Norton Design Lab | July 16, 2026
A summary of the full white paper: three instruments, five runs, 1,100+ scored outputs, one production decision. The complete findings, methods, and a tested MIT-licensed reference implementation are in the paper itself.
If you produce technical writing from source material, whether that’s pharma IC training content, market analysis, or research summaries, you face two problems that compound each other.
The first is trust. Cheap AI models fabricate statistics, and the AI judges commonly used to catch fabrication approve it. In one of my evaluation runs, 27.5% of outputs (66 of 240) scored 4.5 or better out of 5 from an AI judge while failing a deterministic provenance check. That result held across three judge models from three different families. Wholesale fabrication sits above the judge ceiling no matter which judge you pick.
The second is cost. Cloud pricing looks harmless per request, but production workloads are measured in token-days. My first RAG corpus ingestion run burned roughly 500 million tokens in a single morning. At my actual blended OpenRouter rate that week ($3.79 per million), that morning prices out near $1,900. At frontier rates it clears $4,500. On my own workstation it cost electricity, under a dollar. The only reason I can quote the cloud figure as an avoided cost rather than a paid one is a $50 API spend cap that tripped mid-run and forced the switch to local hardware.
The white paper documents what happened when I stopped treating those as two separate problems.
The finding: the model is not the load-bearing component
A deterministic verification layer is. The pattern is a Python checker that extracts every numeric claim from a draft, classifies it against a known evidence pack, gates on zero violations, and feeds itemized failures back to the model for up to two repair rounds. No AI judge participates in the verdict. It costs no tokens and no license.
That checker converted every model family it touched:
Three local models running on my own hardware (gemma4-26b-a4b, gemma4-31b, qwen3.6-27b) reached the frontier bar (100% gate pass, deterministic score 5.00) across nine parity cells, at zero marginal cost. One of them, gemma4-31b, went 10 for 10 on the bare brief with nothing but the checker feeding violations back. No exemplar, no prompt engineering.
The worst fabricator tested, DeepSeek V4 Flash, produced 363 fabricated statistics in one run and a raw deterministic score of 1.02 out of 5. With repair, it converted to 9 of 10 passing at roughly 4.9, and every repaired brief ended at zero invented statistics. Fabrication, the failure mode people fear most, turned out to be the most repairable class on the board.
GLM-5.2, a budget cloud model that had posted 13 fabricated statistics on a three-brief smoke test, went 40 for 40 across four scaffolding strategies once the checker was in the loop. Its cheapest configuration produces a verified draft at $0.047 against the frontier reference’s $0.118, a 60% cost reduction per verified document.
The frontier reference (GPT-5.6) never needed the scaffold: 80 for 80 across two runs, zero repair rounds. That is what a bar is for.
Without the checker, the gap is not closable by prompting. On the hardest instrument, the frontier model passed 38 of 40 briefs while a raw local model passed 2 of 40 and fabricated 209 statistics. A 19x gap on raw prompting, closed to parity by a few hundred lines of standard-library Python.
The efficiency numbers
The most efficient local configuration delivered clean first drafts (deterministic 5.00 before any repair) at 823 tokens per draft, against the frontier reference’s 3,118 tokens for the identical verdict. That is 3.8x fewer tokens, roughly 214x lower cost if priced as cloud calls, and zero actual marginal cost on owned hardware.
One counterintuitive result worth flagging: strategy rankings do not transfer across model families. On the local models, a strict compliance prompt was the efficiency pick. On GLM-5.2, the same strict prompt tripled hidden reasoning tokens (225,134 for the cell against 75,823 for the plan-then-write strategy, all billed) and landed within 16% of frontier pricing. The cheapest path to a verified draft on a reasoning-billed model is the cheap draft plus the token-free checker, not the careful prompt. Measure, do not assume.
Why this matters even if you never run a local model
The verdicts from an AI judge depend on the judge. The verdicts from a deterministic checker depend only on the evidence pack and the code, which means they can be audited. During the program, three separate defects (a judge rubric blind spot, a dropped detection class, a one-line tag-window bug) were caught by cross-checks inside the instrument itself, and in every case the deterministic layer’s transparency is what made the catch possible. A judge’s blind spots are not findable by reading code. A checker’s are.
The production stack this evidence supports: local models for token-heavy work (RAG ingestion, contextual enrichment, high-volume drafting), a budget cloud model like GLM-5.2 where quality margin matters but frontier pricing does not, the same checker wrapped around both, and frontier models reserved for the jobs where blinded writing quality, not provenance, is the differentiator. Parity here means parity on the provenance and structure contract; blinded quality is a separate test, and a frontier-class model won it.
One prediction, on record: cheaper verified output will not shrink your AI budget. It removes the last reason to ration, which is the 161-year-old efficiency-rebound pattern the economics literature has documented since Jevons. Expect more verified output per dollar and a widening set of jobs that were previously not worth doing at all.
Get the paper
The full white paper covers all five runs in detail, including the judge rubber-stamp analysis, the collision-axis difficulty design, the fixture-integrity incident and its fix, the complete GLM-5.2 results table, and the operational case for hard API spend caps as standing infrastructure. Appendix 1 contains a tested, roughly 100-line reference implementation of the checker and repair loop, released under the MIT License. Nothing in the pattern is specific to pharma or to my hardware; if you can define what a violation is for your evidence source, you can build this in an afternoon.
Download the white paper | Questions about applying this to your content pipeline: bradley.norton@nortondesignlab.com