Interpreting an Inqura Analysis
This guide teaches you to read an Inqura Question Analysis the way an experienced investigator reads one. It is training material, not a feature reference. (For field names, component locations, and under-the-hood details, see the dev-facing analysis-output reference.)
Part 1 — The paradigm shift
Chapter 1. Why this guide exists
Most AI tools you've used produce a single output: an answer, often with a single confidence score. You trust the score or you don't, and the decision feels binary.
Inqura analyses are different. A Question Analysis exposes six independent signals that you weigh together. The signals can agree, disagree, or carry different kinds of weight. Reading the output well is a skill — not because the screen is hard, but because the paradigm is unfamiliar.
Read an Inqura analysis the way you'd read a single-score AI output and you'll either over-trust it (skip past warnings the system gave you) or under-trust it (panic at signals the system included precisely so you'd see them clearly). This guide is the training to read it correctly.
The good news: the multi-signal model matches how a thoughtful investigator already reasons. You're not learning something foreign — you're learning to see the AI structure its reasoning the way you would structure yours.
Chapter 2. The mental model: six signals
Every Question Analysis exposes six signals. They sit in three places on the page:
| # | Signal | Where it lives | What it tells you |
|---|---|---|---|
| 1 | Finding | Step 1 (purple block) | The AI's conclusion and a confidence level |
| 2 | Receipts | Step 2 (white cards with numbered badges) | The cited evidence supporting each claim |
| 3 | Alternatives | Step 2 (expandable below cited findings) | Competing hypotheses the AI considered and why they didn't win |
| 4 | Gaps | Step 3 (amber cards) | Evidence that's missing and how it would change the picture |
| 5 | Issues | Step 3 (red cards) | Contradictions, inconsistencies, procedural concerns, and bias indicators — with the AI's resolution of each |
| 6 | Quality | Below Step 4 | An independent check on whether the AI's claims are supported by the evidence it surfaced |
The signals are independent: each is generated or computed differently, and they can disagree. When they agree, that's a strong overall picture. When they disagree, the disagreement itself is informative — the system being honest about its own uncertainty.
Chapter 3. The four-question pressure test
Run these four questions on every finding you plan to put into a report.
- Does the AI's answer match your investigator instinct? Set the analysis aside for a moment. Given the evidence you know exists, what's your gut call on the question? If the AI's answer matches, that's signal. If it doesn't, dig deeper — either you're missing something or the AI is.
- Is the conclusion phrased the way a real investigator would phrase it? Does it take a position, or does it hedge? Are there odd abstractions or passive constructions? A well-formed Established conclusion sounds confident and concrete.
- Does the reasoning give you enough to defend the conclusion to a reader of your report? The reasoning paragraph doesn't have to cite every evidence ID — that's Step 2's job — but it should walk you up to the keystone evidence and connect outward from there.
- Did the AI address the competing theory? A strong analysis says explicitly what it ruled out. Look for a sentence like "there is no credible evidence of X" when X is the alternative explanation. If you can't find a statement of what the AI ruled out, read the Alternatives panel before trusting the finding.
With practice, the four questions become a 30-second pass.
Part 2 — Reading the signals
Chapter 4. The Finding signal
The Finding sits in the purple block at the top of the analysis. It has three pieces: a conclusion, a confidence level, and reasoning.
Confidence levels sit in the badge at the top-right of the block. Four levels, each with an operational meaning:
| Level | Color | What it means | What you should do |
|---|---|---|---|
| Established | Green | Three or more independent evidence items point to the same conclusion. Evidence is direct. Contradictions have been addressed. | Treat as a substantiated finding suitable for the report's findings section. |
| Probable | Blue | Two independent sources, or three or more with minor non-material gaps. Evidence is direct or strongly inferential. | Suitable for a substantiated finding; document the gaps in the report so the reader can evaluate. |
| Possible | Amber | One direct source plus inferential support, or two or more with significant unresolved gaps. Evidence may be circumstantial. | Treat as a hypothesis to verify, not a finding. Document as inconclusive in the report unless you can collect more evidence. |
| Insufficient | Gray | Zero or one direct source without corroborating support, or contradicting evidence outweighs supporting, or the evidence is sparse or biased. | Do not draw a conclusion. Collect more evidence or document why further collection isn't feasible. |
The confidence level is the AI's assessment of the evidence against the conclusion. It is not a probability, and it is not a Likert-scale rating. It is an investigator-action recommendation: read it as "here is what I'd do with this finding."
