For most reporting, root-cause and structured-to-text work, build the templated generator first and make an LLM explanation layer prove it is worth adding. Templates filled from model outputs carry strong faithfulness guarantees by construction. An LLM adds fluency, but in one controlled study its narratives raised the confidence of LLM judges acting as non-expert readers without making them more accurate. So settle the choice with a short feasibility assessment that measures three things against a template baseline: faithfulness to the source data, cost and latency at your volume, and the validation burden the narrative creates downstream. Where a template measurably falls short, one published hybrid does analytics outside the model, checks the neural output and falls back to a template.
It also helps to stop using “generative” to mean “LLM”. Generative AI covers models that produce images, audio and 3D assets as well as text, and large language models are one family within it. A template-based generator and a fine-tuned small language model both produce text from data. The question worth asking is which generator fits the report, not whether to “use generative AI”.
Five levels between mail merge and a written narrative
Ehud Reiter (2016) proposed dropping the “NLG versus templates” binary and talking about levels of sophistication instead. Level 1 is fill-in-the-blank, like Word mail merge. Level 2 adds scripting or business rules. Level 3 adds word-level grammar such as plurals and “a” versus “an”. Level 4 builds sentences dynamically from a representation of their meaning, so edge cases come out sensibly. Level 5 plans the whole document. Reiter notes that everyone calls Level 1 a template and Level 5 NLG, and there is no agreement about the levels in between.
Level 3 and 4 functions make grammatical, sensible text easier to generate, so try them before reaching for an LLM.
Can SHAP or feature-importance templates meet a plain-language explanation requirement?
Yes, as long as the set of scenarios is bounded. A template that names the top features, their direction and rough size cannot state a feature effect absent from the attribution table. The TempLM study (ACL Findings, 2023) puts the trade-off plainly: classic template systems give strong guarantees of faithfulness at the cost of fluency, while pretrained language models are known to produce unfaithful content. In that study, distilling a fine-tuned BART model into a template-based generator cut its out-of-domain unfaithfulness rate from 83% to 0%. The reason is structural: the output can only be a template filled with input data, so if no template contains a hallucination, no output can either. The authors say openly that they give up linguistic diversity to get that guarantee.
The case for LLM narratives is real, though. In the SHAPstories work (2023), users in a credit-scoring setting answered comprehension questions correctly significantly more often with LLM narratives than with SHAP values alone. Over 92% of non-data-scientist respondents in that work found the narratives easy to interpret. A secondary summary of SHAP-to-narrative user studies reports that non-technical participants identified the top driver correctly 85% of the time with LLM narratives, against 55% with raw SHAP. But look at what those numbers compare. They pit a narrative against raw SHAP output, partly through survey judgements. Both compare the narrative with SHAP alone, not with a template.
A controlled 2026 study tested LLM-generated explanations over an XGBoost and SHAP forecasting pipeline. In it, the narrative added nothing beyond the structured data it described: the key contrast gave an odds ratio of 0.83 with p = .53. Placebic narratives, which described a different instance entirely, raised confidence about as much as the real ones. So the confidence gain came from text being present, not from what the text said. On out-of-distribution inputs, the narrative halved detection of unreliable predictions, from 30.0% to 15.0%. The study used LLM judges prompted as non-expert users and a low-stakes domain. Its authors say generalisation to healthcare, credit or legal settings is not established, but they judge the direction unlikely to reverse and the consequences likely to be larger. A separate 2026 user study points the same way: reasoning traces and post-hoc explanations increased user acceptance of LLM answers whether or not those answers were correct.
Fluent text is a credibility cue, so an explanation layer has to be judged on what users get right with it, not on how well it reads.
When does a template library cost more to maintain than an LLM layer?
We would not trust a breakeven template count carried over from another product. The threshold depends on how many distinct scenarios your outputs produce, how often the underlying model changes, and how much variation in wording your readers need. The TempLM study shows where hand-written templates strain. In the TempLM study, human template writers typically ran out of cognitive capacity after reading about 5 to 20 examples, and they missed the less common cases. Templates extracted automatically by TempLM beat the human-written ones. Very large template sets also become less interpretable to the people inspecting them.
