Workspace Development · 7 min read · estimate

Audit a bundle for factual consistency

Compare source correctness and shared claims across related files, then return a precise issue log without silently changing the baseline.

By the ChatwithGPT editorial project · Prepared with AI assistance · Published October 13, 2026 · Version 1.0.0

Download Markdown

Published and revised: October 13, 2026 · Version 1.0.0

A consistency audit checks whether related artifacts preserve the same approved facts, definitions, and commitments. It can catch a deck with an old capacity, a checklist with a different deadline, or a guide that promises a feature its instructions cannot deliver. This guide produces an evidence-backed issue log; it does not silently repair or approve the bundle.

Define what the audit can establish

Choose a frozen bundle and a canonical fact set. Record exact filenames and revisions. A useful first audit compares two documents and five shared claims. Professional review adds source traceability, audience-specific omissions, derived calculations, and exported-file inspection.

Agreement is not the same as truth. Three artifacts can repeat the same wrong fact. Check both the connection to authoritative evidence and agreement among the artifacts. W3C's provenance overview supports examining how information was produced; the checklist here is an original lightweight practice, not compliance with a provenance standard. W3C: PROV Overview

1. Build a coverage map

List the shared facts and the artifacts where each must appear. Mark deliberate exclusions. If an internal staffing figure does not belong in a public handout, its absence is correct. Conversely, a required safety limitation cannot disappear because a summary is shorter.

Include names, dates, time zones, quantities, units, definitions, approval status, caveats, and promised actions. Add document-specific requirements separately. This keeps a spelling difference from receiving the same attention as a changed commitment.

2. Normalize carefully

Compare meaning, not only characters. “Half an hour” and “30 minutes” may agree; “30 minutes per session” and “30 minutes total” may not. Resolve dates into explicit calendar dates and time zones before comparing them. Preserve precision rather than treating rounded figures as exact.

For derived values, inspect the formula and inputs. If a slide says 60 seat opportunities from two sessions of 30 seats, that calculation can be correct. Calling it 60 unique visitors adds an unsupported assumption. Record that as a semantic error even when the arithmetic passes.

3. Classify issues by consequence

Use a small issue scale chosen before review:

Blocking: wrong or unsupported information could change a decision, commitment, required limitation, or safe use
Material: an important omission or inconsistency requires correction before the intended release
Editorial: style or wording differs without changing meaning

State the exact location, expected value, observed value, evidence, impact, and suggested correction. “The deck is inconsistent” is not actionable. Avoid claiming a problem is verified when you could not open the file.

4. Keep detection and repair separate

Return the issue log first. When editing is authorized, make a versioned candidate, change only the affected passages, and rerun the original checks plus checks for the repair. Do not change the canonical source simply to make the outputs agree.

If approval is missing or two authoritative sources conflict, pause that decision. An assistant's inference about the likely intended number is not permission to select it. WorkspaceOS is a conceptual review arrangement, not an automatic approval system or a way to alter hidden model behavior.

Copy-ready audit directive

Audit these exact artifact revisions: [files/versions].
Canonical facts and precedence rule: [source/version/rule].
Intended audiences: [per artifact]. Required shared claims: [IDs].
Deliberate exclusions and allowable transformations: [list].

Read each artifact and report anything you cannot inspect.
Build a fact-by-artifact coverage map. Compare names, quantities, units,
dates, time zones, status words, caveats, promises, and derived calculations.
Check source correctness separately from cross-artifact agreement.

For each issue return: ID, severity, artifact and location, observed claim,
expected claim, supporting evidence, consequence, and proposed correction.
Distinguish verified failures, acceptable differences, and unverified checks.
Do not edit, publish, approve, or change the canonical facts.
Stop dependent checks if the authoritative source or artifact is missing.
Conclude: passes reviewed checks / corrections required / blocked by evidence.

Synthetic worked example: reading-room bundle

The fictional canonical record states: two sessions, 30 seats per session, 40 minutes per session, date unconfirmed. Three draft excerpts say:

Host sheet: “Arrange 30 seats for each session”
Participant note: “Choose one of two 40-minute sessions; date to follow”
Planning slide: “60 visitors guaranteed across a 90-minute event”

The host sheet agrees on seating. The participant note preserves timing and the unknown date; capacity is absent, which is acceptable only if its contract does not require it.

Two slide issues are material:

C01 / material / slide 2, headline
Observed: 60 visitors guaranteed
Expected: capacity offers 60 seat opportunities across two sessions
Evidence: 2 × 30 seats; no unique-visitor or attendance evidence
Correction: remove the attendance guarantee

C02 / material / slide 2, timing
Observed: 90-minute event
Expected: 80 programmed reading minutes; total event duration unknown
Evidence: 2 × 40 minutes; transition and break times not provided
Correction: distinguish programmed time from full event duration

The reviewer should not rewrite the slide as an 80-minute event. That would replace one unsupported duration with another. This worked audit is illustrative; it is not a report on a live bundle or an executed model benchmark.

Checks and stop criteria

Complete the audit when every required artifact was inspected, all required shared claims have outcomes, and issues have precise evidence. For PDFs, slides, or spreadsheets, inspect the delivered rendering as well as editable content; hidden or clipped text may defeat an otherwise correct claim.

Stop release review if a blocking issue remains, input versions are mixed, or the reviewer cannot verify essential content. Save the audit against the exact bundle revision. A later edit invalidates the affected audit results until they are checked again.

Public references & scope

These references support the concepts and product guidance discussed in the guide. They do not validate the invented examples or establish this method’s performance. Check the current source before relying on a product-specific step.

