cschanhniem · Jun 28, 2026 · Jun 28, 2026 · Jun 28, 2026
diff --git a/Makefile b/Makefile
@@ -20,7 +20,7 @@ lint:
 	$(RUFF) check src tests scripts
 
 ci: lint test
-	@echo "CI passed: lint + 1124-test suite green"
+	@echo "CI passed: lint + 1158-test suite green"
 
 bench-leakage:
 	PYTHONPATH=src $(PYTHON) -m openamp_foundry.cli bench leakage \

diff --git a/docs/DISCOVERY_PREDICTION.md b/docs/DISCOVERY_PREDICTION.md
@@ -3,7 +3,7 @@
 **Pipeline:** OpenAMP-Foundry v0.1.0  
 **Date:** 2026-06-28 (updated 2026-06-28)  
 **Status:** Pre-synthesis scientific assessment — for expert review before ordering  
-**Completed improvements:** Serum stability scoring (PR #31/#32), Family diversity cap (PR #31), Reference set expansion 44→73 sequences (PR #33), Net charge pH 7.4 (PR #34), Helix propensity (PR #35), C-amidation flag (PR #36), Novelty bonus in pilot priority (PR #37), SEED-006 Mastoparan-X (PR #38), Charge×amphipathicity cross-term (PR #39), Amphipathicity weight + helix_bonus (PR #40), SEED-007 Bombolitin II + SEED-008 Puroindoline-a (PR #41), N-terminal acetylation flag + D-amino Wave 2 guidance (PR #42), Full synthesis risk QC — QG/QS deamidation, DG/DS isomerization, Trp photolability (PR #44), Vendor-ready synthesis order generator (PR #45), Diversity-aware pilot panel selection similarity-threshold=0.75 (PR #46), Selectivity proxy + HIGH_CYTOTOX_RISK flag — charge/GRAVY-based mammalian cytotoxicity risk detector (PR #47), selectivity_proxy routed into pilot_priority formula — low-cytotox-risk candidates gently demoted (PR #48), Elastase resistance (HNE 3-protease stability model) + aggregation propensity scoring (synthesis feasibility penalty) (PR #49)
+**Completed improvements:** Serum stability scoring (PR #31/#32), Family diversity cap (PR #31), Reference set expansion 44→73 sequences (PR #33), Net charge pH 7.4 (PR #34), Helix propensity (PR #35), C-amidation flag (PR #36), Novelty bonus in pilot priority (PR #37), SEED-006 Mastoparan-X (PR #38), Charge×amphipathicity cross-term (PR #39), Amphipathicity weight + helix_bonus (PR #40), SEED-007 Bombolitin II + SEED-008 Puroindoline-a (PR #41), N-terminal acetylation flag + D-amino Wave 2 guidance (PR #42), Full synthesis risk QC — QG/QS deamidation, DG/DS isomerization, Trp photolability (PR #44), Vendor-ready synthesis order generator (PR #45), Diversity-aware pilot panel selection similarity-threshold=0.75 (PR #46), Selectivity proxy + HIGH_CYTOTOX_RISK flag — charge/GRAVY-based mammalian cytotoxicity risk detector (PR #47), selectivity_proxy routed into pilot_priority formula — low-cytotox-risk candidates gently demoted (PR #48), Elastase resistance (HNE 3-protease stability model) + aggregation propensity scoring (synthesis feasibility penalty) (PR #49), Aggregation-safe mutation generation + balanced K/R charge variants + SynthQC continuous aggregation score (PR #50)
 
