Imo Abstract Polish

Phase 1: Move-by-Move Solver

[!CAUTION] Claim Safety Rule: Before generating ANY Move, read ## 3. Claims Inventory in manuscript_semantic_core.md.

• Claims marked WEAK must NOT appear unless softened (e.g., "generalizable" → "potentially generalizable").
• Claims marked MODERATE must use hedging language (e.g., "suggests", "indicates").

[!IMPORTANT] Source Lock: Solver may ONLY read manuscript_semantic_core.md and writing_brief.md. Do NOT read draft_abstract.md during generation. Numbers must match Fact Base precision exactly.

Move Definitions

For each Move below:

1. Generate 4 candidate sentences (Candidate α, β, γ, δ)
2. Run the Micro-Verifier checks listed on all 4
3. Rank candidates by: (a) micro-verifier pass count, then (b) conciseness (fewer words wins ties)
4. Select winner (highest-ranked candidate)
5. Extract Rheme (the new information introduced) and pass it to the next Move as Theme seed
6. Confidence annotation: Rate the winner on 3 axes (0.0–1.0 each):
   • Fact — How certain the facts are correctly represented?
   • Flow — How well does Theme-Rheme chain from the previous Move?
   • Style — How well does this match Nature editorial DNA? Record as [F: X.X | Fl: X.X | S: X.X] in the Move Workbench.

Move M1 — Context (~20 words, 1 sentence)

Role: Set the domain stage. Name the object/field of study.

Solver Input:

• Logic Graph → Context/Gap (first half)
• writing_brief.md → §1 opening pattern (dominant type)
• §5 Constraint 1: Must open object-first

Generate: Write ONE sentence that names the core scientific object or capability being studied.

Micro-Verifier:

Extract: Identify the Rheme (new concept introduced) → pass to M2.

Move M2 — Gap (~20 words, 1 sentence)

Role: Build tension. Auto-detect Gap type from Logic Graph.

Gap Type Detection (read Logic Graph → Gap node):

• Type A — Improvement: An incumbent method/technique exists but has limitations. Use concession-then-limitation ("However, it suffers from..." or "While X enables Y, it remains limited by Z").
• Type B — Exploration: No incumbent exists; the field simply hasn't achieved X yet. Use frontier gap ("Yet no method has been able to..." or "However, [capability X] has remained elusive").

Solver Input:

• Logic Graph → Gap (determines Type A or B)
• M1 Rheme (Theme seed for this sentence)

Generate: Write ONE sentence matching the detected Gap type. Keep it short — a simple "However,..." is preferred over elaborate subordinate clauses.

Micro-Verifier:

Extract: Identify Rheme (the limitation/gap) → pass to M3.

[!TIP] G3 — M1+M2 Merge Option: After generating M2, check if M1 + M2 combined ≤ 25 words. If so, they MAY be merged into one sentence using:

• "X demands/requires Y, yet Z" (context, yet gap)
• "X is essential for Y, but Z" (context, but gap)

This saves ~5–10 words for later Moves. The merged sentence inherits M2's Rheme for M3. If merged, produce ONE candidate pair for the merged version and micro-verify:

• Opening: object-first ✅
• Contains contrast marker ("yet"/"but") ✅
• Word count: ≤25 ✅
• Theme-Rheme: gap concept is the Rheme ✅

Move M3 — Pivot (~25 words, 1 sentence)

Role: The "Here we" sentence — announce what the paper achieves.

[!TIP] G4 — Architecture-First Detection: Read Fact Base for device/architecture count.

• ≥2 distinct architectures (e.g., two cavity designs, two fabrication methods): M3 should prioritize listing them (mechanism specificity) and defer headline numbers to M4. Pattern: "Here we demonstrate X based on Y, realized in two architectures: [arch 1] and [arch 2]."
• ≤1 architecture: Use default outcome-first pattern: "Here we demonstrate X, achieving [number]."

Solver Input:

• Logic Graph → Mechanism + Key Result
• M2 Rheme (Theme seed)
• §5 Constraint 2: Must include "Here we [verb]"
• §5 Constraint 12: Outcome-first (WHAT, not HOW) — unless ≥2 architectures triggers architecture-first
• writing_brief.md → §1 verb frequency (prefer top verb)

Generate: Write ONE sentence starting "Here we [verb]". If architecture-first: list architectures without numbers. Otherwise: lead with achievement.

Micro-Verifier:

Extract: Identify Rheme (the technique/achievement named) → pass to M3b.

Move M3b — Mechanism (~40 words, 1-2 sentences)

Role: Explain HOW the technique works. Densest information.

