BETO × Evo2 — Epistemic Governance Wrapper for Zero Silent Completions in Genomic Generation

[aramirez-maza]

Hi Arc Institute team and community,

We wanted to share an integration we built on top of Evo2 that addresses a governance problem we believe is relevant to scientific use of DNA language models.

The problem

Evo2 is extraordinarily capable — but like all LLMs, it silently completes missing parameters. Ask it to generate a sequence without declaring organism, biosafety level, or sampling constraints, and it generates anyway, inferring what you did not say. In computational biology, this means sequences with no traceable decision record — a reproducibility and biosafety problem.

What we built

BETO-EVO2 — a governance wrapper using BETO Framework (an epistemic governance protocol for LLM-assisted systems) that enforces:

• 8 critical genomic parameters must be DECLARED before any Evo2 call
• NOT_STATED = execution blocked (no biosafety level declared → no call)
• 3 human approval gates before every NIM API call (topology, full spec, exact payload)
• TRACE_VERIFIED output with unique ID, spec_hash, and full epistemic manifest
• SQLite audit trail: who authorized what, when, with what parameters
• Zero silent completions

Real run result (Evo2-7B via NVIDIA NIM API)

{
  "trace_id":    "BETO-EVO2-20260320-92B0AA3C",
  "trace_status": "TRACE_VERIFIED",
  "model_used":  "evo2-7B",
  "latency_ms":  10645,
  "evo2_output": "GTCCAATTTGGTATTTTTCTACCATGTATTTTTTTTATTTTTTTTTATTTCCATT...",
  "epistemic_summary": {
    "DECLARED":   8,
    "NOT_STATED": 0,
    "INFERRED":   0
  }
}

The 8 critical parameters (all must be DECLARED)

Parameter	Why Critical
organism	Determines biological context and safety protocols
bio_function	Defines the functional target — must be operator-authorized
sequence_length	Directly affects reproducibility
biosafety_level	BSL-1 through BSL-4 — no default is acceptable
evo2_model_size	1B / 7B / 20B / 40B — affects quality and reproducibility
sampling_temp	Undeclared temperature = silent choice
top_k	Undeclared = unauditable diversity
excluded_motifs	Sequences that must not appear — [] is a valid declaration

The Gate-B operator review (real run)

✓ organism             = Homo sapiens        [DECLARED]
✓ bio_function         = modular BRCA1 expression [DECLARED]
✓ sequence_length      = 256                 [DECLARED]
✓ biosafety_level      = BSL-2               [DECLARED]
✓ evo2_model_size      = 7B                  [DECLARED]
✓ sampling_temp        = 0.7                 [DECLARED]
✓ top_k                = 4                   [DECLARED]
✓ excluded_motifs      = []                  [DECLARED]

Why this matters for science

• Reproducibility: Every run produces a spec_hash. Same hash = same experiment. No more "we used standard parameters" in Methods sections.
• Biosafety auditability: SQL query → who authorized a BSL-3 sequence generation, timestamp included.
• Publication-ready trace: The epistemic_manifest can be attached as supplementary material, proving no parameter was silently inferred.

Resources

Resource	Link
Full example (code + spec + real run)	https://github.com/aramirez-maza/beto-framework/tree/main/examples/beto-evo2
Technical FAQ (28 questions, 7 categories)	https://github.com/aramirez-maza/beto-framework/blob/main/examples/beto-evo2/FAQ.md
Real Evo2-7B result (TRACE_VERIFIED)	https://github.com/aramirez-maza/beto-framework/blob/main/examples/beto-evo2/resultado_real.json
BETO Framework	https://github.com/aramirez-maza/beto-framework

Quick start (simulation mode, no API key needed)

git clone https://github.com/aramirez-maza/beto-framework
cd beto-framework/examples/beto-evo2/beto_evo2
pip install requests
python3 main.py --interactive --operator-id your_id

Quick start (with NVIDIA NIM API key)

export NVIDIA_API_KEY="nvapi-your-key"
python3 main.py \
  --intent "organism: Homo sapiens, function: modular BRCA1 expression, 256bp, BSL-2, 7B, temperature: 0.7, top_k: 4, excluded_motifs: []" \
  --task GENERATION \
  --seed "ATGGATTTATCTGCTCTTCGCGTTGAAGAAGTACAAAATGTCATTAATGCTATGCAGAAAATCTTAGAGT" \
  --operator-id your_id \
  --output result.json

What changed since first post (v1.1)

Three resilience fixes applied after community review:

• 429 Rate Limit — now classified as RATE_LIMITED with Retry-After seconds
• Manifest on failure — operator decisions persisted in SQLite even when Evo2 fails
• DRY_RUN mode — --dry-run validates spec + gates without consuming NIM credits

Happy to discuss — and very open to feedback from the Arc Institute team on how this governance layer could be made more useful for the Evo2 community.