 ---
 
@@ -65,6 +65,21 @@ Key improvements since last assessment (PRs #39–#49):
   (proxy=0.30) are now demoted by 0.035 relative to a fully selective peer with equal ensemble
   score — they can still earn their max_per_seed slots if they score well on ensemble, but are
   no longer ranked identically to selective candidates with the same activity profile.
+- **Aggregation-safe mutation generation + balanced K/R variants (PR #50):** Generator-level
+  improvements that make candidate generation consistent with the scoring and QC layers:
+  (a) `generate_aggregation_safe_double_variants()` filters out double substitutions that would
+  create hydrophobic runs ≥4 residues, preventing the generator from producing sequences that
+  are immediately penalised by `synthesis_feasibility_score()` (run risk detected and filtered
+  *before* synthesis ordering, not discovered after). (b) `generate_balanced_charge_variants()`
+  generates both K and R replacements at each polar position (S/T/N/Q), giving the scoring
+  pipeline twice as many charge-enhanced options to rank — K (KD=−3.9) vs R (KD=−4.5) differ
+  in GRAVY and trypsin sensitivity, and the best one for each seed's GRAVY profile can now be
+  selected automatically. (c) `SynthQC.aggregation_propensity_score` (float [0,1]) added to
+  the presynth QC dataclass and to_dict(), providing continuous gradient information alongside
+  the existing `aggregation_risk` boolean — useful for prioritising borderline candidates
+  (e.g. score 0.14 vs 0.0 distinguishes a run-of-4 from a truly safe sequence).
+  Code quality: closes the divergence between the QC boolean flag and the continuous scoring model.
+
 - **Elastase resistance + aggregation propensity (PR #49):** Two new computational features
   addressing the top pipeline gaps identified in post-PR-48 audit:
   (a) `aggregation_propensity()` [0,1] — quantifies on-resin hydrophobic-run and beta-branched
@@ -85,21 +100,22 @@ novel scaffold family in the pilot panel.**
 
 To be publishable as a significant advance in AMP discovery, a candidate must satisfy all of:
 
-| Criterion | Threshold | Original P | After PRs #31–38 | After PRs #39–42 | After PRs #43–47 | After PR #48 | After PR #49 (current) |
-|-----------|-----------|------------|------------------|-----------------|-----------------|--------------|----------------------|
-| Synthesis success (HPLC ≥ 90% purity) | ≥ 90% purity | ~90% | ~90% | ~88% | **~89%** ✓ | **~89%** | **~90%** ✓ (aggregation model) |
-| MIC vs ATCC reference strains | ≤ 32 μg/mL | ~55–65% | ~55–65% | ~60–70% | **~60–70%** | **~60–70%** | **~60–70%** (unchanged) |
-| Excellent selectivity | TI > 10 | ~35–50% | ~35–50% | ~38–52% | **~40–55%** ✓ | **~41–56%** ✓ | **~41–56%** (unchanged) |
-| Serum stability | t½ > 2 h | ~10–20% | ~25–40% | ~28–42% | **~28–42%** | **~28–42%** | **~29–44%** ✓ (elastase model) |
-| Scaffold novelty | Not in APD3/DRAMP | ~10–15% | ~18–25% | ~25–35% | **~26–36%** ✓ | **~26–36%** | **~26–36%** (unchanged) |
-| Potency vs MDR strains | MIC < 8 μg/mL | not tested | not tested | not tested | not tested | not tested | not tested (wet-lab only) |
+| Criterion | Threshold | Original P | After PRs #31–38 | After PRs #39–42 | After PRs #43–47 | After PR #48 | After PR #49 | After PR #50 (current) |
+|-----------|-----------|------------|------------------|-----------------|-----------------|--------------|--------------|----------------------|
+| Synthesis success (HPLC ≥ 90% purity) | ≥ 90% purity | ~90% | ~90% | ~88% | **~89%** ✓ | **~89%** | **~90%** ✓ (agg model) | **~90%** ✓ (agg-safe gen) |
+| MIC vs ATCC reference strains | ≤ 32 μg/mL | ~55–65% | ~55–65% | ~60–70% | **~60–70%** | **~60–70%** | **~60–70%** | **~60–70%** (unchanged) |
+| Excellent selectivity | TI > 10 | ~35–50% | ~35–50% | ~38–52% | **~40–55%** ✓ | **~41–56%** ✓ | **~41–56%** | **~41–56%** (unchanged) |
+| Serum stability | t½ > 2 h | ~10–20% | ~25–40% | ~28–42% | **~28–42%** | **~28–42%** | **~29–44%** ✓ | **~29–44%** (unchanged) |
+| Scaffold novelty | Not in APD3/DRAMP | ~10–15% | ~18–25% | ~25–35% | **~26–36%** ✓ | **~26–36%** | **~26–36%** | **~26–36%** (unchanged) |
+| Potency vs MDR strains | MIC < 8 μg/mL | not tested | not tested | not tested | not tested | not tested | not tested | not tested (wet-lab only) |
 
 **Combined probability of satisfying all criteria simultaneously (original):** ~5–12%  
 **Combined probability after PRs #31–38:** ~16–35%  
 **Combined probability after PRs #39–42:** ~22–42%  
 **Combined probability after PRs #43–47:** ~24–45%  
 **Combined probability after PR #48:** ~25–46%  
-**Combined probability after PR #49 (current):** **~27–47%**
+**Combined probability after PR #49:** ~27–47%  
+**Combined probability after PR #50 (current):** **~27–47%** (generator consistency — no new scoring gates)
 
 **Methodology note:** Combined probability is computed as P(≥1 from 20) under an independent-candidate
 Poisson model: P_individual = P(S0) × P(S1) × P(S2) × P(S3) × P(S4). Using gate midpoints
@@ -436,6 +452,17 @@ Despite the probability gap, the pipeline has done the following correctly:
     content are now correctly penalised, ensuring the pipeline's stability gate reflects in-vivo
     biology rather than only in-vitro serum dilution assay conditions.
 