[!TIP] Merge Option: If M3 + M3b combined ≤ 50 words AND the paper has ≤1 architecture (G4 not triggered), they MAY be merged into a single sentence. Use a participial phrase: "...which [outcome], [mechanism]-ing [details]." If merged, skip M3b's micro-verifier and move directly to M4. If ≥2 architectures (G4 triggered), keep M3 separate (architecture description) and use M3b for mechanism/fabrication details.

Solver Input:

• ★★★ facts from Fact Base (mechanism-related)
• M3 Rheme (Theme seed)
• §5 Constraint 3: Must include ≥1 quantitative result

Generate: Write 1-2 sentences explaining the technical mechanism, citing specific numbers from the Fact Base.

Micro-Verifier:

Extract: Identify Rheme (key technical detail) → pass to M4.

Move M4 — Key Result (~20–35 words, 1 sentence) ⚡ CONDITIONAL

Role: The headline number(s) — concentrated impact.

[!IMPORTANT] Conditional Move: Generate M4 ONLY if there is a standalone quantitative result that was NOT already stated in M3 or M3b. If the main headline number is already in M3, and no additional independent metric justifies its own sentence, SKIP M4 and proceed to M5.

[!TIP] G5 — Number-Packing Option: If there are ≥2 independent headline numbers NOT yet stated (e.g., efficiency for weak coherent + efficiency for single photons + multimode count), M4 MAY pack them using parallel structure:

• "A for X and B for Y"
• "A for X, B for Y, alongside C with D"

When packing, relax the single-claim rule:

• Standalone check → allow up to 3 parallel quantitative claims
• IU count → allow up to 5 (instead of 3)
• Word count → allow up to 35 (instead of 25)

This matches Nature's editorial preference for concentrated impact sentences.

Solver Input:

• ★★★ facts → quantitative result(s) NOT yet mentioned in M3/M3b
• M3b Rheme (Theme seed)
• §5 Constraint 9: Standalone sentence for headline achievement

Generate: Write ONE sentence with quantitative result(s) not yet presented. If ≥2 numbers available, use parallel packing (G5). Otherwise, single-number standalone.

Micro-Verifier:

Extract: Identify Rheme (the specific achievement) → pass to M5.

Move M5 — Application / Secondary Result (~25 words, 1 sentence)

Role: Present the application or a secondary result that broadens the paper's impact.

Application Type Detection (read Fact Base + Logic Graph):

• Type A — Clinical/Biomedical: Paper has disease/patient/diagnostic results → use domain-shift marker "As a medical/clinical application,..." + name specific conditions/organ systems.
• Type B — Materials/Characterization: Paper demonstrates technique on materials, devices, or systems → use domain-shift marker "As a demonstration of [material/device] characterization,..." or "Applied to [material system],..." + name specific materials/substrates.
• Type C — Secondary Result (no application): Paper is purely fundamental with no applied demonstration → present a different ★★★ fact not yet covered. Drop domain-shift marker. Keep Theme chain.

Solver Input:

• ★★★ facts with applied/secondary context
• M4 Rheme or M3b Rheme (Theme seed — depends on whether M4 was skipped)
• §5 Constraint 13: Domain-shift marker (Type A or B only)

Generate: Write ONE sentence. For Type A/B: open with a domain-shift marker. For Type C: open with Theme chain from previous Move.

Micro-Verifier:

Extract: Identify Rheme (the applied/secondary capability) → pass to M6.

Move M6 — Impact (~25 words, 1 sentence)

Role: Close with forward-looking promise. Measured confidence.

Solver Input:

• Logic Graph → Impact + Novelty
• M5 Rheme (Theme seed)
• §5 Constraint 4: Promise closing with forward-looking language
• §5 Constraint 10: Frame as enabling OR supporting (context-dependent)

Generate: Write ONE closing sentence that conveys broad significance with appropriate hedging.

Micro-Verifier:

Move Workbench (tracking table)

After completing all 7 Moves, produce this workbench:

Phase 2: Assembly + Global Verification

Step 2a: Concatenate + Smooth

1. Concatenate all 7 Move winners in order (M1 → M6).
2. Smoothing pass: Read the full assembled text and make ONLY these changes:
   • Fix connective words (add "thereby", "thus", "crucially" if transitions are abrupt)
   • Adjust articles/prepositions for grammatical flow
   • Do NOT change any factual content, numbers, or core vocabulary
   • Do NOT merge or split sentences
   • Do NOT exceed 200 words total
3. Record changes made in the Smoothing Log.

Step 2b: Global 5-Dimension Scoring

Score the assembled abstract on 5 dimensions:

Flow Sub-checks

Theme-Rheme chain verification (from Move Workbench — should already be ≥80%):

Chain Score: __/6. If ≥5 → +1 Flow bonus. If <3 → -2 Flow penalty.

Information density check (IU per sentence):

Density shape: [bell/flat/spiky/front-loaded/back-loaded]. Target: bell.