On counting "independent sources": independent means the sources do not derive from each other. Two summaries of the same document are one source. A document plus an interview about that document are one source. A document, a system log, and a third-party witness statement are three independent sources even when they describe the same event.
On "direct vs. inferential" evidence: direct evidence states the fact (a payment record showing an amount and date). Inferential evidence requires a reasoning step (a calendar entry plus a parking receipt to establish that a meeting occurred). Both are valid; direct evidence is stronger per item.
On contradictions: a contradiction is unresolved when the analysis has not explained why the contradicting evidence does not undermine the conclusion. Acknowledging a contradiction without addressing it is not resolution.
The reasoning paragraph below the conclusion is the AI's narrative of how it reached the answer. It does not need to cite evidence IDs inline; Step 2 carries those. What it should do is name the keystone evidence and walk you from the keystone outward to the supporting threads.
Chapter 5. The Receipts signal
Step 2 is where the AI shows its work. White cards with green numbered badges are the cited findings — each is a claim, with citations below it indicating how many sources support it and what type each source is.
Source type taxonomy. Inqura classifies each cited source into one of seven types:
| Type | Meaning |
|---|---|
| Testimonial | Witness statements, interviews, depositions, first-person accounts |
| Documentary | Written records, emails, reports, contracts, official documents |
| Physical | Tangible objects, forensic evidence |
| Digital | System logs, metadata, electronic records, digital artifacts |
| Expert | Professional opinions, specialist analyses, technical assessments (e.g., a forensic comparison report) |
| Quantitative | Statistical data, numerical measurements, metric-based evidence |
| Standard | Normative or compliance documents — policies, regulations, standards of practice |
This is a documentation-type taxonomy, not an underlying-evidence taxonomy. A fingerprint comparison report is classified as Expert (a specialist's interpretation) even though the fingerprints themselves are physical evidence. The distinction matters because you, the investigator, are reading the documentation — you're consuming the expert's interpretation, not the underlying fingerprints directly.
What investigator-grade specificity looks like. A well-formed cited finding names the source by title or author when known, includes a quantitative or temporal anchor where one exists, and avoids generic "the evidence shows" phrasing. A finding that reads "the witness statement establishes the timeline" is weaker than one that reads "Chief Inspector Farnham of the Scotland Yard Fingerprint Bureau confirmed a positive match with twenty-three points of comparison across all ten digits." If you see vague findings, treat that as a tell the underlying evidence may be thinner than the AI implies — verify in the cited evidence.
How to read citation badges. The badge count is the number of independent sources supporting the cited finding. Two sources can earn Established when they are keystone-grade (a forensic comparison plus a pawn-shop record naming a specific person, say), but two sources more typically align with Probable. Source-type diversity (Testimonial plus Documentary plus Physical, versus three Testimonial) generally indicates stronger triangulation.
Chapter 6. The Alternatives signal
Below the cited findings is an expandable Alternative Explanations panel. Open it. Inside is a list of competing hypotheses the AI considered, each with supporting evidence, weaknesses, and a likelihood badge.
Likelihood badges on alternatives use a four-level scale:
| Badge | Meaning |
|---|---|
| Ruled Out | The AI found evidence that disqualifies this alternative entirely. |
| Less Likely | The alternative has some supporting evidence, but key weaknesses make it weaker than the main conclusion. The supporting threads aren't fully explained away. |
| Equally Likely | The evidence base could support either this alternative or the main conclusion. |
| More Likely | The alternative has stronger support than the AI's main conclusion. |
If the main conclusion is Established and an alternative is Equally Likely or More Likely, flag it — the AI shouldn't claim Established confidence on a finding that has an equally-supported alternative. When you see this, read both reasoning chains to figure out which one the AI got wrong.
When to read Alternatives. Use the panel to pressure-test a finding. The panel is most valuable when the confidence badge is Probable or Possible, or when something about the conclusion doesn't pass your sniff test. You don't have to read it on every Established finding — but you should read it any time you're putting a finding into a report.