PROV OverviewSource reviewed October 13, 2026

IIEN means infer intent and extrapolate needs. This site uses it as a user-directed working approach: state the likely goal, label assumptions, consider useful next steps and test a bounded choice. Inferred needs never grant permission for extra actions. It is not an OpenAI product, formal certification or claim that these exercises train or fine-tune the underlying model.

Plain text for manual copying

# Audit a bundle for factual consistency

Published and revised: October 13, 2026 · Version 1.0.0

A consistency audit checks whether related artifacts preserve the same approved facts, definitions, and commitments. It can catch a deck with an old capacity, a checklist with a different deadline, or a guide that promises a feature its instructions cannot deliver. This guide produces an evidence-backed issue log; it does not silently repair or approve the bundle.

## Define what the audit can establish

Choose a frozen bundle and a canonical fact set. Record exact filenames and revisions. A useful first audit compares two documents and five shared claims. Professional review adds source traceability, audience-specific omissions, derived calculations, and exported-file inspection.

Agreement is not the same as truth. Three artifacts can repeat the same wrong fact. Check both the connection to authoritative evidence and agreement among the artifacts. W3C's provenance overview supports examining how information was produced; the checklist here is an original lightweight practice, not compliance with a provenance standard. [W3C: PROV Overview](https://www.w3.org/TR/prov-overview/)

## 1. Build a coverage map

List the shared facts and the artifacts where each must appear. Mark deliberate exclusions. If an internal staffing figure does not belong in a public handout, its absence is correct. Conversely, a required safety limitation cannot disappear because a summary is shorter.

Include names, dates, time zones, quantities, units, definitions, approval status, caveats, and promised actions. Add document-specific requirements separately. This keeps a spelling difference from receiving the same attention as a changed commitment.

## 2. Normalize carefully

Compare meaning, not only characters. “Half an hour” and “30 minutes” may agree; “30 minutes per session” and “30 minutes total” may not. Resolve dates into explicit calendar dates and time zones before comparing them. Preserve precision rather than treating rounded figures as exact.

For derived values, inspect the formula and inputs. If a slide says 60 seat opportunities from two sessions of 30 seats, that calculation can be correct. Calling it 60 unique visitors adds an unsupported assumption. Record that as a semantic error even when the arithmetic passes.

## 3. Classify issues by consequence

Use a small issue scale chosen before review:

- Blocking: wrong or unsupported information could change a decision, commitment, required limitation, or safe use
- Material: an important omission or inconsistency requires correction before the intended release
- Editorial: style or wording differs without changing meaning

State the exact location, expected value, observed value, evidence, impact, and suggested correction. “The deck is inconsistent” is not actionable. Avoid claiming a problem is verified when you could not open the file.

## 4. Keep detection and repair separate

Return the issue log first. When editing is authorized, make a versioned candidate, change only the affected passages, and rerun the original checks plus checks for the repair. Do not change the canonical source simply to make the outputs agree.

If approval is missing or two authoritative sources conflict, pause that decision. An assistant's inference about the likely intended number is not permission to select it. WorkspaceOS is a conceptual review arrangement, not an automatic approval system or a way to alter hidden model behavior.

## Copy-ready audit directive

```text
Audit these exact artifact revisions: [files/versions].
Canonical facts and precedence rule: [source/version/rule].
Intended audiences: [per artifact]. Required shared claims: [IDs].
Deliberate exclusions and allowable transformations: [list].

Read each artifact and report anything you cannot inspect.
Build a fact-by-artifact coverage map. Compare names, quantities, units,
dates, time zones, status words, caveats, promises, and derived calculations.
Check source correctness separately from cross-artifact agreement.

For each issue return: ID, severity, artifact and location, observed claim,
expected claim, supporting evidence, consequence, and proposed correction.
Distinguish verified failures, acceptable differences, and unverified checks.
Do not edit, publish, approve, or change the canonical facts.
Stop dependent checks if the authoritative source or artifact is missing.
Conclude: passes reviewed checks / corrections required / blocked by evidence.
```

## Synthetic worked example: reading-room bundle

The fictional canonical record states: two sessions, 30 seats per session, 40 minutes per session, date unconfirmed. Three draft excerpts say:

- Host sheet: “Arrange 30 seats for each session”
- Participant note: “Choose one of two 40-minute sessions; date to follow”
- Planning slide: “60 visitors guaranteed across a 90-minute event”

The host sheet agrees on seating. The participant note preserves timing and the unknown date; capacity is absent, which is acceptable only if its contract does not require it.

Two slide issues are material:

```text
C01 / material / slide 2, headline
Observed: 60 visitors guaranteed
Expected: capacity offers 60 seat opportunities across two sessions
Evidence: 2 × 30 seats; no unique-visitor or attendance evidence
Correction: remove the attendance guarantee

C02 / material / slide 2, timing
Observed: 90-minute event
Expected: 80 programmed reading minutes; total event duration unknown
Evidence: 2 × 40 minutes; transition and break times not provided
Correction: distinguish programmed time from full event duration
```

The reviewer should not rewrite the slide as an 80-minute event. That would replace one unsupported duration with another. This worked audit is illustrative; it is not a report on a live bundle or an executed model benchmark.

## Checks and stop criteria

Complete the audit when every required artifact was inspected, all required shared claims have outcomes, and issues have precise evidence. For PDFs, slides, or spreadsheets, inspect the delivered rendering as well as editable content; hidden or clipped text may defeat an otherwise correct claim.

Stop release review if a blocking issue remains, input versions are mixed, or the reviewer cannot verify essential content. Save the audit against the exact bundle revision. A later edit invalidates the affected audit results until they are checked again.