v1.10.0 — Variant C figure_ids hard constraint + normalize BS3+BS4

Highlights since v1.9.2

• Variant C — figure_ids hard constraint ships as v1.10's headline change. Picks the winner of a 3-variant × 9-paper × 3-audit-cycle test (33 LLM runs across 3 git worktrees).
• normalize_ocr_latex BS3+BS4 lifts verifier accuracy across all variants by collapsing LaTeX escape sequences (\% → %) and Unicode super/subscripts (³ → 3).
• 3-cycle audit committee (each cycle = 3 specialist auditors in parallel) caught 5 bugs in variant C, all fixed before ship.
• 280/280 tests pass (+25 from v1.9.2's 255).

Full report: docs/v1_10_variant_comparison.md · External reference survey: docs/v1_10_external_reference.md

Variant C — figure_ids hard constraint

Before v1.10, figure embedding in the rendered output depended on the LLM voluntarily writing Fig. N literally in the prose. On figure-rich papers (ML papers with ≥10 figs, materials reviews with ≥20 figs), the LLM cited <30% of available figures, and the s09 binding silently dropped the rest.

Variant C makes figure citation a schema-level + verifier-level constraint:

• GroundedClaim.figure_ids: list[str] — schema field tells the binding which figures the claim cites.
• Prompt rule — for each figure listed in section_figures, the LLM must emit a claim with figure_ids=["Fig. N"] AND include the literal Fig. N / 图N in text.
• Verify advisory — accepted claims whose figure_ids don't appear literally in text get recorded as figure_hint_unmet for audit visibility (claim still kept).
• Figure-retry pass — when ≥50% of available figures aren't mentioned post-verify, one strengthened LLM call adds them. Three swap guards (parity with retry-when-short β#3): more figures AND ≥1 verifier-accepted claim AND required-mention coverage not regressed.
• Env-gated whitelist — LAZY_PAPER_FIGURE_ID_WHITELIST=1 strips unknown fig_ids; default OFF because cycle 1+2 evidence (reading original PDFs) shows "unknown" fig_ids are usually s04_figures OCR-vs-paper-actual numbering misalignment, NOT LLM hallucination.

Validation — 33 LLM runs across 3 variants

M2 figure embed ratio (true distinct-figure count, not panels)

Paper	available	A (env tuning)	B (cap tiering)	C (figure_ids)
ali2025_flash	26	5 (19%)	4 (15%)	26 (100%)
he2023	8	4 (50%)	7 (88%)	8 (100%)
meng2024	7	4 (57%)	5 (71%)	7 (100%)
hif_1 (Adv Mat 62p)	20	2 (10%)	4 (20%)	20 (100%)
hif_2 (DALL-E 2)	17	3 (18%)	5 (29%)	17 (100%)

M1 zero-variance probe (meng2024 × 3 runs)

Variant	per run	mean	stdev
baseline (v1.9.2)	10117 / 12742 / 13666	12175	1503
A	12819 / 12065 / 12521	12468	310
B	13740 / 13012 / 12002	12918	713
C	12703 / 13525 / 12849	13026	358

All 3 variants stay well below the spec §7 floor (baseline stdev = 1503).

M4 TestCase scores

TestCase (max)	baseline	A	B	C
meng2024 T1 ch01 benchmark recovery (17)	9 / 9 / 9 (stdev 0)	5 / 9 / 9 (stdev 1.88)	5 / 17 / 15 (stdev 5.25)	9 / 9 / 9 (stdev 0)
ali2025_flash T4 ch14 comparison depth (5)	4 / 3	4	3	5
chai2026 T6 ch01 (4)	4 / 4	3	3	4
yang2025 T2 ch01 (3)	3 / 3	3	3	3
meng2024 T3 ch10 (5)	3 / 3 / 5	4 / 5 / 4	5 / 4 / 4	4 / 4 / 4

Variant C is the only variant that preserves the meng2024 T1 = 9/9/9 zero-variance achievement of v1.9.x AND breaks baseline on ali2025_flash T4 (4 → 5).