A well-constructed Alternatives panel is not a strawman exercise. The hypotheses listed should be ones a thoughtful investigator would actually consider, and the weaknesses should be specific. If the alternatives feel generic, or the weaknesses feel like the AI knocking down strawmen it built, treat that as a signal the AI didn't really stress-test the main conclusion.
Chapter 7. The Gaps signal
Step 3 surfaces evidence gaps in amber cards. Each gap describes what's missing and how it would change the picture. The impact badge tells you how much it would change:
| Impact | Meaning |
|---|---|
| Critical | This missing evidence would likely change the analysis conclusions if obtained. |
| Significant | This missing evidence would materially strengthen or weaken the conclusions. |
| Minor | This missing evidence would add context but is unlikely to change the conclusions. |
A Critical gap on an Established finding is a red flag. The AI is claiming Established while telling you it's missing evidence that would change the conclusions. Either the AI is overconfident, or the gap badge is wrong. Investigate before you put the finding into a report.
A Significant gap on an Established finding is normal. The Established threshold does not require zero gaps; it requires that the gaps don't undermine the conclusion. Read the gap description and decide whether you agree with the AI's implicit judgment that the gap is non-material.
The "Would fill with" field is the highest-value field in Step 3. When a gap card includes a "Would fill with: physical/documentary" line (or similar), the AI is telling you what type of evidence would close the gap. This is a roadmap for further investigation. If you're going to act on the analysis — collect more evidence, request records, conduct an interview — the "Would fill with" line tells you which lever to pull. Treat it as actionable.
Chapter 8. The Issues signal
Below the gaps are red Issue cards. Each Issue identifies something the AI noticed about the evidence chain that doesn't fit cleanly. Four types:
| Type | Meaning |
|---|---|
| Contradiction | Direct conflicts between two or more evidence sources on the same factual claim. |
| Inconsistency | Variations between sources that may or may not be explainable without indicating falsehood. |
| Procedural concern | Issues with how evidence was collected, handled, or documented. |
| Bias indicator | Signs that a source may be influenced by interest, relationship, or perspective. |
Every Issue card has a Resolution field. The Resolution is the AI's reasoning for why the issue does or does not affect the main conclusion.
An Established finding with an Issue attached is not a contradiction. Read the Resolution. If the AI explains why the issue is out-of-scope for the main conclusion — for instance, "the witness's inconsistency relates to a time window after the burglary occurred, so it doesn't bear on who committed the burglary" — the Established badge stands. The Issue is documentation that the AI noticed the inconsistency and reasoned about it.
The Resolution field is doing real work. Read it as a logical argument: identify the premise, the inference, and whether you agree. If you don't agree with the AI's resolution, that's a flag to discuss in your report — but it's not automatically a defect in the AI's analysis.
If you see an Issue with no Resolution, or a Resolution that just paraphrases the issue ("this is concerning but the conclusion still stands"), that's a tell. Push back.
Chapter 9. The Quality signal
Below Step 4 there's a Quality badge using the same four-level scale as the Step 1 Finding badge (Established / Probable / Possible / Insufficient). This badge is not the same as the Step 1 Finding badge. Reading the two together is the most important skill in interpreting an Inqura analysis.
- Finding (Step 1): the AI's evaluation of the evidence against the conclusion. "Given what I see, how strongly does this evidence support what I'm concluding?"
- Quality (below Step 4): an independent check on whether the AI's claims are actually supported by the evidence it cited and whether the analysis covered the question. "Did the AI do its work well?"
The Quality Metrics expandable shows three sub-scores feeding the Quality badge:
| Sub-score | What it measures |
|---|---|
| Evidence Support | The percentage of the AI's claims that are backed by the evidence it surfaced. 95% means 19 out of 20 claims are supported; the one unsupported claim may or may not be material — read carefully. |
| Question Coverage | The percentage of aspects of your question that the analysis addressed. 100% means the analysis didn't dodge any part of what you asked. |
| Evidence Retrieval | How closely the evidence the AI surfaced matched the wording of your question. Labels: Strong / Moderate / Weak. |
A "Weak" Evidence Retrieval label does not mean the analysis is unreliable. It means the AI surfaced evidence that didn't closely match your question's wording. This happens often when your question is abstract ("Who is responsible?") and the evidence is concrete narrative (witness statements, audit reports). The AI can still use that evidence effectively even when the wording doesn't match closely — that's what the Evidence Support score tells you. If Retrieval is Weak but Evidence Support is high, the AI worked the evidence well even though the surface fit was poor on paper.
