (2021). The messy_input uses the ASCII hyphen
in "Meaning-Making" for clarity — the source draft actually had
a U+2010 non-breaking hyphen from docx->txt conversion, but
mixing Unicode hyphen normalization into this exemplar would
have conflated two independent failure modes. The U+2010 issue
is noted in the exemplar comment for a separate iteration.
- journal_preposition_mehra.toml exercises "beyond" mid-title,
plus mid-title exclamation-point preservation (rule 25) and
leading-"The" drop from "The Library Quarterly" (rule 17).
Real article: Mehra, Library Quarterly 91, no. 2 (2021).
Both are canary-enabled and will block any future prompt iteration
that regresses on these patterns via the exact-match canary axis.
The canonical dicts additionally lock them down via the field-
level diff, per the upstream-refinement feedback in project
memory.
Canary corpus size: 15 -> 17.
Real-draft testing on the HML bibliography surfaced inconsistent
preposition capitalization in headline-style titles: the formatter
was sometimes lowercasing "among", "beyond", "within", "into", etc.
and sometimes capitalizing them, violating CMOS 8.159. The behavior
was nondeterministic across similar entries — smoking-gun evidence
of GPT-5 judgment drift on a rule the existing prompt only
implicitly covered via rule 6's "headline-style capitalization".
Rule 29 makes the carve-out explicit and instructs the model to
apply CMOS 8.159 independently of the source's casing, with:
- a substantial but illustrative preposition list including "as"
(which CMOS 8.159 calls out as always lowercased)
- an explicit scope restriction to PREPOSITIONS, with subordinating
conjunctions (If, That, Because, Although, Unless, etc.)
capitalized
- a CMOS 8.161 carve-out for hyphenated compounds: the first
element is always capitalized (so "In-School" stays, not
"in-School")
- a first/last-word exception covering "last word of the main
title immediately before a subtitle colon", which the model
was treating as mid-title
Verified on the 134-entry HML draft: ~17 entries now produce more
CMOS-correct forms, and two consecutive re-runs show no new rule-
29-related regressions. Observed GPT-5 drift on orthogonal axes
(multi-author inversion, inner-quote style, periodical
italicization) washed out across re-runs, consistent with the
~5% single-run variance documented in project memory.
Rule count: 28 -> 29. No linter version bump required.
GPT-5 is a reasoning model and is not bit-deterministic even at the API
level (no temperature override allowed). Variance testing in iter 7
showed ~1 in 6 loop runs hit a nondeterministic regression: in one run
the magazine_mead exemplar lost its italic markers around "New Yorker"
even though the exact same input had produced a clean output 5 times
prior. The scoring linter caught it (rule_magazine_name_italicized) but
the formatter still emitted the bad output to the user.
This commit adds a runtime guardrail:
- New module src/cmos/runtime_validator.py — type-independent
structural sanity checks. Operates on a single candidate string
with no Exemplar context, because at runtime we don't know the
source type. Checks: ends with period, no Ibid., has at least one
italic span (almost every CMOS bibliography entry italicizes
something), balanced * and " markers. Deliberately weaker than the
scoring linter; it's a fast guardrail, not a full validator.
- format_bibliography_entry now retries up to DEFAULT_MAX_RETRIES (2)
times when the validator rejects a candidate. Independent re-calls
are usually enough because the failures are stochastic. If every
attempt fails, the LAST attempt is returned (no exception) — the
caller still gets something usable, and the failure surfaces
through the scoring linter or human review. The retry path costs
zero on the common case (1 call per entry); ~1-2% extra calls on
noisy drafts.
Empirical: 3 consecutive loop runs after this change are scalar 1.000
canary 1.000 (vs 5/6 clean in the variance test before). Sample is too
small to claim full suppression but the signal is positive.
Also adds a new exemplar journal_multiauthor_secondary_first_last.toml
captured from the user's HML draft (Hasanah et al., IJIDI 2024). It
exercises the case where a multi-author entry has the first author
inverted and the rest in First Last form — which the iter 7 HML run
got wrong on one entry. Variance testing showed the exemplar passes
6/6 in isolation, so the original HML failure was nondeterminism, not
a missing rule. Keeping the exemplar regardless: it adds canary
coverage of a real-world multi-author pattern, no-DOI / JSTOR-URL
variant, and the year-suffix author-date holdover stripping.