The LLM side has its own maintenance bill. Designing and refining domain-specific prompts takes skilled people and repeated testing, and the prompts often do not generalise beyond their task. Every model or prompt change also means rerunning the faithfulness evaluation. The fair comparison sets template upkeep against prompt upkeep plus repeated evaluation, with both estimated from the same list of scenarios. Templates are losing when conditional branches multiply faster than new scenarios arrive, or when readers keep calling the text mechanical. Template-based time-series narration has long been criticised for exactly that mechanical quality. There is also a middle route: TempLM-style distillation uses a language model to write the templates and then runs only the templates.
How do we test whether users understand and trust templated text before committing engineering effort?
Run a small task-based study with the template as the baseline arm. Do it before building the generative layer, and fix the protocol before you see any results. Van der Lee et al. (INLG 2019) start from the observation that there is little agreement on how NLG output should be evaluated. That is a good reason to set your criteria, rater instructions and analysis plan in advance. A workable sequence:
- Sample real cases, including predictions known to be wrong and inputs outside the training distribution, because that is where narratives have been shown to mislead.
- Prepare four arms: structured output only, the template, the LLM narrative, and a placebic narrative from a different case.
- Report task outcomes, such as flagging an unreliable prediction, alongside quality ratings.
- Record confidence separately and check whether it tracks correctness. In the 2026 study it tracked the presence of a narrative instead.
- Ask about preference. In the TCube expert review (n = 21), 76.2% of experts chose TCube over templated narratives for deployment, while those who chose templates cited the needs of mission-critical data.
In a pre-registered study of 759 US participants, contextualised LLM explanations improved perceived transparency and trust in movie recommendations. An earlier study it cites (N = 97) found that LLM explanations beat template baselines on trust and satisfaction. A gain in trust is still not a gain in accuracy, and that is the gap the placebic arm is there to expose. Small samples of agronomists, technicians or operators may not generalise, so repeat the result in a wider rollout before fully committing. Our piece on designing task-specific evaluation harnesses for LLM deployment decisions covers how to make a comparison like this repeatable.
Build-vs-skip decision table for an explanation layer
| What the assessment shows | What to build |
|---|---|
| The scenario space can be enumerated, templates cover it, and users complete the task with template text | Templated generator at Level 2 or 3 |
| Templates cover the scenarios, readers find the text mechanical, task outcomes are unaffected | Raise template sophistication (Level 3 or 4) before any LLM |
| Scenario combinations outgrow the library, or users do measurably better on task accuracy with generated text | Constrained LLM or RAG layer with audits and a template fallback |
| The LLM narrative wins on preference only, or worsens detection of unreliable predictions | Keep templates and re-test after changes |
| Faithfulness must be guaranteed in a regulated record and the false-explanation rate is unbounded | Templates in the record; LLM text at most as a reviewed draft |
Skipping this comparison is not the neutral option. Building a generative layer without it means taking on extra model risk with nothing on record to justify it.
Can a RAG pipeline over historian data give root-cause narratives the traceability of SPC templates?
It can come close, but only if traceability is checked record by record rather than assumed from the architecture. A templated report is faithful by construction if its templates are, because every output is a filled template. A RAG narrative is only as traceable as its citations turn out to be.
TechnoLynx can build both the RAG narrative layer and its evaluation through Generative & Agentic AI R&D, starting from a labelled set of historian cases scored against your existing SPC template.
Useful numbers come from a study of LLM-assisted anti-money-laundering triage, a regulated workflow with a similar provenance problem. An LLM working without retrieval reached citation validity of only 0.62 and evidence support of 0.54. Forcing it to cite within a bounded evidence bundle raised these to 0.96 and 0.83. Adding a verifier and counterfactual checks lifted them to 0.98 and 0.88. Numerical inconsistency fell from 0.11 without retrieval to 0.06 with RAG and 0.03 with the full pipeline. Policy hallucination fell from 0.14 to 0.05 and then 0.02. That is a different domain from batch manufacturing, so read these as evidence of how much the architecture matters, not as the rates your historian will produce.
Citations alone do not prove faithfulness. The same study found that a model can cite relevant-looking passages while basing its conclusion on something else. Its evidence counterfactuals remove a cited item and check whether the rationale changes, which is a direct test for decorative citations. For historian data, a citation should resolve to a specific tag, timestamp and batch, and a verifier should reject any citation that does not exist or any value that disagrees with the record. Retrieval also has a ceiling: if the indexed records are incomplete, out of date or inconsistent, grounding can fail or introduce new errors. Generation, repair and counterfactual re-runs add latency, so the study’s authors suggest routing only ambiguous or high-risk cases through the full pipeline. Our article on constraining and auditing generative outputs in production covers the verifier and audit-log side in more depth.