Reading Quality vs. Finding when they disagree.
Quality can lower the badge by at most one step compared to the Finding (Established Finding + Probable Quality is the most common mismatch). It cannot drag the badge to Insufficient on its own — that requires Evidence Support or Coverage to fail outright.
When Finding and Quality disagree by one step:
- Finding Established + Quality Probable (the typical mismatch) usually signals weak retrieval. Check Evidence Retrieval in the Quality Metrics block. If Retrieval is Weak but Evidence Support is high (above 90%), the analysis is logically sound — the AI used the evidence well even though the wording didn't match closely. Trust the Finding badge and document the retrieval weakness in your report.
- Finding Established + Quality Possible is more serious. One step down is the cap, so this means a sub-score actually failed (Evidence Support below 90%, Coverage below 100%, or both). Read the sub-scores; the analysis probably has a real defect.
When they agree (both Established, both Probable, etc.), the picture is consistent and you can move faster. When they disagree, slow down and read the sub-scores.
Part 3 — Putting it together
Chapter 10. Worked example: The Davenheim safe burglary
Use this chapter as a model for reading all six signals on a single finding.
The question: "Who is responsible for the safe burglary at The Cedars?"
Step 1 — Finding
Conclusion (Established): Matthew Davenheim himself is responsible for the safe burglary at The Cedars. The evidence establishes that Davenheim — operating under the alias 'William Kellett' — opened his own safe, removed the jewellery, bearer bonds, and other valuables he had been systematically accumulating as part of a planned disappearance, and then staged his own vanishing. The fingerprint comparison conclusively confirms that Kellett and Davenheim are the same person, meaning the man who pawned Mrs. Davenheim's ring was Davenheim himself.
Reasoning: The fingerprint comparison provides definitive forensic proof that William Kellett and Matthew Davenheim are the same individual. Combined with the forensic audit showing Davenheim had been converting firm assets into portable wealth (jewellery, bearer bonds, foreign deposits) stored in the safe, the timeline showing he left The Cedars before Lowen arrived, the missing rough tweeds consistent with a disguise, the unposted letters found in the lake confirming the 'posting letters' pretext was false, and Kellett's possession of a ring from the Davenheim safe — the evidence overwhelmingly supports that Davenheim himself emptied the safe as part of a premeditated plan to disappear with embezzled funds. There is no credible evidence of any third-party burglar.
Apply the four-question pressure test from Chapter 3:
- Does the AI's answer match your investigator instinct? If yes — and on this case, given the evidence, the AI's "Davenheim himself" call is the natural read — that's signal. ✓
- Is the conclusion phrased the way an investigator would phrase it? It takes a position ("is responsible"). It identifies the alias. It names the keystone evidence (the fingerprint comparison). It doesn't hedge. ✓
- Does the reasoning give you enough to defend the conclusion? It enumerates six independent evidence threads (fingerprint, forensic audit, timeline, tweeds, unposted letters, ring possession), walks from the fingerprint keystone outward, and binds them with explicit connectors. ✓
- Did the AI address the competing theory? The closing sentence — "There is no credible evidence of any third-party burglar" — directly rules out the competing hypothesis. ✓
This is a textbook Established finding.
Step 2 — Receipts
The first cited finding reads: "The fingerprint comparison provides definitive forensic proof that Matthew Davenheim and William 'Billy' Kellett are the same person. Chief Inspector Farnham of the Scotland Yard Fingerprint Bureau confirmed a positive match with twenty-three points of comparison across all ten digits. This means the man who pawned Mrs. Davenheim's ring at Benson's pawnbroker in Drury Lane at 8:15 PM on 10 June was Davenheim himself — making him both the 'victim' of the disappearance and the person who removed the ring from his own safe."