+14. **Generator–scorer–QC consistency (PR #50):** Three layers of the pipeline now use the same
+    aggregation threshold (VILMFW run ≥ 4) consistently: (a) `generate_aggregation_safe_double_variants()`
+    in the generator filters out unsafe mutations before they enter the candidate pool; (b)
+    `aggregation_propensity()` in the scorer penalises unsafe sequences that reach synthesis
+    feasibility scoring; (c) `aggregation_risk` boolean and the new `aggregation_propensity_score`
+    float in `SynthQC` report both to the wet-lab operator. All three layers use the same definition
+    of "aggregation-risky", removing any possible divergence between what the generator creates,
+    what the scorer penalises, and what the QC reports. Additionally, `generate_balanced_charge_variants()`
+    now generates both K and R charge-enhanced options for each polar position, letting the scoring
+    pipeline select the variant with the better GRAVY/selectivity profile automatically.
+
 ---
 
 ## Highest-Probability Bets in the Current Panel
@@ -482,19 +509,19 @@ Executing all four actions on the best Wave 1 hits would push the combined proba
 
 ## Summary Table: Probability by Gate
 
-| Stage | Gate | Original | After PRs #31–38 | After PRs #39–42 | After PRs #43–47 | After PR #48 (current) | Primary limiting factor |
-|-------|------|----------|-----------------|-----------------|-----------------|----------------------|------------------------|
-| 0 | Synthesis success | ~90% | ~90% | ~88% | **~89%** ✓ | **~89%** | SEED-008 W-rich; PR #44 catches more risks early |
-| 1 | MIC ≤ 32 μg/mL | ~55–65% | ~55–65% | ~60–70% | **~60–70%** | **~60–70%** | AUROC 0.811; 8 scaffold families |
+| Stage | Gate | Original | After PRs #31–38 | After PRs #39–42 | After PRs #43–47 | After PRs #48–50 (current) | Primary limiting factor |
+|-------|------|----------|-----------------|-----------------|-----------------|--------------------------|------------------------|
+| 0 | Synthesis success | ~90% | ~90% | ~88% | **~89%** ✓ | **~90%** ✓ (agg model + agg-safe gen) | SEED-008 W-rich; all aggregation risks now modelled |
+| 1 | MIC ≤ 32 μg/mL | ~55–65% | ~55–65% | ~60–70% | **~60–70%** | **~60–70%** | AUROC 0.814; 8 scaffold families |
 | 2 | TI > 10 (selectivity) | ~35–50% | ~35–50% | ~38–52% | **~40–55%** ✓ | **~41–56%** ✓ | sel_proxy in pilot ranking demotes SEED-004 |
-| 3 | t½ > 2 h (serum) | ~10–20% | ~25–40% | ~28–42% | **~28–42%** | **~28–42%** | Wave 2 D-amino plan machine-readable |
+| 3 | t½ > 2 h (serum) | ~10–20% | ~25–40% | ~28–42% | **~28–42%** | **~29–44%** ✓ (3-protease model) | Wave 2 D-amino plan machine-readable |
 | 4 | Scaffold novelty | ~10–15% | ~18–25% | ~25–35% | **~26–36%** ✓ | **~26–36%** | Diversity filter removes cross-seed near-dups |
-| All | "Breaking news" hit | ~5–12% | ~16–35% | ~22–42% | ~24–45% | **~25–46%** ✓ | MDR strains + Wave 2 D-amino = path to 50%+ |
+| All | "Breaking news" hit | ~5–12% | ~16–35% | ~22–42% | ~24–45% | **~27–47%** ✓ | MDR strains + Wave 2 D-amino = path to 50%+ |
 
 **Probability of ≥1 active AMP from pilot panel (Stage 1 only):** ~90–97%  
 (Probability of zero active from 20 candidates with ~65% individual hit rate ≈ 3–10%)
 