What false-explanation rate should we measure before trusting an LLM narrative in a batch record?
Your quality function should set the acceptable rate. The statistics behind it then fix how large the evaluation has to be. A pilot protocol for LLM-generated care summaries at Children’s National Hospital sized its sample at 100 summaries. That gives 95% confidence that the true hallucination rate is below 3% if zero hallucinations are observed, using the Clopper-Pearson exact method. The same protocol sorts each inaccuracy as an omission, an incorrect statement or a hallucination, and grades severity as mild, moderate or severe. It also re-checks performance after 20 and 40 evaluated summaries.
Worked example: sizing the labelled evaluation set
Assume n independently sampled historian cases, each reviewed against source records, with zero false explanations found. The one-sided 95% Clopper-Pearson upper bound on the true rate is then 1 − 0.05^(1/n).
| Cases reviewed (zero errors) | Upper bound on false-explanation rate |
|---|---|
| 100 | 1 − 0.05^(1/100) ≈ 2.95% |
| 300 | about 1.0% |
| 3,000 | about 0.1% |
A single observed error widens the bound, so the set has to grow to win the margin back. The cases should also represent the deviations that actually reach batch records, including the rare ones. Otherwise the bound describes an easier population than the one you will deploy on.
Has anyone validated an LLM-generated explanation in a GxP-regulated report?
Plan as if you have no accepted precedent to lean on, and treat that as an audit risk to manage. Until the false-explanation rate is bounded and documented, we would keep the templated SPC report as the record of truth. In that design the RAG narrative is a draft a qualified person reviews before anything enters the GxP record. This matches the guidance in a 2025 hallucination survey that healthcare LLM outputs serve as decision support, with detection acting as a safeguard and humans in the loop. TechnoLynx covers the assurance side of that work through its AI Governance and Trust offer. Assembling the evidence package an auditor will accept is a separate problem from the narrative choice discussed here.
Prompting a general LLM or fine-tuning a smaller model: which hallucinates less on structured-to-text?
Neither wins by default. At low volume, with output formats that change often, template-conditioned prompting of a general model is the quicker baseline. At high volume, with enough curated field-to-text pairs, a fine-tuned smaller model is the candidate for more faithful, cheaper output. When neither meets the hallucination bar, we would try a fine-tuned small model inside structural constraints next.
TechnoLynx’s LLM Selection Pack is built for this decision. It frames the faithfulness criteria and a cost model for your volume, benchmarks a template-conditioned prompting baseline against a fine-tuned smaller model and a hybrid, and delivers the choice backed by measured hallucination, cost and latency.
Prompting helps, but explicit prompts do not stop invented narratives. In a 2024 study of LLM-generated SHAP narratives, a detailed prompt that told the model to respect feature rank and sign consistently beat a minimal prompt on faithfulness metrics. Templates with placeholders inside the prompt are a recognised way to reduce hallucination. Even so, LLMs can still invent feature-effect narratives despite explicit prompts, which is why continuous logging and random audits are advised. The same 2024 study found a worrying pattern. When fed deliberately manipulated, implausible SHAP signs, models “corrected” the sign in up to 30% of cases, against a naive expectation of around 4%. For an explanation layer this is the worst failure to have: the narrative stops describing the model at exactly the moment the model is doing something odd. Zero-shot baselines are weak on novel data. Kasner and Dušek (2024) found that more than 80% of zero-shot outputs from open LLMs on novel data-to-text inputs contained at least one semantic error.
Fine-tuning gives you a different lever. A 2024 study across BLOOM, OPT and Llama 2 families found that larger models within a family gained readability and informativeness but tended to lose faithfulness. Smaller models coped better when sources and references diverged, and the authors recommend smaller models where faithfulness is critical. In a 2026 study, Llama 3.1 8B was fine-tuned with LoRA on 219 curated examples. It reached 100% JSON structural validity, against 85–95% for zero-shot approaches, and 83.0% human-validated accuracy. Large models are not hopeless at following instructions: GPT-4 reached a 0% slot error rate on ViGGO with selected in-context examples, although the same authors still regard supervised training as essential for data-to-text. Fine-tuned models struggle to generalise to domains outside their training data. A curated set that misses some listing or report types will reproduce that gap.