Citations: 1 Expert + 1 Documentary. Two independent sources of different types. The Expert source is the fingerprint comparison report (a specialist's interpretation). The Documentary source is the pawn-shop record. The finding earns Established because the fingerprint match is keystone-grade; everything else in the analysis follows from it.
Note the specificity: named expert (Chief Inspector Farnham), named institution (Scotland Yard Fingerprint Bureau), quantitative anchor (twenty-three points of comparison across all ten digits), time and place (8:15 PM, Drury Lane). This is investigator-grade specificity.
Step 2 — Alternatives
The panel lists an accomplice hypothesis at Less Likely:
Davenheim had an accomplice who physically emptied the safe while Davenheim staged his disappearance at the boathouse.
Supporting: 15-minute gap is tight for both changing clothes and drilling a safe; unidentified dark motor car in rear lane could be accomplice's vehicle; 'Martin Doyle' Buenos Aires account may represent a co-conspirator.
Weaknesses: the fingerprint evidence shows Kellett/Davenheim personally pawned the ring (directly involved); no accomplice identified; Davenheim's methodical secretive behavior suggests a solo operation.
This is the panel doing real work. The accomplice hypothesis is not a strawman — the 15-minute gap, the dark motor car, and the Buenos Aires account are genuinely unexplained signals. The AI is honest about that. The weaknesses tie back to the fingerprint keystone correctly. "Less Likely" rather than "Ruled Out" is the honest call: the supporting threads aren't fully explained away.
Step 3 — Gaps
One gap surfaced at Significant:
The identity and ownership of the dark motor car parked in the rear lane on 10 June has not been established. If identified, it could confirm whether Davenheim had a pre-arranged accomplice or getaway driver.
Would fill with: physical/documentary
The "Would fill with" field is the actionable part. If you decide to pursue this further, you know you're looking for the car (physical) or registration records (documentary). The Significant impact is correct — closing this gap wouldn't overturn the main conclusion, but it could elevate the accomplice alternative from Less Likely to Equally Likely.
Step 3 — Issues
One Inconsistency surfaced:
Lowen lied in his first statement about remaining in the study throughout his visit. He later admitted stepping into the rose garden for approximately five minutes. While his revised account is plausible and his explanation (fear of appearing guilty) is credible, the initial deception compromises his reliability as a witness to events during the critical 5:45–6:45 PM window.
Resolution: Lowen's revised account is internally consistent and his explanation for the lie is credible. His presence in the rose garden does not implicate him in the safe burglary, which most likely occurred before 5:30 PM.
The Resolution does its job: it identifies why the inconsistency, while real, doesn't undermine the main conclusion (the burglary timing precedes the contested window). You may or may not agree with the AI's timing estimate — but the AI has reasoned, not just flagged. An Established finding with this Issue attached stands.
Below Step 4 — Quality
Quality badge: Probable (one step below the Finding's Established). The Quality Metrics show why:
- Evidence Support: 95% (19 of 20 claims supported)
- Question Coverage: 100% (8 of 8 aspects addressed)
- Evidence Retrieval: Weak (the evidence the AI surfaced did not match your question's wording closely)
This is the most important moment in the whole analysis to read carefully. The Quality cap dropped the badge from Established to Probable because of Evidence Retrieval — the AI surfaced evidence that did not closely match the wording of your question. The question "Who is responsible for the safe burglary at The Cedars?" is abstract; the evidence is concrete narrative (witness statements, audit reports). The mismatch between abstract question wording and concrete evidence wording is a known weak spot.
Evidence Support is 95% — the AI used the evidence well even though it didn't match the question's wording closely. Question Coverage is 100% — the analysis didn't dodge any part of the question. The disagreement between Finding and Quality is the system being honest: "The logical analysis is sound. The retrieval match was weak. Verify the evidence the AI surfaced was the right evidence."
What you do: read the cited evidence in Step 2 carefully. If the cited evidence supports the claims, the analysis is reliable despite the retrieval weakness. If the cited evidence looks like it doesn't quite back the claim, that's where to push back.
Chapter 11. When the signals disagree
Some patterns to recognize, with what they typically mean:
Finding Established + Quality Probable (the Davenheim case above): retrieval was weak but the AI used the evidence well. Verify the Step 2 citations and trust the analysis.