Hedge density check: Count hedges in second half (M4–M6). Target: ≥1 hedge. If 0 → deduct 1 from Impact.

Step 2c: Source Lock Verification

W4 — 5-gram overlap check against draft_abstract.md:

If ≥6 words shared → deduct 1 from Novelty. If ≥10 → reject and revise.

W4b — Paraphrase Detection: For each sentence, estimate semantic similarity to closest draft_abstract.md sentence. If max >0.85 → flag.

Step 2d: Iteration

IF global_score ≥ 8.0 AND Coverage ≥ 7 → ACCEPT, go to Phase 3

ELSE:
    Identify lowest-scoring dimension
    Identify which Move(s) contribute to the weakness
    Re-generate ONLY those Move(s) (max 3 retries per Move)
    Re-assemble and re-score

Max 3 global iterations. Accept best available if limit reached.

[!CAUTION] If Accuracy < 8 → identify which facts are wrong and which Move contains them. If Coverage < 7 → list which ★★★ facts are missing — this BLOCKS acceptance. If a WEAK claim appears unqualified → Coverage = 0.

Phase 2.5: Adversarial Red Team

[!CAUTION] Role Switch: You are now a critical Nature reviewer who wants to REJECT this paper. Your goal is to find the WEAKEST sentence in the abstract and attack it. Priority targets: sentences whose Confidence scores have any axis < 0.8.

Step RT1: Attack Surface Scan

For each sentence S1–S(n), evaluate against 5 attack vectors:

Attack Vector Definitions:

Step RT2: Triage and Fix

IF no HIGH-severity attacks → PASS. Record "Red Team: CLEAN" and go to Phase 3.

IF ≥1 HIGH-severity attack:
    1. Identify which Move(s) produced the attacked sentence(s)
    2. Re-generate ONLY those Move(s) with the critique as an ADDITIONAL constraint
       (e.g., "Reviewer says S3 overclaims — soften 'establishes' to 'suggests'")
    3. Re-assemble → re-score → re-Red-Team
    Max 2 Red Team iterations. Accept best available if limit reached.

Step RT3: Red Team Log

Record in abstract_scoring_matrix.md:

## Red Team Report
| S# | Attack | Severity | Action Taken | Resolved? |
|:--|:--|:--|:--|:--|
| S3 | Overclaim: "establishes" for MODERATE claim | HIGH | Changed to "suggests" | ✅ |
| S5 | Vague: "significant improvement" | MED | Acceptable — kept | N/A |

Phase 3: Save Output

[!CAUTION] You MUST use the write_to_file tool to save BOTH output files:

1. abstract_candidates.md — Move Workbench + assembled abstract
2. abstract_scoring_matrix.md — full scoring breakdown

Traceability Table

Output Templates

abstract_candidates.md

# Abstract — Move-by-Move Generation

## Move Workbench

| Move | Role | Winner | Sentence |
|:--|:--|:--|:--|
| M1 | Context | [α/β] | [sentence text] |
| M2 | Gap | [α/β] | [sentence text] |
| M3 | Pivot | [α/β] | [sentence text] |
| M3b | Mechanism | [α/β] | [sentence text] |
| M4 | Key Result | [α/β] | [sentence text] |
| M5 | Application | [α/β] | [sentence text] |
| M6 | Impact | [α/β] | [sentence text] |

## Smoothing Log
- [list of minor changes made during smoothing pass]

---

## 🏆 Polished Abstract (Final)
[full assembled + smoothed abstract, ≤200 words]

## Iteration Log
- Round 1: Score = X.X. Weakness: [dimension]. Action: [which Moves revised].
- ...

abstract_scoring_matrix.md

# Abstract Scoring Matrix

| Dimension | Score | Rationale |
|:--|:--|:--|
| Novelty | X/10 | [rationale] |
| Accuracy | X/10 | [rationale] |
| Coverage | X/10 | [rationale] |
| Flow | X/10 | [rationale] |
| Impact | X/10 | [rationale] |
| **Average** | **X.X** | |

## Theme-Rheme Chain
| Pair | Chains? | Detail |
|:--|:--|:--|
| M1→M2 | ✅/❌ | [rheme] → [theme] |
| ... | | |

## Information Density
| Move | IU count | ≤6? |
|:--|:--|:--|
| M1 | __ | ✅/❌ |
| ... | | |

Shape: [bell/flat/spiky]

## Source Lock
| Check | Result | Pass? |
|:--|:--|:--|
| Max 5-gram overlap | [N] words | ✅/❌ |
| Max semantic similarity | [0.XX] | ✅/❌ |

## Fact Check
- [F1] "[metric]" → ✅ present / ❌ missing
- [F2] ...

## Traceability
[copy from Phase 3 traceability table]