Methodology · 15 min read
The Textimony method for source-linked message evidence.
Textimony begins with the supplied export, keeps participant assumptions visible, separates direct measurements from software candidates, and gives a reviewer the context needed to make their own decisions.
This page explains how Textimony validates a supported two-person message record, records uncertainty, surfaces review candidates, and keeps analysis output connected to the supplied source.
Long messages and model windows
Long messages can exceed the input window of a configured classifier. Where a component supports windowing, the message can be evaluated in overlapping token segments and the resulting candidate stays attached to the original message rather than becoming a detached quotation. Window size and overlap affect what a model can see. They are implementation settings, not measures of truth. A reviewer should still read the complete message and the surrounding exchange because a phrase that looks significant inside one window may read differently in the full conversation.
What changes with record size
An empty or structurally unusable upload should stop with a clear reason. A short record may not provide enough examples for every analyzer, so an abstention or zero-candidate result must not be rewritten as proof that a behavior did or did not occur. A larger record takes more time and computing resources. The workspace reports run stages and failure states, but Textimony does not promise that an arbitrary file size will complete on every configuration. The review contract remains the same: software output is a route into the record, and the underlying messages remain the material a person must inspect.
The Textimony Evidence Method
The method is a sequence of review controls: preserve the supplied export, validate its structure, confirm the two participants, state the case timezone, run available analyzers, inspect candidates in context, record reviewer decisions, and rebuild a curated report when appropriate. Federal Rules of Evidence 901, 106, and 1006 can be relevant to authentication, completeness, and summaries in federal proceedings. They do not make software output automatically admissible, and they do not replace jurisdiction-specific advice. Textimony helps organize information a lawyer, examiner, or other reviewer may use when addressing those questions. A careful report says what the source contains, what the software measured, what the software only suggested, and what a person actually reviewed.
Rules, models, and assistant features do different jobs
Deterministic checks look for defined language or structural conditions. Their strength is inspectability; their weakness is literalness. Configured classifiers can surface paraphrases and related wording, but their scores are probabilistic and can be wrong. Neither type of output determines intent or legal significance. Assistant-style exploration is separate from the classification pipeline. Generated language should be evaluated through its cited messages and should never be treated as a substitute for reading the record. If an analyzer is unavailable or has too little usable text, the honest result is an abstention or failure state, not a fabricated answer. Direct counts describe rows, dates, sender direction, or other fields in the supplied record. Rule candidates identify messages that met a documented language or structure check. Model candidates identify messages that crossed a configured model threshold. Reviewer decisions record what a person accepted, rejected, reclassified, or left unresolved.
Source references and checksums
At intake, Textimony records a SHA-256 checksum for the supplied file. Recalculating that checksum later can show whether another file is byte-for-byte identical to the uploaded file. It does not prove who created the file, whether the export was complete before upload, or whether a message was deleted before collection. Parsed messages keep available source identifiers such as a row label, message identifier, sender value, timestamp, and source artifact reference. Not every export provides every field, and Textimony does not invent a byte offset, device identifier, or delivery fact that the source does not contain. Upload checksum — Whether two available files contain the same bytes. — Authorship, completeness before upload, or legal admissibility. Source reference — Which supplied record and available row or message identifier produced an item. — Facts the export never contained. Context view — Messages around a selected candidate for human reading. — A fixed amount of context that is sufficient in every dispute.
Analysis and report output
Federal Rule of Evidence 1006 addresses summaries, charts, and calculations of voluminous admissible materials and the availability of underlying originals or duplicates. Whether a particular summary qualifies is a legal question; a chart generated by Textimony is not certified by the rule. The signed-in workflow presents run-bound daily aggregates, candidate lists, source context, and report output. Charts help a reviewer see volume and direction over time; the companion table exposes the plotted values. Candidate records and selected excerpts provide the path back into message context. The interface should not imply that every aggregate chart point is itself a legal citation. A candidate count is a count of software output. It becomes useful only when a reviewer can inspect the messages behind it and explain the decisions made.
Participant confirmation before analysis
The current analysis workflow expects one conversation with two participants. It uses sender labels, direction fields, addresses, and user confirmation to orient those participants. A group or mixed-participant record is rejected rather than reduced to two people. When the supplied structure does not support a reliable assignment, a sender can remain UNKNOWN and the workflow can request confirmation. Tone, pronouns, relationship role, or the apparent meaning of a message are not reliable identity evidence.
Message-level span candidates
When configured and available, a token-level span tagger can propose a character range inside one message. That range helps a reviewer see which words contributed to a candidate without replacing the complete message. A highlighted phrase is not a diagnosis, a finding of harassment, or proof of a multi-message campaign. Some span classes may not have a verified human-readable semantic mapping. In that situation, the class should remain generic, and a reviewer should rely on the source text rather than an invented label.
Thresholds, abstention, and limitations
A confidence score describes a model’s output under one configuration; it is not the probability that an allegation is true. Thresholds control which candidates surface, and changing a threshold changes the tradeoff between missed candidates and false positives. Textimony does not publish a universal accuracy guarantee across every category, export style, language, and conversation. Components can misread sarcasm, quotations, shared accounts, coded language, and context outside the record. When a component cannot run, its status should say so. “Could not analyze” and “analyzed but found no candidate” are different results.
A careful review workflow
The same discipline helps a self-represented user, attorney, mediator, investigator, or expert understand a large conversation without giving every role the same task. Textimony organizes the record; the professional using it decides what additional collection, authentication, redaction, or legal review is required. Start with the best supported export available. Keep an untouched copy outside the working review. Confirm that the file contains one two-person thread, resolve participant orientation, state the case timezone, wait for the run to complete, inspect candidates and ordinary context, record decisions, and then rebuild a curated report if that report fits the matter. Use a currently supported CSV, XML, timestamped text, line-delimited JSON, or EML record. Keep the uploaded file and its recorded checksum available for comparison. Confirm two-participant orientation from source fields and user knowledge. Treat rule and model output as candidates, including low-information and abstention states. Read enough surrounding messages to understand each selected excerpt. Use available exports as working material for the appropriate professional review.
Published by
Textimony. Editorial status: Current product-methodology explanation. It is not a forensic certification or a legal opinion. Updated: 2026-07-23.
Sources
Federal Rule of Evidence 901 — Legal Information Institute, Cornell Law School; Federal Rule of Evidence 902 — Legal Information Institute, Cornell Law School; Federal Rule of Evidence 1006 — Legal Information Institute, Cornell Law School; Federal Rule of Evidence 106 — Legal Information Institute, Cornell Law School; Guidelines on Mobile Device Forensics — National Institute of Standards and Technology; SWGDE Best Practices for Mobile Device Evidence Collection and Preservation — Scientific Working Group on Digital Evidence