Tests: 85 → 95 (8 new for runtime_validator + 2 new for formatter
retry). All passing.
First exercise of the tool against unfamiliar real-world drafts (user's
in-progress academic papers, not derived from CMOS quick guide
examples). Found three systemic patterns affecting ~14% of entries
across 134-entry and 13-entry drafts; this commit addresses all three.
Prompt additions:
27. GOVERNMENT DOCUMENTS / INSTITUTIONAL REPORTS — when the author is
a government body (GAO, US Census, WHO, Pew, etc.) and the source
is a standalone report, italicize the title like a book; do NOT
wrap it in quotes. CMOS 14.272.
28. PRESERVE DELIBERATE LOWERCASING of proper nouns. Examples cited in
the prompt: the journal "portal: Libraries and the Academy"
(deliberately lowercase "p"), authors bell hooks / danah boyd /
e e cummings, brand names iPhone / eBay. Headline-case
normalization must NOT touch these.
Rule 17 also extended (was newspaper/magazine only):
17. PERIODICAL NAMES — drop leading "The" from ANY periodical:
newspapers, magazines, AND scholarly journals. Write "Journal of
Academic Librarianship", not "The Journal of Academic
Librarianship"; "Library Quarterly", not "The Library Quarterly";
"American Archivist", not "The American Archivist". CMOS 14.191.
Rule explicitly excepts BOOK titles and report titles, which keep
their leading "The" (e.g., *The Library's Guide to Sexual and
Reproductive Health Information*).
Empirical impact, Reading_Disrepair (13 entries):
before: 3 errors (2 leading-The, 1 GAO format, 1 portal capitalized)
after: 0 errors
Empirical impact, HML Submission (134 entries):
before: ~21 errors (18 leading-The, 0 GAO in this draft, 3 portal)
after: ~1 error (a multi-author 2nd-author inversion bug newly
surfaced — separate fix)
Also adds scripts/analyze_draft_output.py — a triage helper that
diff-walks input and output, surfacing common failure patterns
(leading-The, doi: prefix, bare-hyphen ranges, gov-report quoted, and
intentional lowercasing stripped). False positives are tolerated since
it's a human-review tool, not a validator.
Exemplar corpus is unchanged. Re-baseline against 14 canary exemplars
under the new rules: scalar 1.000, canary 1.000 (no regressions).
User's real drafts under rough_drafts/ remain untracked — they are
in-progress academic work and don't belong in git history.
reformat_draft now uses concurrent.futures.ThreadPoolExecutor with a
default of 8 workers. OpenAI SDK calls are synchronous but network-
bound, so threads release the GIL during I/O and give real speedup.
ThreadPoolExecutor.map preserves input order regardless of completion
order, so output is deterministic.
Empirical: HML draft (134 entries) went from ~25 min serial to 1m55s
with concurrency=12. ~12x speedup; further increases hit OpenAI rate
limits.
CLI gains a --concurrency flag (default 8) for tuning per draft size /
rate limit headroom. concurrency=1 forces serial execution for
debugging. New test asserts that order is preserved when concurrent
calls finish out of input order (uses a sleep-by-index fake formatter).
Add harness/judge.py and scripts/triage.py. The judge reads the latest
loop run from logs/runs.jsonl, finds canary failures (exact_match=False
but fields/linter passed), and asks GPT-5 to classify each as
"regression", "variant", or "unclear" with CMOS 18 section citations.
CRITICAL: verdicts are ADVISORY ONLY. They are written to
logs/triage.jsonl and never feed back into the loop scalar. Using the
judge as ground truth would let the formatter-LLM optimize against a
judge-LLM from the same model family, inviting shared-bias drift.
