When we tell IR managers that Orbiton calibrates to your company's voice from prior filings, the next question is almost always some version of: "What does that actually mean?" It is a fair question. The phrase "AI voice calibration" gets used loosely, and IR professionals have every reason to be skeptical about what a tool is doing with sensitive investor documents they feed into it.
This is my attempt to explain the technical process in terms that are specific enough to be useful, without oversimplifying or overstating what the extraction step can and cannot achieve.
What the calibration input actually is
The starting point for voice calibration is a corpus of your prior investor communications. For most listed mining and energy companies with 12-18 months of available documents, this means quarterly production reports, earnings press releases, MD&A sections from annual reports, and earnings call transcripts. Investor day presentations add useful signal when available.
The quality of the calibration output is proportional to the quality and consistency of the input. If your company has had multiple different IR managers writing quarterly reports over the past two years, the corpus will contain multiple distinct voices. The model will average across them rather than learning any single coherent style. Companies with stable IR teams and consistent writers produce better calibration results.
The documents are processed as plain text with structural metadata preserved where possible. Section headers, footnote designators, and tabular number blocks are tagged differently from narrative prose so the model can learn patterns within each type of content separately. You do not want the model learning your commodity price sensitivity disclaimer as part of your narrative voice.
What the model extracts
Voice calibration is not summarization and it is not fine-tuning a model on your documents. It is closer to constructing a profile of stylistic features that characterize your writing. The features that matter most for IR text fall into several categories.
Lexical preferences: which terms does your company use for specific concepts? "Attributable gold production" versus "payable gold production" versus "total gold poured." "All-in sustaining costs" versus "AISC." "Revenue" versus "net revenue" versus "realized revenue." These are not interchangeable. A gold miner that has used "payable gold ounces" consistently for three years in filings cannot switch to "produced ounces" in a press release without creating a question from any analyst who reads closely.
Sentence-level patterns: what is your typical sentence length in narrative sections? Do you use passive voice for cost variance explanations or active voice? Do you open with the result or the explanation? For example: "Production of 42,000 oz fell below Q1 guidance primarily due to lower mill throughput" versus "Lower mill throughput in the period reduced production to 42,000 oz, below Q1 guidance." These carry the same information but signal different register habits.
Structural conventions: how do you order sections? Do you present production results before financial results or after? When you give guidance, do you state the range before explaining the basis or after? These structural patterns are remarkably consistent within a single company over time and remarkably inconsistent across companies.
Qualifier vocabulary: what hedge language does your company use around forward-looking statements beyond the required legal boilerplate? "Remains on track" versus "is expected to remain" versus "should remain." "Approximately" versus "roughly" versus nothing at all. These small word choices accumulate into a recognizable register.
What the model cannot learn from documents alone
This is the part where it is important to be direct. Voice calibration from prior filings captures stylistic patterns in approved, published text. It does not capture:
Judgment about what to say. The model can learn how your company typically phrases guidance revisions. It cannot know whether the geotechnical update from the mine site last week requires a guidance revision, or how material that information is under your continuous disclosure obligations. That is an IR judgment call that requires the actual facts, not a style pattern.
The register difference between internal drafts and published text. If you feed the model a mix of published filings and internal drafts, the internal drafts will pull the learned style toward less formal patterns. Published filings are better training material than working drafts for this reason.
Recent changes in company strategy or communication approach. If your company changed its capital allocation framework in Q1 2026 and is now emphasizing free cash flow yield over growth capex, the model calibrated on 2024 filings will not reflect that strategic shift. The calibration corpus has to be periodically refreshed and any significant strategic shifts flagged so the model can be given context about what has changed.
How the calibration is used in drafting
When Orbiton generates a draft section, it does not simply replicate phrases from prior filings. The calibration profile constrains the generation toward your company's established stylistic patterns. The output is new text that reads like your company wrote it, not a collage of prior sentences.
For the sections with the highest repetition rate, the production metrics table structure, the cost comparison language, the forward-looking statement boilerplate, the calibration has the most leverage. These sections are structurally consistent and the model produces output that requires minimal editing.
For narrative sections that depend on current facts, the model provides a structure and a stylistic frame, but the IR manager has to supply the current-period context. A variance explanation like "production was below guidance primarily due to X" requires the IR manager to specify what X was. The model knows how your company typically phrases that sentence; it does not know what X is for this quarter.
Practical implications for onboarding
When an IR team at a listed resource company onboards to Orbiton, the calibration step typically requires uploading 4-8 prior quarterly reports and 4-8 prior earnings press releases. The model processes these and produces a calibration profile that the IR manager can review. The review involves looking at a set of generated sample sentences and flagging any that do not sound like the company's actual voice.
This review step is not optional. The model can extract patterns but it cannot know when a pattern from the corpus is an anomaly (a one-off phrasing from a guest contributor or a prior IR manager who has since left) versus a genuine house style element. The IR manager's feedback on the samples corrects the profile before it is used in production drafting.
Companies that have gone through an IR head transition in the past 18 months should think carefully about the corpus period they use for calibration. If the current IR head wants to establish a new voice, it may be more productive to calibrate on a smaller sample of documents they have already written in the new role rather than averaging across the prior and current style.
A note on data handling
The practical concern that comes up most often with IR managers is what happens to the documents used for calibration. The answer: they remain in your organization's data environment. Calibration is run in a tenant-isolated context and the documents are not used to train models that serve other customers. The calibration profile lives in your account and is not accessible to other users of the platform.
This is worth stating clearly because the concern is legitimate. Public filings are already available in EDGAR and on company websites, so the calibration corpus is not confidential in the way that internal strategy documents are. But pre-release drafts and transcripts of investor calls that have not been publicly released would be. The correct input for calibration is approved, published material only.
