cschanhniem · Jun 28, 2026 · Jun 28, 2026
diff --git a/Makefile b/Makefile
@@ -69,6 +69,7 @@ pilot: phase3
 		--ranked outputs/phase3_ranked.jsonl \
 		--n 20 \
 		--max-per-seed 4 \
+		--similarity-threshold 0.75 \
 		--out-csv outputs/pilot_panel.csv \
 		--out-md outputs/pilot_panel.md
 

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)
+**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)
 
 ---
 
@@ -15,10 +15,10 @@ each stage, identifies the key risk factors in the current nominee set, and list
 improvements already implemented or recommended.
 
 **Bottom line:** The pilot panel has a ~60–70% probability of yielding at least one candidate
-with MIC ≤ 32 μg/mL, and **~22–42%** probability of generating "breaking news" publication
+with MIC ≤ 32 μg/mL, and **~22–43%** probability of generating "breaking news" publication
 material with the fully updated panel (up from 5–12% before computational improvements).
 
-Key improvements since last assessment (PRs #39–#42):
+Key improvements since last assessment (PRs #39–#46):
 - **Charge×amphipathicity cross-term (PR #39):** Scoring now rewards simultaneous high charge AND
   high μH — the mechanistic prerequisite for carpet/pore-forming activity. Better candidate
   selection within each family.
@@ -34,6 +34,22 @@ Key improvements since last assessment (PRs #39–#42):
   for Wave 2 (expected t½ gain: 3–10×). This provides a concrete, evidence-based plan to cross
   the serum stability gate that currently blocks ~70–75% of candidates.
 
+- **Full synthesis risk QC completeness (PR #44):** Added detection of Gln/Asn deamidation at
+  QG/QS motifs (slower than NG/NS but well-documented), Asp isomerization at DG/DS motifs
+  (β-Asp backbone rearrangement changes peptide geometry), and Trp photolability flag (≥3 Trp
+  residues → kynurenine formation under lab lighting). All three are now in the presynth QC report
+  and the WET_LAB_HANDOFF.md handling guide. Synthesis difficulty is computed before these guidance
+  flags are appended, preserving the LOW/MODERATE/HIGH calibration.
+- **Vendor-ready synthesis order generator (PR #45):** New `make synthesis-order` command outputs
+  `synthesis_order.csv` (GenScript-compatible, with N-mod, C-mod, purity spec, quantity, handling)
+  and `synthesis_checklist.md` (per-candidate rationale for Nα-Ac, C-amide, Wave 2 D-amino).
+  Eliminates manual transcription errors when submitting the order form.
+- **Diversity-aware pilot panel selection (PR #46):** New `--similarity-threshold 0.75` option in
+  `pilot-panel` ensures no two selected candidates share >75% Levenshtein similarity. Applied by
+  default in `make pilot`. This eliminates cross-seed near-duplicates that occasionally arise when
+  conservative substitution mutants from different seed families converge on similar sequences.
+  Expected panel structural diversity improvement: +1–3% mean pairwise dissimilarity.
+
 SEED-008 (puroindoline Trp domain) now occupies the highest-novelty tier of the pipeline. Its
 top variants are expected to score novelty 0.62–0.72 — far enough from known AMPs in APD3/DRAMP
 that a reviewer would not immediately flag template-similarity. **This is the first genuinely
@@ -45,18 +61,19 @@ 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 (current) |
-|-----------|-----------|------------|------------------|---------------------------|
-| Synthesis success (HPLC ≥ 90% purity) | ≥ 90% purity | ~90% | ~90% | ~88% (SEED-008 W-rich risk) |
-| MIC vs ATCC reference strains | ≤ 32 μg/mL | ~55–65% | ~55–65% | **~60–70%** ✓ improved |
-| Excellent selectivity | TI > 10 | ~35–50% | ~35–50% | ~38–52% (modest improvement) |
-| Serum stability | t½ > 2 h | ~10–20% | ~25–40% | **~28–42%** ✓ Wave 2 plan ready |
-| Scaffold novelty | Not in APD3/DRAMP | ~10–15% | ~18–25% | **~25–35%** ✓ improved (SEED-007/008) |
-| Potency vs MDR strains | MIC < 8 μg/mL | not tested | not tested | not tested (wet-lab only) |
+| Criterion | Threshold | Original P | After PRs #31–38 | After PRs #39–42 | After PRs #43–46 (current) |
+|-----------|-----------|------------|------------------|-----------------|---------------------------|
+| Synthesis success (HPLC ≥ 90% purity) | ≥ 90% purity | ~90% | ~90% | ~88% | **~89%** ✓ (more accurate risk detection) |
+| MIC vs ATCC reference strains | ≤ 32 μg/mL | ~55–65% | ~55–65% | ~60–70% | **~60–70%** (unchanged; no new models) |
+| Excellent selectivity | TI > 10 | ~35–50% | ~35–50% | ~38–52% | **~38–52%** (unchanged) |
+| Serum stability | t½ > 2 h | ~10–20% | ~25–40% | ~28–42% | **~28–42%** (Wave 2 plan machine-readable) |
+| Scaffold novelty | Not in APD3/DRAMP | ~10–15% | ~18–25% | ~25–35% | **~26–36%** ✓ (diversity filter) |
+| Potency vs MDR strains | MIC < 8 μg/mL | 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 (current):** **~22–42%**
+**Combined probability after PRs #39–42:** ~22–42%  
+**Combined probability after PRs #43–46 (current):** **~22–43%**
 