What are the cost and latency differences at our volume?
Only a measurement at your own volume answers that. The published figures show the shape of the trade-off, not your numbers. The 2026 hybrid study reports about 2 seconds per inference on a single NVIDIA A100, at $0.013 per evaluation against $0.025–$0.055 for proprietary APIs. A 2023 customer-service case study priced a distilled GPT-2 response on an A100 at ¢0.0011 and a GPT-3 Davinci response at ¢1.10. Even though GPT-3’s suggestions were accepted somewhat more often, for the case-study brand the yearly difference in net savings came to only about $3,000. The same case study found that low traffic makes upfront R&D cost the main factor, while high traffic makes inference cost dominate. Reiter (2023) noted that a GPT-4 response could cost $0.10 or more and take 5 to 10 seconds. On the self-hosted side, running Mistral 7B needs at least 16 GB of VRAM unless quantised or served remotely. Each of these numbers comes from one study’s own setup. Re-measure whenever volume changes materially.
Can we fine-tune a small model and still constrain it with a structured template?
Yes, and a hallucination survey rates hybrids of complementary techniques as most promising. The 2026 hybrid study describes the principle: give contextual reasoning to the neural model and constraint enforcement to a deterministic layer, which corrects rule-bound fields at no cost. Reiter (2023) describes a version for data-to-text. The neural system handles only the language, analysis and content selection happen elsewhere, and a checker discards any suspect output and falls back to a template generator. He reports that practitioners he spoke to favour doing analytics and insight creation outside the LLM and passing the results in as input. Verification can be cheap. A verification-and-correction loop on a fine-tuned T5 cut the slot error rate on E2E from over 2.5% to almost zero. It worked even though over 90% of the errors its exact-match checker flagged on ViGGO were false positives. Combine layers that fail in different ways. Stacking methods that share a failure mode, such as several confidence-based checks, does not add robustness.
Do vet reviewers need generated text, or do calibrated classifier scores already meet the workflow need?
Assume they might not need it until there is documented evidence that calibrated probabilities, perhaps wrapped in templated sentences, fail the reviewers’ actual task. The nearest comparisons come from time-series experts and movie recommendations, and they measure preference and perceived trust rather than diagnostic accuracy.
The minimal non-generative version is a Level 1 or 2 template. It maps each finding’s calibrated score to a fixed phrase for its probability band, names the finding, and states what the reviewer should check next. Every word can be traced to the score and the band definition. The gate then runs in order: classifier scores alone, then scores with templated summaries, and a generative narrative only if the templated version demonstrably fails the vets’ stated need in a comparison like the study above. Two cautions apply to that comparison. Templated text can read as mechanical even when it does the job, so measure task outcomes alongside acceptance. And document and version any generative component from its first release, with explanations linked to each version.
Who can help us decide whether our model outputs need an LLM narrative at all?
TechnoLynx helps teams decide whether a narrative layer should be a template, a prompted general LLM, a fine-tuned smaller model or a constrained hybrid. The decision rests on measured hallucination, cost and latency at the team’s own volume, delivered through the LLM Selection Pack, with Generative & Agentic AI R&D building the option that wins. Bringing in a specialist pays off when the comparison has to stand up to a reviewer or auditor, or when fine-tuning and hybrid constraints are new to the team. When choosing a partner, ask for a measured comparison against your own template baseline on your own cases, not a demonstration of fluent output.
This kind of engagement does not suit everyone. It does not replace regulatory or GxP sign-off, which belongs with your quality function and, where needed, regulatory specialists. It is more than you need if your template already passes the task test, or if you only want a single prompt written. Realistic alternatives are an in-house evaluation built on the procedure above.
FAQ
Is there an AI that is not generative?
Yes. Classifiers, regressors and forecasting models predict labels or values rather than generating content. A template-based report generator produces text without any generative model at all.
What does GxP regulated mean?
GxP is the family of “good practice” regulations in life sciences, such as good manufacturing, laboratory and clinical practice. Systems that create or hold GxP records must be validated for their intended use.
What is the main advantage of using a RAG system?
RAG retrieves relevant external documents at inference time, so the model can ground its answer in that evidence rather than only in what it learned in training. The grounding only goes as far as the quality and coverage of what is retrieved.
