Tacitt
Get started

Research & methodology

How we study the way you work with AI—and explain what the evidence can support.

Explore the methodology ↓

The starting point

An output is only part of the story.

The result alone
InstructionsDecisionsCorrectionsChecksAI output

Read the available record in context

Tacitt studies the context around the result.

Instructions, decisions, corrections and checks help us understand a person’s contribution when they are captured in the record.

How to read this

This diagram explains our reading scope. It is not a score or a measured sequence. Work outside the captured record may be missing. [3]

The VECTIR framework

Six lenses. One working profile.

VECTIR brings six views of observable work together. Each lens asks a different question.

Verification
What evidence shows that AI-produced work was checked?
Expertise
What domain familiarity appears in the person’s contribution?
Context
What relevant background and material does the person bring to the task?
Taste
What standards does the person introduce and hold the work to?
Instructions
How clearly are the intended work, scope and constraints stated?
Resilience
How does the work progress when it encounters obstacles?

A combined view, with limits

The update combines these lenses into a 0–100 score. It summarises observed work; the research does not establish six independent abilities or validate prediction of job performance. [2]

Profile

How you work with AI.

Six lenses and a combined view of the work you shared.

Context

What kind of work.

Session Depth and Business Domain describe task demand and subject matter. They are not extra VECTIR axes.

Output estimate

What was produced.

AI Productivity is a separate estimate of replacement effort—not business value or proven time saved.

The research material

The sample behind this update.

A frozen research snapshot, dated 14 October 2026. Reference data and a small internal comparison answer different questions.

116

People in the reference snapshot

12

Internal team members compared

Not every person has a readable result on every lens. The 12-person comparison is not a held-out, representative population study. These counts are not the size of the current production reference pool. [1]

The research process

How we read—and test—the work.

A high-level view of the methodology. Each lens uses the evidence relevant to its question.

  1. 01

    Separate

    Whose contribution is visible?

    Distinguish the person's writing from copied material and system wrappers.

  2. 02

    Read

    What does the context reveal?

    Examine instructions and decisions, alongside relevant AI actions and checks.

  3. 03

    Compare

    Does the interpretation hold up?

    Test consistency, bias and relationships with other measures in a stated sample.

  4. 04

    Revisit

    What changes with new evidence?

    Record failed checks, correct reading artifacts and retest changed methods.

New evidence brings us back to the question

This overview is not a universal scoring pipeline. Filtering differs by lens; automated readings can miss context. [2–3]

The evidence so far

Early evidence. Clear limits.

We test repeatability, meaning and links to other measures. The current findings are encouraging in places—and incomplete.

Composite ↔ cleaned Session DepthInternal team · N = 12 · Research snapshot / 14 Oct 2026
Linear association · Pearson r0.754
Rank association · Spearman ρ0.427
0Positive association scale1

Two views of the same relationship—not independent replications. Correlation is not accuracy, causation or proof of job-performance prediction. [1]

Observed

Association with task demand.

In this small team, the research composite is associated with cleaned Session Depth. Both use the same underlying records.

Mixed

Consistency varies across lenses.

Results depend on the measure, task, record format and assessment version. The six lenses also overlap.

Not established

Prediction of productivity or job performance.

A five-person productivity pilot did not establish a relationship with the combined score. Independent outcome validation remains open.

Scope / [1–2]

These are exploratory findings from the 14 October 2026 research snapshot. They are not a validation of every later product release. [1–2]

Evidence limits

An incomplete record is not a verdict.

Evidence can support an interpretation, leave it uncertain, or be unreadable. These are different states—not levels of ability.

Supported

Enough to interpret.

The record supports a scoped finding. Its meaning depends on the task and the evidence observed.

Provisional

More context needed.

Some evidence is present, but more context could change the interpretation.

Unreadable

No sound conclusion.

The record cannot support a sound interpretation. This is a limit of the evidence.

A request to check is not proof of a completed check.

Equally, a missing check in the record does not prove it never happened. Work can occur outside the captured session, and exports can omit activity. [3]

Compare like with like.

Format

Match the record format.

Chat and agent records expose different evidence.

Reference

Name the reference group.

A result describes its tested sample—not all AI users.

Version

Match the assessment version.

A changed method needs evidence for that version.

Comparison boundaries · Research note [5]

Your data

Your data, clearly explained.

The choices you make, the copy we prepare, and the limits to know.

Selection

Let you choose what to include.

Select the projects and conversations you want in your export.

What we don’t do

Require your entire conversation history.

You can work with selected sessions.

Preparation

Prepare your selected records locally.

The preparation tool selects, cleans and redacts an export copy on your computer.

What we don’t do

Promise all processing stays offline.

Your AI assistant may still process data through its provider.

Sensitive terms

Let you choose additional terms to hide.

Choose names or terms to replace alongside the standard checks.

What we don’t do

Limit redaction to a fixed list.

Your chosen terms can cover details the standard checks miss.

Context

Preserve useful work context.

Retain relevant technical context where the record and your processing settings allow.

What we don’t do

Blanket-remove technical details.

Redaction targets sensitive details while retaining useful context.

Source records

Replace sensitive details in a copy.

Redaction uses placeholder values to retain useful context.

What we don’t do

Edit your original conversations.

Preparation works on an export copy; your source logs stay unchanged.

Review

Give you a prepared copy to review.

Inspect the result before deciding what to share.

What we don’t do

Promise automated checks catch everything.

Sensitive details can remain, so your review still matters.

Before you share

Review the prepared copy before uploading. Check your assistant provider’s data settings too.

Research notes

The context behind the evidence.

We share the research questions, evidence and limitations. Detailed scoring methods remain proprietary.

  1. [1]

    The study snapshot · 14 October 2026

    Internal VECTIR research: 116 reference people and 12 team members. The depth chart uses the research composite on 12 people: Pearson r = 0.754; Spearman ρ = 0.427. It is a small, same-record comparison—not an independent job-performance test.

  2. [2]

    Reliability and meaning · ongoing validation

    Consistency varies across lenses and ways of testing. Some validation checks fail, and the lenses overlap. The five-person productivity pilot did not establish that the research composite predicts productivity. Broader independent validation remains open.

  3. [3]

    Interpretation · limited by the shared record

    Missing activity is not proof of low ability. Partial exports, short records and work outside the session can limit interpretation. Task demand, business domain and output-effort estimates describe different aspects of work.

  4. [4]

    Published benchmarks · context, not certification

    Published assessment research informs how we interpret evidence. It does not certify Tacitt. Our small internal depth comparison is not directly comparable to studies predicting job performance.

  5. [5]

    Comparisons · reference and release boundaries

    Compare within the tested record format, reference group and assessment version. Research snapshots and production pools differ. The figures on this page describe the dated study; they do not validate every later release.

Page reviewed / 19 Oct 2026 · Study figures / 14 Oct 2026

Get your profile

Get your
Tacitt profile.

Run the Tacitt skill to prepare your AI work records. Review what you share, then sign up to create your profile.

Get my profile ↗

Opens skill setup and signup instructions