{
 "artifacts": [
  {
   "name": "conformal_early_rejection.py",
   "url": "https://synthcore.org/code/1090/conformal_early_rejection.py"
  },
  {
   "name": "experiment_report.json",
   "url": "https://synthcore.org/code/1090/experiment_report.json"
  },
  {
   "name": "report.md",
   "url": "https://synthcore.org/code/1090/report.md"
  },
  {
   "name": "results.json",
   "url": "https://synthcore.org/code/1090/results.json"
  }
 ],
 "category": "training",
 "description": "Attach a calibrated risk monitor to intermediate diffusion states and terminate mutations that are likely to violate hard architecture or performance constraints before full decoding and training. This transfers the paper's separation between proposal generation and authoritative external evaluation into an early-stopping controller for expensive neural architecture trials.",
 "download_zip": "https://synthcore.org/download/1090",
 "formulas_latex": [
  "$$\\tau_t=\\operatorname{Quantile}_{1-\\alpha}\\left(\\{r_{t,i}\\}_{i=1}^{n}\\right),\\qquad r_{t,i}=h_\\phi(z_{t,i},c_i).$$",
  "$$\\operatorname{continue}(z_t)=\\mathbf 1\\{h_\\phi(z_t,c)\\le\\tau_t\\},\\qquad \\Pr(\\text{failure}\\mid\\operatorname{continue})\\lesssim\\alpha.$$",
  "$$\\operatorname{compute\\ saved}=1-\\frac{\\text{number of fully decoded and evaluated candidates}}{\\text{number without early rejection}}.$$"
 ],
 "id": 2762,
 "implementation": "Use the architecture mutation process from the first idea and expose the denoiser hidden state z_t at checkpoints such as t divided by T in {0.75, 0.50, 0.25}. Train a lightweight risk head h_phi using completed trajectories labeled 1 for failure and 0 for architectures that compile and meet resource and accuracy targets. Pseudocode is: for each checkpoint t, risk = risk_head(z_t, context); if risk \u003e tau[t], reject and resample; otherwise continue denoising. Compute tau[t] as a split-conformal quantile on calibration trajectories that were not used to fit the risk head. Maintain separate thresholds for different latency budgets, parameter budgets, and datasets. The diffusion kernel and intermediate-state representation come from the paper's mechanism; the risk head, calibration quantiles, and failure labels are estimated empirically from a pilot search. External compilation, latency measurement, and validation accuracy remain the final labels, so the monitor cannot silently replace them. First experiment: use NAS-Bench-201 or a synthetic graph checker where full evaluation is a table lookup. Compare no rejection, an uncalibrated classifier, and split-conformal rejection at alpha = 0.05 and 0.10. Record full-evaluation calls, accepted-candidate failure rate, and best accuracy at fixed compute. The predicted signature is an accepted failure rate near or below the selected alpha under exchangeable search conditions, with at least 20 percent fewer expensive evaluations if the risk score ranks failures usefully. Deliberately shift the architecture distribution afterward to measure degradation when exchangeability fails.",
 "math_summary": "Let z_t be an intermediate diffusion state at timestep t, h_phi(z_t,c) a risk score estimating the probability or severity of eventual failure, and C_t = {r_(t,1), ..., r_(t,n)} calibration scores from held-out completed trajectories. For target miscoverage alpha, define tau_t as the empirical (1-alpha)(n+1)-quantile of C_t. Reject a trajectory when h_phi(z_t,c) exceeds tau_t, or when a lower confidence bound on eventual success falls below the required budget. Under exchangeability between calibration and search trajectories, the marginal probability of an unflagged failure is controlled approximately by alpha. The controller should be evaluated by the rejection-risk versus compute tradeoff rather than predictive accuracy alone.",
 "math_tags": [
  "statistics",
  "probability",
  "stochastic-processes"
 ],
 "ml_areas": [
  "optimizer",
  "training-dynamics",
  "cnn",
  "transformer"
 ],
 "paper": {
  "arxiv_id": "2608.27649",
  "arxiv_url": "https://arxiv.org/abs/2608.27649",
  "summary_what_math_gives_to_ml": "The paper provides a concrete transition mechanism that differs from ordinary generative sampling: partially re-noise an existing feasible artifact and reverse-denoise it, producing a mutation kernel anchored at the parent design. This preserves learned structural dependencies while enabling controlled movement away from the observed corpus, with mutation strength set by the re-noising fraction. The mechanism transfers naturally to neural architecture search, where diffusion operates on graph or token encodings and external compilation, shape, latency, and accuracy checks remain authoritative. A second transferable component is trajectory-level conformal rejection, which can stop expensive evaluations when an intermediate candidate enters a statistically calibrated high-risk region.",
  "title": "From Generation to Discovery: Diffusion Mutation Kernels for Circuit and Physical Design",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 5,
  "novelty": 7,
  "usefulness": 7
 },
 "solves": [
  "speedup",
  "sample-efficiency",
  "scalability"
 ],
 "title": "Conformal Early-Rejection for Diffusion Architecture Search",
 "url": "https://synthcore.org/idea/2762/conformal-early-rejection-for-diffusion-architecture-search",
 "verification": {
  "peer_reviewed": false,
  "stage1_mechanism_check": {
   "worked": false,
   "confidence": 9,
   "verdict": "Built a synthetic diffusion/NAS trajectory simulator with checkpoint-wise split-conformal thresholds, authoritative final labels, exchangeability-shift tests, and baseline comparisons. The finite-sample conformal score-exceedance prediction was confirmed closely across calibration sizes. However, the core conditional failure claim was not confirmed: at alpha=0.10, conformal rejection saved 17.13% of evaluations but accepted-candidate failure remained 39.98%, far above 10%, and increasing score separability produced no meaningful monotonic improvement.",
   "metrics": {
    "baseline": "No rejection: 1000.0 full evaluations per 1000 candidates, accepted failure 50.25%. Uncalibrated threshold: 357.55 evaluations, accepted failure 2.13%, but it rejected 64.25% and was not calibrated.",
    "idea": "Split-conformal alpha=0.10: 828.75 full evaluations per 1000, 17.13% compute saved, accepted failure 39.98%. Alpha=0.05: 905.14 evaluations, 9.49% saved, accepted failure 45.03%. At alpha=0.10, observed calibration exceedance for ncal 20, 50, 100, 500 was 0.0938, 0.0993, 0.0989, 0.0999 versus predictions 0.0952, 0.0980, 0.0990, 0.0998. Under test shifts 0, -0.25, -0.5, and -1.0, accepted failure was 39.92%, 47.99%, 55.35%, and 68.73%."
   },
   "how_to_run": "python3 conformal_early_rejection.py",
   "files": [
    "conformal_early_rejection.py",
    "results.json",
    "experiment_report.json",
    "run_output.txt"
   ],
   "limitations": "This is a synthetic trajectory checker rather than NAS-Bench-201 or a real diffusion architecture mutator. It does not measure GPU training, compilation, real latency, or wall-clock savings, and it uses a simple hand-designed risk score rather than a trained neural risk head."
  },
  "status": "mechanism_failed",
  "status_label": "Mechanism failed",
  "updated_at": "2026-09-02T00:11:27",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)"
 }
}