-**Probability of ≥1 candidate satisfying ALL gates (current panel, PRs #31–#48):** ~25–46%
+**Probability of ≥1 candidate satisfying ALL gates (current panel, PRs #31–#50):** ~27–47%
 
 ---
 

diff --git a/src/openamp_foundry/features/physchem.py b/src/openamp_foundry/features/physchem.py
@@ -193,7 +193,7 @@ def aggregation_propensity(sequence: str) -> float:
     1. Hydrophobic run length: consecutive residues from {V,I,L,M,F,W} ≥ 4 drive
        self-association during SPPS, solubilisation, and assay buffers.
        Risk onset at run ≥ 4 (same threshold as QC HYDROPHOBIC_RUN_RE flag; same
-       residue set AGG_HYDROPHOBIC). Ramp: 0.0 at run=4 → 1.0 at run ≥ 8.
+       residue set AGG_HYDROPHOBIC). Ramp: 0.0 at run < 4, 0.20 at run=4, 1.0 at run ≥ 8.
     2. Beta-branched residue density (Val, Ile, Thr) > 20% promotes β-strand over
        α-helix and intermolecular β-sheet aggregation in concentrated solutions.
 

diff --git a/src/openamp_foundry/generators/template_mutator.py b/src/openamp_foundry/generators/template_mutator.py
@@ -14,6 +14,8 @@
   - Preserve backbone length (no insertions/deletions in this generator)
   - Never target charged residues for hydrophobic substitution (or vice versa)
   - Use deterministic seeding so output is reproducible
+  - Aggregation-safe double substitutions avoid creating hydrophobic runs ≥4
+    (matching the QC HYDROPHOBIC_RUN_RE threshold used in presynth_check.py)
 
 This generator does NOT claim to produce biologically active peptides.
 All generated sequences must pass through the full pipeline safety filters
@@ -23,6 +25,8 @@
 
 import random
 
+from openamp_foundry.features.physchem import AGG_HYDROPHOBIC as _AGG_HYDROPHOBIC
+
 CANONICAL_AA = "ACDEFGHIKLMNPQRSTVWY"
 
 # Conservative substitution groups: within-group swaps preserve physicochemical character
@@ -36,6 +40,8 @@
     frozenset("C"),          # cysteine — singleton; substitutions restricted
 ]
 
+# _AGG_HYDROPHOBIC is imported from physchem.py (single source of truth for the VILMFW residue set).
+
 
 def _conservative_substitutes(aa: str) -> list[str]:
     """Return amino acids that can conservatively substitute for `aa`."""
@@ -45,6 +51,19 @@ def _conservative_substitutes(aa: str) -> list[str]:
     return []
 
 
+def _max_hydrophobic_run(sequence: str) -> int:
+    """Return the length of the longest consecutive _AGG_HYDROPHOBIC run in sequence."""
+    max_run = 0
+    current = 0
+    for aa in sequence:
+        if aa in _AGG_HYDROPHOBIC:
+            current += 1
+            max_run = max(max_run, current)
+        else:
+            current = 0
+    return max_run
+
+
 def generate_single_substitution_variants(sequence: str) -> list[str]:
     """Generate all single-substitution conservative variants of a template.
 
@@ -99,6 +118,55 @@ def generate_double_substitution_variants(
     return variants
 
 
+def generate_aggregation_safe_double_variants(
+    sequence: str,
+    n_samples: int = 20,
+    seed: int = 42,
+    max_run: int = 4,
+) -> list[str]:
+    """Like generate_double_substitution_variants but rejects variants with a
+    consecutive _AGG_HYDROPHOBIC run >= max_run.
+
+    Default max_run=4 matches the QC HYDROPHOBIC_RUN_RE threshold ([VILMFW]{4,})
+    and the aggregation_propensity() run component in physchem.py. This prevents
+    the generator from producing sequences that the synthesis scoring pipeline
+    would penalise for on-resin aggregation.
+
+    Returns up to n_samples unique aggregation-safe variants.
+    """
+    rng = random.Random(seed)
+    substitutable = [
+        (i, _conservative_substitutes(aa))
+        for i, aa in enumerate(sequence)
+        if _conservative_substitutes(aa)
+    ]
+    if len(substitutable) < 2:
+        return []
+
+    seen: set[str] = set()
+    variants = []
+    attempts = 0
+    # Larger attempt budget: some hydrophobic seeds need many tries to find safe pairs.
+    while len(variants) < n_samples and attempts < n_samples * 50:
+        attempts += 1
+        pos_a, subs_a = rng.choice(substitutable)
+        pos_b, subs_b = rng.choice(substitutable)
+        if pos_a == pos_b:
+            continue
+        sub_a = rng.choice(subs_a)
+        sub_b = rng.choice(subs_b)
+        seq_list = list(sequence)
+        seq_list[pos_a] = sub_a
+        seq_list[pos_b] = sub_b
+        new_seq = "".join(seq_list)
+        if _max_hydrophobic_run(new_seq) >= max_run:
+            continue
+        if new_seq not in seen:
+            seen.add(new_seq)
+            variants.append(new_seq)
+    return variants
+
+
 def generate_charge_enhanced_variants(
     sequence: str,
     n_samples: int = 10,
@@ -108,6 +176,10 @@ def generate_charge_enhanced_variants(
 
