Multi-Teacher Verification & Rejection Sampling Pipeline
How Solstice-AI filters, validates, and aligns reasoning traces using compiler sandboxes and mathematical assertion engines.
Overview
High-quality distillation requires eliminating hallucinated shortcuts and circular reasoning paths. The Solstice-AI verification harness executes every candidate trace through a 4-stage automated pipeline before inclusion into the Solace dataset.
Candidate Traces (2.2B Tokens)
│
▼
┌───────────────────────────────────────┐
│ Stage 1: Syntax & Trace Normalization │
└───────────────────┬───────────────────┘
│ (Pass: 86.2%)
▼
┌───────────────────────────────────────┐
│ Stage 2: Compiler & Sandbox Execution │
└───────────────────┬───────────────────┘
│ (Pass: 54.1%)
▼
┌───────────────────────────────────────┐
│ Stage 3: Formal Mathematical Checker │
└───────────────────┬───────────────────┘
│ (Pass: 71.3%)
▼
┌───────────────────────────────────────┐
│ Stage 4: Multi-Teacher Consensus Rank │
└───────────────────┬───────────────────┘
│ (Pass: 85.0%)
▼
Solace 1.0 Verified (630M+ Tokens)
1. Stage 2: Compiler & Execution Sandbox
All code synthesis samples are executed inside an ephemeral isolated gVisor sandbox with strict resource constraints (2 CPU cores, 4GB RAM, 10s timeout, zero network access).
from solstice_verify import SandboxRunner
runner = SandboxRunner(
language="python",
timeout_seconds=5,
memory_limit_mb=1024,
network_enabled=False
)
result = runner.execute_test_suite(
code=candidate_trace.generated_code,
unit_tests=candidate_trace.ground_truth_tests
)
if not result.all_passed:
# Reject candidate or trigger automated self-correction
reject_trace(candidate_trace.id, reason=result.error_log)
2. Stage 3: Formal Mathematical & Proof Assertion
Mathematical steps are decomposed into intermediate assertions:
- Symbolic Equivalence Check: Evaluated using SymPy and CAS engines to ensure algebraic transformations hold true at each step.
- Lean 4 Proof Verification: Olympiad-grade proofs are checked against the Lean 4 formal math kernel.
-- Lean 4 verification harness example
import Mathlib.Data.Nat.Basic
theorem solace_verified_prime_property (p : ℕ) (hp : Nat.Prime p) (h : p > 3) :
24 ∣ (p^2 - 1) := by
-- Verified formal proof steps extracted from consensus traces
sorry
3. Stage 4: Cross-Teacher Consensus Scoring
When multiple frontier models solve the same problem, we compute the semantic overlap of their reasoning graphs:
$$\text{Consensus Score}(T_1, \dots, T_k) = \frac{1}{\binom{k}{2}} \sum_{i < j} \text{GraphSim}(\mathcal{G}{T_i}, \mathcal{G}{T_j})$$
Traces with high agreement on intermediate reasoning landmarks receive higher sampling weights during student fine-tuning.