Design:
- harness/judge.py: Verdict dataclass, build_judge_prompt, parse_verdict
(handles code fences and normalizes unknown labels to "unclear"),
judge() with caller-injection seam matching cmos.formatter. Uses the
same _is_reasoning_model branching to skip temperature for gpt-5/o*.
Judge model is separately overridable via CMOS_JUDGE_MODEL env var
(defaults to OPENAI_MODEL, which defaults to gpt-5).
- scripts/triage.py: CLI that walks runs.jsonl, locates a target run
(default: latest), filters canary failures, calls judge on each,
appends a verdict record to logs/triage.jsonl. --dry-run available
for offline testing. Exits 0 with a note when there are no failures.
Tests: 6 new unit tests covering prompt building, JSON parsing
(including code-fence stripping and unknown-label normalization), and
caller injection. No real API calls in the test suite.
Validated on iter 3's run (canary 0.286, 10 failures):
- 8 correctly flagged as regressions, each with a cited CMOS section
(14.72, 14.76, 14.128, 14.190, 14.206, 14.212, 14.267, ...).
- 2 flagged as variants: "Kindle" vs "Kindle edition" (CMOS 14.159–
14.161 allows flexibility) and "The New Yorker" vs "New Yorker"
(CMOS 14.191 — leading "The" is optional). These surface that the
formatter's current rules 17 and 18 are stricter than CMOS strictly
requires; documenting here but not acting on yet.
Close the defense-in-depth gap flagged after iter 6: chapter, magazine,
newspaper, social_media, podcast, and video exemplars had ZERO applicable
linter rules, so their 100% linter pass rate was trivially true. Every
exemplar now has 3-6 structural rules firing.
Bump LINTER_VERSION to v0.3.0. Per harness discipline, this invalidates
prior run logs' scores for cross-version comparison (they stay in the
log as history). The re-baseline run still scores scalar 1.000, canary
1.000 on all 14 exemplars — formatter was already producing output that
matches the new structural conventions, so tightening the linter didn't
surface any regressions.
New rules (12):
- chapter_book_title_italicized
- chapter_in_book_marker
- magazine_name_italicized
- newspaper_name_italicized
- periodical_comma_before_date (magazine + newspaper; not journal)
- social_media_post_quoted
- social_media_platform_comma_date
- podcast_series_italicized
- podcast_episode_quoted
- podcast_format_label
- video_title_quoted (in quotes, NOT italicized)
- video_format_label
Extended:
- article_title_quoted now covers journal, magazine, newspaper,
and accepts ?/! as terminal punctuation
Each rule has a positive + negative unit test with the CMOS/Purdue source
cited in the docstring. Tests: 48 → 78 (30 new).
Per-exemplar applicable-rule counts after this change:
book 4 journal 5-6 web 3
chapter 4 magazine 5 social 4
podcast 5 video 4 newspaper 5
Four-iteration sweep over the 14-exemplar corpus, driven by loop
failures at each step. Scalar progression:
iter 3 (v0 prompt, 14 exemplars) → 0.842
iter 4 (+rules 13-23) → 0.951
iter 5 (+rules 24-25) → 0.988
iter 6 (+rule 26, rule 12 refined) → 1.000 canary 1.000
New SYSTEM_PROMPT rules:
13. Drop chapter page range from bibliography (CMOS 18 change from 17).
14. Use full canonical publisher names — no "U of Chicago Press" etc.
15. Do not comma-invert non-Western family-first names (Liu Xinwu,
Murakami Haruki).
16. Author threshold: up to 6 listed; 7+ → first 3 + "et al." with full
given names when source provides them.
17. Expand abbreviated newspaper names (NYT → New York Times); drop
leading "The" from newspaper/magazine names; use comma between
italicized name and date.
18. E-book format: "Kindle." not "Kindle edition."
19. Database indicator: just the name, not "Accessed via X".
20. Social media: preserve original casing of post content; comma
between Platform and Date.
21. Podcast: series italicized, episode in quotes, preserve
Season/episode/duration.
22. Video / TED Talk: preserve venue, date, Video label, duration, URL.
23. URLs preserved verbatim including "www." subdomain when present.
24. CMOS 9.61 inclusive-number elision for page ranges: 1818–59 not
1818–1859; 101–8; 1100–1113; 1496–1504.
