DefendableLedger · Royal Jelly Tiers

The five tiers

SwarmJelly assigns one of five Royal Jelly tiers to every extracted training pair, based on SwarmCurator’s 4-dimension grade.

tier	score band	what it is	use
Apex	4.5–5.0	Operator-grade ground truth · principal-level signal	Primary fine-tune corpus for Atlas + Curator + specialist Hacks
Honey	3.5–4.5	Production-ready · strong	Production fine-tune blends · validator chain training
Jelly	2.5–3.5	Solid corpus material · workhorse	Breadth coverage · DPO pair candidates
Pollen	1.5–2.5	Broad coverage · breadth signal	Diversity sampling · low-weight inclusion
Propolis	< 1.5	Edge cases · failure modes · repair lift	SwarmFixer corpus · repair-pair training

How the tier is assigned

score = mean([accuracy, cre_judgment, format, score_overall])

if score >= 4.5:  tier = "apex"
elif score >= 3.5: tier = "honey"
elif score >= 2.5: tier = "jelly"
elif score >= 1.5: tier = "pollen"
else:              tier = "propolis"

The 4-dim rubric is held by SwarmCurator-9B (Qwen 3.5 base · in-house cook). No external grader. No hosted-LLM tax.

Why tiering matters

A flat corpus is a wasted corpus. Mixing apex signal with propolis edge cases at equal weight dilutes both ends — your fine-tune learns muddy ground truth AND fails to focus on the actual repair lift. Tiering preserves signal density.

Apex is what you train the first pass on. Propolis is what you train SwarmFixer on. Pollen + Jelly are the breadth corpus. Honey is the production-ready blend.

This is how the corpus compounds toward operator-grade Hacks without drifting.

What lives in each tier directory

data/swarmjelly/
├── apex/
│   ├── SJP-20260524-01KSE...AKVHNB.json
│   └── SJP-20260524-01KSE...XYZ.json
├── honey/
├── jelly/
├── pollen/
├── propolis/
└── corpus_index.jsonl        # streaming index of all pairs (one line each)

Each pair file:

{
  "pair_id": "SJP-YYYYMMDD-ULID",
  "receipt_id": "DRR-YYYYMMDD-ULID",
  "verdict_id": "TRIB-YYYYMMDD-ULID",
  "tier": "apex|honey|jelly|pollen|propolis",
  "input": "<raw client language>",
  "output": { "assignment_type": "...", "expected_outputs": [...] },
  "rubric_scores": { "accuracy": 4.5, "cre_judgment": 4.7, "format": 4.8, "score": 4.6 },
  "created_at": "2026-05-24T22:33:26Z",
  "pair_sha256": "<sha256>"
}

The 88K-pair compounding math

At one receipt every ~6 minutes (a steady principal cadence), the eco system mints ~88,000 training pairs per year — entirely sovereign · zero marginal cost · automatically tiered. Five years compounds to ~440K pairs / ~242K in Apex+Honey ready for fine-tune blends.

Compare to outsourced labeling at $2/pair market rate: $96K/yr of equivalent value · captured for $0 marginal · with full provenance · without sending operator-grade data to a third-party labeler.

🐝 The corpus compounds · the cook gets sharper · to the shed.