Why not Variant A or B

• A (env tuning) introduces M4 variance (stdev 1.88 on T1) and Auditor 1 caught a content-quality error in cycle 2 (called PbZrO₃ "lead-free" — chemical-naming mistake). The env tuning is real lift; it'll be recommended in .env.example as a LAZY_PAPER_MIN_SECTION_CHARS=1200 LAZY_PAPER_BEST_OF_N=3 combo, not as a separate variant flag.
• B (cap tiering) had the worst M4 variance (stdev 5.25 — can reach 17/17 peak but unstable) AND lowered M3 coverage rate (meng2024: A 71% vs B 64%; gaur2022: A 53% vs B 34%). cap=12 spreads LLM attention too thin. Deferred to v1.11 for redesign as dynamic cap = min(comparator_count, 10).

normalize_ocr_latex BS3+BS4

Auditor 2 (cycle 1) inventoried 4 normalize blindspots causing 41-74 verifier false-rejects per s08 run. Cycle 2 broke them into BS1-4 with asymmetric fix difficulty:

• BS3 — LaTeX escape sequences: \%, \&, \_, \^, \$ lose their leading backslash so they match the LLM's unescaped quote.
• BS4 — Unicode super/subscript folding: NFKD decomposes ³ → 3, ₂ → 2. Greek letters (α/β/π) NOT decomposed. NFKD doesn't fold U+2212 / U+2013 / U+2014 to ASCII -, so an explicit _UNICODE_DASH substitution follows.
• BS1+BS2 (letter-spaced subscript) deferred to v1.11 due to inherent OCR↔LLM asymmetry.

11 unit tests in stages/_common/tests/test_normalize.py lock in the new behavior.

3 audit cycles — caught & fixed bugs in variant C

Cycle	Auditor	Bug	Fix commit
1	Auditor 3	figure_ids advisory ran on rejected claims (double-report)	2e30e1c
1.5	Auditor 3.2	figure-retry missing C-1 coverage guard	0cc056d
1.5	Auditor 3.2	figure-retry missing C-2 min-accepted guard	0cc056d (same)
2	Auditor 2	advisory entries inflated audit log's "verifier_rejects" count	535d035
3	Final ship gate	env-gated whitelist needed for cycle 3 visibility	e3cb9e1

3 cycles ran the audit committee in parallel (each cycle = 3 specialist auditors with non-overlapping scopes, reading original PDFs + code + reference docs).

Side effects

• M2_figures_embedded metric (scripts/collect_variant_metrics.py) was inflating values by counting <img> tags (which includes every panel of a multi-panel figure). Fixed to count <figure> blocks (one per distinct figure). Cycle 1 caught this — earlier "17/17 = 100%" readings on multi-panel-heavy papers were panel-count artifacts.
• New M2_figures_hallucinated_count field surfaces s04_figures ↔ LLM numbering misalignment (v1.11 #2 candidate fix).
• M4 TestCase scores are deterministic (regex + size + fuzzy substring — no LLM judge), so baseline recheck took 0 LLM cost and confirmed zero drift from v1.9.x scores.

Deferred to v1.11 (ranked by ROI)

#	Item	Complexity
1	normalize_ocr_latex BS1+BS2 (letter-spaced subscript)	M
2	s04_figures caption-aware numbering (fix OCR vs paper number misalign)	M
3	prompt comparator gap — Jiang/Ma et al. systematically missed	M
4	template-vs-paper subject-mismatch graceful degrade	L
5	Variant B redesign: dynamic cap = min(comparator_count, 10)	S
6	real-time LLM cost meter into metrics.yaml (M6)	S
7	DOCX HYPERLINK dead-code fix (thread sources into renderers)	M
8	_merge_drafts 60-char prefix + (author, value) dedup	S
9	6 hardcodes → env vars	S

Release trail (since v1.9.2)

• v1.9.2 — 8-bug fix + 18-paper validation
• v1.10.0 — Variant C ship + normalize BS3+BS4 + 3-cycle audit

Full details: CHANGELOG.md