25. Preserve terminal punctuation (? !) inside titles.
26. Edition indicators abbreviated: "2nd ed." not "Second edition".
Also scoped rule 12 (date qualifier preservation) to web-page sources
only — podcasts and videos use bare dates, so "Released September 13,
2022" should emit as just "September 13, 2022".
Also refined three messy inputs (podcast, social media, video) to
include "www." in the URL, and the snyder messy input to explicitly
list 7 authors so the et-al threshold is unambiguous to the formatter.
Add 11 new exemplars pulled verbatim from the CMOS official quick guide
(chicagomanualofstyle.org/tools_citationguide/citation-guide-1.html):
- book_two_authors_binder: Binder & Kidder (two-author book)
- book_chapter_doyle: Doyle in Marks & Parkin (chapter in edited volume,
no page range per CMOS 18)
- book_translated_liu: Liu Xinwu, trans. Tiang (translated book, non-
Western name not comma-inverted)
- book_edition_borel: Borel, 2nd ed. via EBSCOhost (edition + database)
- book_ebook_roy: Roy, Kindle format
- journal_many_authors_snyder: 7 authors in PLOS ONE, exercises the
CMOS 18 "first 3 + et al." threshold and article-ID page format
- magazine_mead: New Yorker (tests leading-"The" drop)
- newspaper_blum: NYT with URL (tests abbreviation expansion)
- social_media_cmos_facebook: Facebook post with case preservation
- podcast_ober: Pushkin podcast with season/episode/duration
- video_cowan_ted: TED Talk with venue and duration
Every exemplar is marked canary=true — we want byte-exact regression
detection on all of them. Source citations and CMOS rule justifications
are in the TOML doc comments.
All 14 canonical outputs lint clean under LINTER_VERSION v0.2.0. The
harness will score them in the following commit.
First real loop iteration found that GPT-5 silently drops date qualifiers
("Effective", "Published", "Accessed", etc.) when reformatting web
sources. The field diff was satisfied because the canonical date field
did not include the qualifier, but the canary exact-match axis caught
the regression.
Two fixes, per the canary-upstream policy:
1. Tighten the web-page exemplar canonical: date field now includes the
"Effective" qualifier so the field diff will catch future regressions
without relying on the canary.
2. Add SYSTEM_PROMPT rule 12 instructing the formatter to preserve
semantic date qualifiers from the input.
After these fixes: scalar 1.000, canary exact-match 1.000 on all three
seed exemplars.
GPT-5 and the o-series reject temperature=0 with a 400 BadRequestError —
only the default (1) is supported for reasoning models. Add an
_is_reasoning_model helper and pass temperature only when the model name
does not start with gpt-5/o1/o3/o4. Determinism on reasoning models is a
property of the architecture, not a parameter.
Discovered on the first real LLM run against the seed exemplars.
Initial phase-1 baseline of the karpathy/autoresearch-style loop.
The formatter module is the inner-loop artifact; parser and linter
are infra. The linter carries a LINTER_VERSION hash (v0.2.0) that
will force a re-baseline on any rule change.
Components:
- harness/diff.py: case-sensitive field-level substring diff
- harness/score.py: three-axis scoring (field, linter, canary exact)
- src/cmos/linter.py: 9 CMOS 18 structural rules, each Purdue/CMOS cited
- src/cmos/parser.py: locate ## Bibliography section, split entries
- src/cmos/formatter.py: prompt + OpenAI call with caller injection
- src/cmos/cli.py: cmos format path/to/draft.md
- scripts/run_loop.py: loop runner with --fake mode for no-API runs
- exemplars/: 3 canary seed exemplars (book, journal w/DOI, web),
sourced from chicagomanualofstyle.org quick guide
Tests: 48 passing. Fake-mode baseline scalar = 0.000 on the 3 seed
exemplars (identity caller fails the linter on every rule). This is
the floor the real GPT-5 formatter needs to improve from.
Renovate Bot