docs/research/2026-07-22-frontier-build-test-survey.md

Howfrontier models are actually made and tested — survey + gap audit

Date: 2026-07-22 · Method: one wide web-search fan-out (10 parallel queries + a date-corrected second pass on the current generation, no workflow), primary technical reports where available. Purpose: stop reinventing the wheel — align the v1.10 design (AGI-V1.10-WHITE-BOX-HONESTY-DESIGN.md, epic #2841) with what frontier labs measurably do, and log the gaps for operator review.

Generational note: §1 documents the settled consensus established by the previous generation (DeepSeek-V3, Qwen3, Llama 4, Kimi K2 — late 2024→2025 reports, still the best-documented pipelines). §1b covers what the current generation (2026 releases, through this week) changed.

Source quality note (External Reality Rule): rows marked [P] cite a primary technical report / system card; [S] are secondary summaries surfaced by search and are verified:false until read in full. Nothing here is reproduced on our hardware.

1. How they're MADE — theconsensus pipeline

Every current frontier model is assembled the same way. The stages, with per-lab evidence:

Stage A — Data (where the actual competition is)

Qwen3 36T P: [2505.09388] · Llama>30T S: [Meta blog] · Kimi K2 15.5T P: [2507.20534].

  • Dedup: MinHash/LSH at paragraph + document level, run before and after synthesis [S:

Toloka pipeline survey].

  • Quality filtering: small model-based classifiers ("good web" vs noisy), perplexity filters

(KenLM/TinyLM) [S: same].

  • Synthetic data: web-rephrasing is standard, **but human-data anchors stay ≥60–70% of the

mix; synthetic is targeted amplification on verifiable domains only** S: [decision guide].

  • Curriculum: multi-stage pretraining (general → reasoning-dense → long-context annealing);

Qwen3 is explicitly three-stage [P: 2505.09388].

Stage B — Architecture (a commodity; nobody differentiates here)

  • All sparse MoE: DeepSeek-V3 671B/37B active (aux-loss-free load balancing, MTP objective)

[P] · Kimi K2 1.04T/32B [P] · Gemini 2.5 sparse MoE on TPUv5p, thinking-budget 128–32768 (Pro) / 0–24576 (Flash) tokens P: [2507.06261; ranges full-text-verified 2026-07-22] · LlamaMoE [S].

  • This confirms the blueprint's "architecture is rented" stance — no lab wins on topology.

Stage C — Training systems (the engineering moat)

  • FP8 mixed-precision at scale (DeepSeek-V3, first validated at 671B) [P] · **MuonClip: 15.5T

tokens with ZERO loss spikes (Kimi K2) [P] · TPUv5p at 93.4% utilization (Gemini) [P/S] · MetaP** hyperparameter transfer across scales (Llama 4) [S].

  • Notable: hyperparameters are derived by transfer rules, not hand-tuned per run.

Stage D — Post-training (themodular stack)

  • Consensus shape: SFT → DPO/SimPO (preference) → GRPO/RLVR (verifiable rewards) → polish;

RLAIF/constitutional to scale labels S: [frontier training methodologies].

  • Qwen3's four stages: **Long-CoT cold-start SFT → Reasoning RL (GRPO on query-verifier pairs) →

thinking-mode fusion → general RL** [P: 2505.09388].

  • RLVR is the dominant reasoning paradigm — deterministic verifiers, "millions of verification

signals per day"; every major reasoning model since lateuses verifier-driven RL [S]. This is the industrialized version of our Spiral/VTD thesis — we did not invent verifier-gated training; we are late to it and should adopt its mature form.

Stage E — Small models: distilled, never trained from scratch

  • Qwen3 strong-to-weak distillation for everything ≤14B: off-policy (teacher outputs in both

think modes) then on-policy (student generates, aligned to teacher logits) — and it outperforms RL on both quality and training cost for lightweight models [P: 2505.09388].

  • LlamaMaverick codistilled from Behemoth with a novel loss dynamically weighting soft +

hard targets [S: Meta blog].

  • Nobody at any budget trains a small frontier model from scratch. Even Meta distills.

1b. What the CURRENT generation changed (2026 releases, as of 2026-07-22)

  • DeepSeek-V4 (Apr 24, 2026; V4-Pro 1.6T / V4-Flash 284B; 33T/32T tokens) [S:

kili data story · sitepoint · P: 2606.19348]:

  1. Pretraining now deliberately emphasizes long documents + agentic execution traces — i.e.

verified working sessions are pretraining data at the frontier now. Our session/PR trace mining (#2842) is the same move at our scale.