     Targets positions occupied by non-charged hydrophilic residues (S, T, N, Q)
     and replaces with K (higher charge density → better activity likeness).
+
+    Deprecated: generate_all_variants() now calls generate_balanced_charge_variants()
+    which generates both K and R for each polar position. This function is retained
+    for backward compatibility (used by existing tests and any external callers).
     """
     rng = random.Random(seed)
     targetable = [
@@ -127,6 +199,40 @@ def generate_charge_enhanced_variants(
     return variants
 
 
+def generate_balanced_charge_variants(
+    sequence: str,
+    n_positions: int = 10,
+) -> list[str]:
+    """Generate both K and R replacements for each polar (S/T/N/Q) position.
+
+    Unlike generate_charge_enhanced_variants(), which randomly picks K or R,
+    this function generates one K and one R variant per targetable position.
+    The scoring pipeline can then rank both options and select the better one
+    based on GRAVY / selectivity_proxy trade-offs:
+      - K (KD = -3.9): slightly less hydrophilic than R, slightly lower trypsin rate
+      - R (KD = -4.5): more hydrophilic, marginally higher GRAVY improvement
+
+    Args:
+        n_positions: maximum number of polar positions to target. Returns up to
+            2 × n_positions unique variants (one K variant and one R variant per position).
+            Named n_positions (not n_samples) to avoid confusion with other generator
+            functions where n_samples bounds the total output count directly.
+    """
+    targetable = [i for i, aa in enumerate(sequence) if aa in "STNQ"]
+    if not targetable:
+        return []
+
+    seen: set[str] = set()
+    variants = []
+    for i in targetable[:n_positions]:
+        for replacement in ("K", "R"):
+            new_seq = sequence[:i] + replacement + sequence[i + 1:]
+            if new_seq not in seen:
+                seen.add(new_seq)
+                variants.append(new_seq)
+    return variants
+
+
 def generate_all_variants(
     seed_sequence: str,
     n_double: int = 15,
@@ -135,13 +241,25 @@ def generate_all_variants(
 ) -> list[str]:
     """Generate all candidate variants for a template sequence.
 
-    Combines single substitutions, double substitutions, and charge-enhanced variants.
-    Deduplicates and removes the seed itself.
+    Combines single substitutions, aggregation-safe double substitutions, and
+    balanced (K+R) charge-enhanced variants. Deduplicates and removes the seed.
+
+    Double substitutions use generate_aggregation_safe_double_variants() to avoid
+    creating hydrophobic runs ≥4 that would be penalised by synthesis scoring.
+    Single substitutions are also filtered for the same threshold — a conservative swap
+    of an adjacent residue can extend a near-threshold run into aggregation territory.
+    Charge enhancements use generate_balanced_charge_variants() to produce both
+    K and R options at each polar position, letting the scoring pipeline rank them.
     """
     variants: set[str] = set()
-    variants.update(generate_single_substitution_variants(seed_sequence))
-    variants.update(generate_double_substitution_variants(seed_sequence, n_double, seed))
-    variants.update(generate_charge_enhanced_variants(seed_sequence, n_charge_enhance, seed))
+    # Single subs are filtered: a conservative A→L substitution adjacent to a run-of-3
+    # can create run-of-4, which would be flagged by synthesis scoring and QC.
+    variants.update(
+        v for v in generate_single_substitution_variants(seed_sequence)
+        if _max_hydrophobic_run(v) < 4
+    )
+    variants.update(generate_aggregation_safe_double_variants(seed_sequence, n_double, seed))
+    variants.update(generate_balanced_charge_variants(seed_sequence, n_charge_enhance))
     variants.discard(seed_sequence)
     return sorted(variants)