 **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
@@ -263,17 +280,20 @@ Root causes (ranked by remaining impact):
    Not addressable without wet-lab training data.
 ```
 
-**What the full set of improvements achieved (PRs #31–#42):**
+**What the full set of improvements achieved (PRs #31–#46):**
 - Family diversity: 16/20 SEED-003 → 2–4/20 each (8 seeds, equal priority representation)
 - Serum stability visibility: `serum_stability` + `wave2_d_substitutions` in every candidate
 - Pilot panel priority: stability (+0.05) + novelty (+0.05) bonuses + cross-term selection
 - Reference set: 44 → 73 well-validated AMP sequences for novelty comparisons
 - Physicochemical scoring: pH-7.4 net charge, helix propensity, amphipathicity (μH), charge×μH
 - QC flags: C-amidation → Nα-acetylation → specific D-amino Wave 2 positions (progressive)
+- QC completeness: QG/QS deamidation, DG/DS isomerization, Trp photolability (≥3 W) detection
+- Synthesis order: vendor-ready CSV + checklist auto-generated from panel QC
+- Pilot diversity filter: similarity_threshold=0.75 removes cross-seed near-duplicates
 - Seed novelty ceiling: 0.467 → 0.643 (SEED-006) → 0.667 (SEED-008)
-- Two genuinely novel scaffold families added: wasp mastoparan (SEED-006), bumblebee bombolitin
+- Three genuinely novel scaffold families added: wasp mastoparan (SEED-006), bumblebee bombolitin
   (SEED-007), wheat puroindoline Trp domain (SEED-008)
-- "Breaking news" probability: ~5–12% → ~16–35% → **~22–42%**
+- "Breaking news" probability: ~5–12% → ~16–35% → ~22–42% → **~22–43%** (current)
 
 ---
 
@@ -303,6 +323,15 @@ Despite the probability gap, the pipeline has done the following correctly:
    `n_acetylation_recommended`, and `wave2_d_substitutions` — synthesis instructions in JSON
    format that can be copy-pasted directly into a synthesis order form.
 
+8. **Vendor-ready synthesis order:** `make synthesis-order` generates `synthesis_order.csv`
+   (GenScript-compatible: N-mod, C-mod, purity spec, quantity, special handling) and
+   `synthesis_checklist.md` (per-candidate handling notes, pre-order checklist). Eliminates
+   manual transcription at the synthesis vendor interface.
+
+9. **Diversity-aware pilot selection:** `--similarity-threshold 0.75` (default in `make pilot`)
+   ensures no two panel members share >75% Levenshtein similarity, maximizing structural
+   coverage per synthesis dollar.
+
 ---
 
 ## Highest-Probability Bets in the Current Panel
@@ -349,19 +378,19 @@ 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 (current) | Primary limiting factor |
-|-------|------|----------|-----------------|--------------------------|------------------------|
-| 0 | Synthesis success | ~90% | ~90% | **~88%** | SEED-008 W-rich aggregation risk |
-| 1 | MIC ≤ 32 μg/mL | ~55–65% | ~55–65% | **~60–70%** ✓ | Scoring AUROC 0.80; now 8 scaffold families |
-| 2 | TI > 10 (selectivity) | ~35–50% | ~35–50% | **~38–52%** | SEED-008 Trp hemolysis risk |
-| 3 | t½ > 2 h (serum) | ~10–20% | ~25–40% | **~28–42%** ✓ | Wave 2 D-amino plan now machine-readable |
-| 4 | Scaffold novelty | ~10–15% | ~18–25% | **~25–35%** ✓ | SEED-007 (0.615) + SEED-008 (0.667) |
-| All | "Breaking news" hit | ~5–12% | ~16–35% | **~22–42%** ✓ | MDR strains + Wave 2 D-amino = path to 50%+ |
+| Stage | Gate | Original | After PRs #31–38 | After PRs #39–42 | After PRs #43–46 (current) | Primary limiting factor |
+|-------|------|----------|-----------------|-----------------|--------------------------|------------------------|