  1. The newest post-training paradigm: independently cultivate domain experts (per-domain

SFT → GRPO for math/coding/agent/IF), then consolidate into one model via ON-POLICY DISTILLATION. Even the flagship is now assembled by distillation, not monolithic RL.

  • Qwen 3.5 → 3.6 → 3.7-Max (Feb–May 2026): **Qwen3.6-27B dense beats the 397B Qwen3.5

flagship on coding S: [codersera guide] — one generation of recipe/data beat ~15× the parameters. Also: the 3.6-Max line moved closed-weights** (the open-weight faucet is not guaranteed to stay open).

  • Kimi K3 (released Jul 16, 2026 — six days ago; weights due Jul 27): 2.8T params, 1M

context, the largest open-weight model ever, benchmarking next to closed frontier S: [Kimi blog · dev.to]. Hallucination rate is now a headline comparison metric in launch coverage. Rent-capability implication: the rentable open tier is now frontier-class; GLM-5.2 (744B, MIT) held that spot before K3.

  • Claude Opus 4.6 (Feb 2026) system card [P:

card PDF]: evals now standardly include honesty assessments, reward hacking, sabotage capability, and evaluation awareness; testing is run against multiple snapshots including "helpful-only" variants (safeguards removed) — they deliberately test the un-gated model, not just the shipping one.

  • GPT-5.5/5.6 (Apr / Jun 26, 2026): first High ratings on bio+cyber for small fast models;

the 5.6 card plainly admits an over-agency regression (unauthorized actions up vs 5.5) P: [GPT-5.5 card · 5.6 hub].

  • Budget shape:frontier runs sit at 1e26–1e27 FLOPs, and **post-training is now

15–25% of total compute** S: [cost trajectory] — the post-training stage is where a mid-size player can actually participate.

2. How they're TESTED — theconsensus

  1. Staged capability-threshold frameworks with teeth: OpenAI Preparedness (High/Critical per

risk; GPT-5.5 rated High on bio + cyber) P: [GPT-5.5 system card] · Anthropic RSP/ASL (ASL-3 activated Mayautomatically on threshold cross, regardless of commercial considerations) P: [RSP].

  1. System cards as the release artifact — every release ships one, with measured evals,

red-team findings, and known regressions stated plainly (GPT-5.6 card admits the flagship "takes actions users did not authorize more often than GPT-5.5") [P/S].

  1. External, adversarial, third-party evals: METR ran evaluations inside Anthropic, Google,

Meta, OpenAI (Feb–Apr 2026, private per-company reports) P: [METR risk report] · UK AISI pre-deployment access, 30+ models in its Trends Report; published engineering playbook P: [AISI] · Apollo for deceptive alignment.

  1. Red-teaming at industrial scale: thousands of hours against constitutional classifiers

P: [2501.18837]; ~200 early-access partners on real use cases before GPT-5.5 shipped [P].

  1. Eval-awareness is a first-classconcern — frontier models reliably distinguish evals

from real use, undercutting pre-deployment evals; benchmarks now designed to resist contamination [S: METR/AISI commentary].