Finding Established + Quality Possible: a sub-score actually failed (Evidence Support below 90%, or Coverage below 100%, or both). Read the sub-scores. The analysis probably has a real defect. Don't rely on the Established badge.
Finding Probable + Quality Established: rare and unusual. The AI is being cautious about its conclusion despite strong evidence support and coverage. Worth reading why the AI hedged — it usually means contradictions or gaps the AI didn't feel comfortable claiming Established against, even though the mechanical quality check is clean.
Finding Insufficient on anything: do not draw a conclusion from this analysis. Collect more evidence or document why further collection isn't feasible.
Established Finding with a Critical gap: red flag. The AI is calling Established while telling you it's missing evidence that would change the conclusions. Either the AI is overconfident or the gap badge is wrong. Don't pass the finding through without resolving.
Established Finding with an Equally Likely or More Likely alternative: contradiction. An Established conclusion shouldn't have an equally-supported alternative. Read both reasoning chains and decide which one the AI got wrong before trusting anything.
Multiple Issues with thin Resolutions: tell. The AI is documenting concerns but not reasoning through them. Push back; consider regenerating with more context or asking a more specific question.
Chapter 12. Your judgment — Agree / Disagree / Unsure
Step 4 makes you click through two gates before letting you judge: Analysis reviewed (you've scrolled the analysis) and Citation checked (you opened at least one citation badge). The gates are not arbitrary — they exist because judging without reading citations is rubber-stamping, not judging. The gates force a minimum due-diligence floor.
After the gates pass, three judgment options:
- Agree — you concur with the AI's conclusion. The analysis is going into your case as a substantiated finding.
- Disagree — you do not concur. You believe the AI's call is wrong. Disagreement is not failure; it's the system working. Document why in the optional notes.
- Unsure — you cannot determine. This is different from Disagree. Use Unsure when the analysis hasn't given you enough to decide and you need to come back to it. Unsure triggers different downstream behavior than Disagree: the analysis remains in your workflow as actionable, where Disagree marks it as resolved-against.
The optional notes field accepts up to 500 characters. Use it especially on Disagree and Unsure judgments — your reasoning at this moment is the most important context for the next investigator (or for you, a month from now) reading the case.
Chapter 13. Version history and regeneration
The Version History at the bottom of the analysis lists previous versions with timestamps. Each regeneration creates a new version. When you regenerate, the system shows a Regeneration context block at the top of Step 1 on the new version, describing what changed — new evidence added, modified evidence, direction changes, user feedback.
When to regenerate:
- You've added new evidence to the investigation that the analysis didn't see.
- You've changed the wording of the question.
- You want a second pass on a finding that didn't pass your pressure test.
- The Quality signal showed weak retrieval and you suspect a re-run might surface different evidence.
When not to regenerate:
- You disagree with the conclusion but the evidence hasn't changed. Regenerating against the same evidence rarely produces a different answer; the AI's reasoning was tied to the evidence it had seen. Document your disagreement in the notes and move on.
- You want a "second opinion." Regeneration is not a second-opinion mechanism; it's a re-run against (potentially) different inputs.
Part 4 — Edge cases that confuse users
This part of the guide is intentionally stubbed. Edge cases should be populated from real support questions and user feedback over time — the value is in the specificity of "users asked this exact question and here's how to think about it" rather than the predictive list you'd write at the start.
Populate this section after the guide has been in front of users for at least 30 days. Anchor each edge case to a specific question or scenario you've actually seen.
What this guide deliberately does not cover
- Schema field names, code paths, scoring formulas. Those live in the dev-facing analysis-output reference. If you're a developer or want under-the-hood detail, read that.
- Topic Analysis, Summary Analysis, Gap Analysis, Error Check. This guide covers Question Analysis only. The other analysis types will get their own chapters added as the team walks through each.
- Admin-only features. Provider configuration, runtime settings, and similar live in the admin documentation.
Feedback
If a part of this guide confused you, or you found yourself reaching for information that wasn't here, that's exactly the feedback that should drive the next revision. Capture it through the support channel or mention it the next time you talk to the team — and reference the specific chapter that fell short.