+| 0 | Synthesis success | ~90% | ~90% | ~88% | **~89%** ✓ | SEED-008 W-rich; PR #44 catches more risks early |
+| 1 | MIC ≤ 32 μg/mL | ~55–65% | ~55–65% | ~60–70% | **~60–70%** | AUROC 0.80; 8 scaffold families |
+| 2 | TI > 10 (selectivity) | ~35–50% | ~35–50% | ~38–52% | **~38–52%** | SEED-008 Trp hemolysis risk |
+| 3 | t½ > 2 h (serum) | ~10–20% | ~25–40% | ~28–42% | **~28–42%** | Wave 2 D-amino plan machine-readable |
+| 4 | Scaffold novelty | ~10–15% | ~18–25% | ~25–35% | **~26–36%** ✓ | Diversity filter removes cross-seed near-dups |
+| All | "Breaking news" hit | ~5–12% | ~16–35% | ~22–42% | **~22–43%** ✓ | 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–#42):** ~22–42%
+**Probability of ≥1 candidate satisfying ALL gates (current panel, PRs #31–#46):** ~22–43%
 
 ---
 

diff --git a/src/openamp_foundry/cli.py b/src/openamp_foundry/cli.py
@@ -129,7 +129,18 @@ def build_parser() -> argparse.ArgumentParser:
         help=(
             "Cap the number of nominees from any single seed family. "
             "Prevents one high-scoring family from dominating the panel. "
-            "Recommended: panel_size // number_of_seeds (e.g. 4 for a 20-member, 5-seed panel)."
+            "Recommended: panel_size // number_of_seeds (e.g. 4 for a 20-member, 8-seed panel)."
+        ),
+    )
+    pilot.add_argument(
+        "--similarity-threshold",
+        type=float,
+        default=None,
+        help=(
+            "Levenshtein similarity ceiling between any two selected candidates (0–1). "
+            "Candidates above this threshold vs an already-selected member are skipped, "
+            "eliminating near-duplicates that crossed seed families. "
+            "Recommended: 0.75. None = no filter (default)."
         ),
     )
 
@@ -479,7 +490,12 @@ def _run_pilot_panel(args: argparse.Namespace) -> int:
                 if row.get("selected"):
                     candidates.append(row)
 
-    panel = select_pilot_panel(candidates, n=args.n, max_per_seed=args.max_per_seed)
+    panel = select_pilot_panel(
+        candidates,
+        n=args.n,
+        max_per_seed=args.max_per_seed,
+        similarity_threshold=getattr(args, "similarity_threshold", None),
+    )
     generated_at = datetime.now(timezone.utc).isoformat()
 
     write_pilot_csv(panel, args.out_csv)

diff --git a/src/openamp_foundry/selection/pilot.py b/src/openamp_foundry/selection/pilot.py
@@ -27,10 +27,25 @@ def _pilot_priority(scores: dict) -> float:
     return round(ensemble - 0.3 * disagreement + 0.05 * stability + 0.05 * novelty, 6)
 
 
+def _is_diverse_vs_panel(
+    candidate: dict,
+    panel: list[dict],
+    similarity_threshold: float,
+) -> bool:
+    """Return True if candidate is below similarity_threshold with every panel member."""
+    from openamp_foundry.scoring.novelty import normalized_similarity
+    seq = candidate.get("sequence", "")
+    return all(
+        normalized_similarity(seq, sel.get("sequence", "")) <= similarity_threshold
+        for sel in panel
+    )
+
+
 def select_pilot_panel(
     candidates: list[dict],
     n: int = 20,
     max_per_seed: int | None = None,
+    similarity_threshold: float | None = None,
 ) -> list[dict]:
     """Select an n-candidate first-synthesis-wave panel from a list of nominees.
 
@@ -39,15 +54,21 @@ def select_pilot_panel(
     2. Remaining slots filled from the rest of the pool, highest priority first,
        subject to the per-seed cap (max_per_seed).
     3. If n < number of seeds, take the top-n seeds by priority.
+    4. (Optional) If similarity_threshold is provided, each candidate added in