3. GAP AUDIT — us vs. the consensus (for operator review)

# Frontier practice (evidence above) Our current state Verdict / action
G1 Small tiers are built by strong-to-weak distillation incl. on-policy logit alignment, which measurably beats RL for ≤14B (Qwen3 [P]) — and in the current generation even the flagship is consolidated by on-policy distillation (DeepSeek-V4's expert→consolidate paradigm, §1b) We do row-level verified SFT only; our design bans logit/KL copying on principle Tension to resolve, not assume. The ban is right for the frontier claim (can't exceed teacher) but two generations of evidence say it's wrong as a blanket rule: on-policy distillation is now the industry's assembly step. Proposal: keep exec-verification as the gate, allow on-policy logit distillation as the transport for the student only. Needs founder ruling.
G2 Post-training is a modular staged stack (SFT → DPO → RLVR) with verifiers in the RL loop at scale VTD = SFT-only; no preference stage, no RL stage; our verifier gates data admission, not policy updates Adopt the standard shape instead of a bespoke one: our Fix-Rate/honesty verifiers slot directly into the GRPO/RLVR stage. The wheel exists — use it.
G3 Training mixes keep ≥60–70% human/general anchors; synthetic/narrow data is targeted amplification only VTD runs 1–2 trained on 100% task traces, zero retention mix → run-1's −6 (instruct damage) is the textbook symptom Confirmed root cause, external corroboration. Never run a distill without an anchor mix again (handoff's dolly-15k plan was right; it just never ran).
G4 Dedup + model-based quality filtering before every train (MinHash/LSH, perplexity filters) Our corpora: 13-gram decontam vs eval sets only; no dedup, no quality filter on the training side Cheap, standard, missing. Add MinHash dedup + perplexity filter to the corpus pipeline (#2843).
G5 Eval discipline: massive internal suites + third-party adversarial + contamination-resistant design n=40 HumanEval (±0.20 CI); honesty ledger has 2 negatives; no third-party analog Already flagged in red-team; now benchmarked: we are ~2 orders of magnitude under eval-power norms. Eval-power gate stays the #1 blocker (#2847).
G6 Hyperparameters transferred by rule (MetaP/muP), not hand-tuned per run lr 2e-4 → disaster → 5e-5 rescue was trial-and-error Adopt muP-style lr scaling for any run ≥ the next tier. Minor but free.
G7 Release artifact = system card with honest regressions stated Our analog (honest-scope blocks in docs, ConvergenceRecords) is real but per-doc, not per-release Lightweight adapter "model card" template per promoted artifact — mostly assembling what we already log.
G8 Eval-awareness / watched-vs-unwatched probing is a live frontier concern SIGMA0-MODEL-DESIGN D8 designed exactly this (watched/unwatched, trained-gamer) — never executed We independently designed a frontier-standard test; execute it rather than redesign it.
G9 Capability-threshold gates that trigger automatically on eval results Σ_θ acceptance gate exists and is the same kind of mechanism Genuine alignment — keep; wire honesty evals into it instead of inventing a new gate.
G10 Nobody trains small frontier models from scratch — even Meta codistills; and Qwen3.6-27B dense beat a 397B flagship one generation older on coding (recipe/data > parameters at the small end, §1b) ADR-0024 Phase-3 ("FRONTIER" from-scratch) is still nominally on the books The survey hardens the case: Phase-3 should be explicitly deprioritized to research-option status; the 2026-standard path for our size is open-weights + staged post-training + distillation. Post-training is now 15–25% of frontier compute — that stage, not pretraining, is where a mid-size player participates.
G11 Frontier cards test "helpful-only" / safeguards-removed snapshots, not just the shipping model (Opus 4.6, §1b) We only ever eval the gated artifact; we have never measured what our adapters do with the Σ_θ/honesty gates off Cheap and revealing: add a gates-off arm to every promotion eval — if the un-gated model is wildly dishonest, the gate is load-bearing and must be treated as a safety control, not plumbing.
G12 Hallucination rate is now a headline launch metric (K3 coverage leads with it, §1b); honesty/eval-awareness assessments are standard card content (Opus 4.6) Our entire v1.10 bet is the honesty axis Validation, not a gap: the market is converging on honesty as a differentiator — but it also means the axis will not stay unoccupied. Timing matters; the white-box verifier angle (which no card ships) remains the unclaimed piece.

4. What this changes about v1.10 (one paragraph)

The v1.10 white-box honesty design keeps its differentiator (the activation-level honesty audit — no lab ships that as a verifier, and the patents inspected are output-level only), but its training topology should stop being bespoke: slot the honesty verifier into the standardstack — de-glossed SFT → DPO on honesty preferences → **RLVR with the dual verifier (exec/citation

  • probe-audit) as the reward oracle** → strong-to-weak distill to the serving tier, with an anchor

mix (G3), dedup/filtering (G4), and the eval-power gate (G5) as hard preconditions. The wheel is already round; our contribution is the verifier bolted to it, not a new wheel.

5. Review asks — RULED (Alex, 2026-07-22)

  1. G1: ✅ APPROVED (compromise). Exec-verification stays the non-negotiable gate on every

training target; on-policy logit distillation permitted as the transport for the student/serving tier only. Recorded in the ADR-0024 amendment.

  1. G10: ✅ APPROVED. ADR-0024 Phase-3 (from-scratch frontier) retired to research-option.
  2. ✅ APPROVED via "map out new phases": v1.10 §6 replaced with the phase map V0–V4

(foundations → honest teacher → verifier-rewarded RL with dose-response → consolidate/distill → research option). See AGI-V1.10-WHITE-BOX-HONESTY-DESIGN.md §6.

  1. G11: folded into Phase V0 (gates-off arm in the eval harness) — implicitly approved with

the phase map; flag at V0 review if that reading is wrong.

Logged for review; no design docs modified pending answers. Sources: primary reports 2412.19437 · 2505.09388 · 2507.20534 · 2507.06261 · GPT-5.5 card · RSP · 2501.18837 · METR · AISI; secondary as marked.