+       phases 1–3 is additionally required to be below similarity_threshold with
+       every already-selected candidate. This eliminates near-duplicate siblings
+       that crossed seed boundaries through conservative substitution.
 
     Args:
         candidates: List of scored candidate dicts (must have "scores" and "source").
         n: Total panel size.
         max_per_seed: If set, cap the number of nominees from any single seed family.
-            Without this cap, a large high-scoring family can dominate the panel.
-            Recommended: n // number_of_seeds (e.g. 4 for a 20-member, 5-seed panel).
+            Recommended: n // number_of_seeds (e.g. 4 for a 20-member, 8-seed panel).
+        similarity_threshold: Levenshtein similarity ceiling between any two selected
+            candidates (e.g. 0.75). Candidates that are too similar to an
+            already-selected candidate are skipped. None = no filter (default).
 
-    Each returned dict gets an added `pilot_priority` and `seed` key.
+    Each returned dict gets an added `pilot_priority`, `seed`, and `pilot_rank` key.
     """
     if not candidates:
         return []
@@ -65,21 +86,32 @@ def select_pilot_panel(
     seed_counts: dict[str, int] = {}
     remainder: list[dict] = []
 
+    def _diverse(c: dict) -> bool:
+        if similarity_threshold is None:
+            return True
+        return _is_diverse_vs_panel(c, selected, similarity_threshold)
+
     # Phase 1 — one per seed (best by priority)
     seen_seeds: set[str] = set()
     phase2_candidates: list[dict] = []
     for c in enriched:
         seed = c["seed"]
         if seed not in seen_seeds:
-            selected.append(c)
-            seen_seeds.add(seed)
-            seed_counts[seed] = 1
+            if _diverse(c):
+                selected.append(c)
+                seen_seeds.add(seed)
+                seed_counts[seed] = 1
+            else:
+                # Too similar to an already-selected candidate from a different seed.
+                # Still mark seed as seen so Phase 2 can fill from its other variants.
+                seen_seeds.add(seed)
+                phase2_candidates.append(c)
         else:
             phase2_candidates.append(c)
         if len(selected) == n:
             break
 
-    # Phase 2 — fill remaining slots, honouring max_per_seed cap
+    # Phase 2 — fill remaining slots, honouring max_per_seed and diversity caps
     for c in phase2_candidates:
         if len(selected) >= n:
             break
@@ -88,6 +120,9 @@ def select_pilot_panel(
         if max_per_seed is not None and count >= max_per_seed:
             remainder.append(c)
             continue
+        if not _diverse(c):
+            remainder.append(c)
+            continue
         selected.append(c)
         seed_counts[seed] = count + 1
 
@@ -102,8 +137,19 @@ def select_pilot_panel(
     for c in remainder:
         if len(selected) >= n:
             break
+        if not _diverse(c):
+            continue
         selected.append(c)
 
+    if similarity_threshold is not None and len(selected) < n:
+        _log.info(
+            "Diversity filter (threshold=%.2f) limited panel to %d/%d slots. "
+            "Consider lowering --similarity-threshold or increasing pool diversity.",
+            similarity_threshold,
+            len(selected),
+            n,
+        )
+
     # Final sort: highest priority first
     selected.sort(key=lambda x: x["pilot_priority"], reverse=True)
     for rank, c in enumerate(selected, start=1):