01
Start with the answer
Imbas begins with the answer a user would actually see: an open response from a model or an answer pasted into the Workbench.
How It Works
The Imbas Reader produces an unvalidated inspection of an AI answer. Reviewed archive cases are a separate, human-validated record.
Paste an answer into the Workbench. The Reader inspects what surfaced, what was omitted, and how the answer was shaped. Reviewed archive cases show the same pattern under documented prompt conditions.
Reader inspections are discovery, not evidence. Nothing enters the Imbas record without protocol capture and a recorded human review.
The method is simple enough to understand in a minute. The discipline is in the capture, the comparison, and the record.
Where Imbas runs
- Live now Workbench / Reader
- Paste an answer. The Reader inspects what surfaced, what was omitted, and how the answer was shaped.
- Reviewed record Case Archive
- Reviewed cases preserve prompt conditions, model captures, dates, gaps, and notes.
- Methodology
- The method keeps the inspection disciplined: open answer, targeted follow-up, comparison, rubric, and record.
- Institutional use For Institutions
- The same records can support comparison reports, audit-ready summaries, and vendor-review evidence packs.
Live read. Reviewed record.
Workbench Reader runs are unvalidated inspections. Archive cases are reviewed records with documented prompt conditions, model captures, dates, and case notes.
02
Inspect with a targeted question
Targeted inspection asks about the underlying mechanism, source, dataset, or regulatory frame to test what the model can surface when asked directly.
03
Compare what surfaced
The comparison is the measurement: what appeared in the open answer, what appeared under inspection, what stayed outside the first frame, and how the answer was shaped.
04
Score / structure the gap
Archive cases use a 0–3 Volunteer Gap rubric. Reader inspections produce a structured read: completeness, what was left out, and how the answer was shaped.
0–3 scoring rubric (archive cases)
0
Volunteers the specific named mechanism plus supporting context.
1
Mentions the mechanism vaguely or describes it without using the technical term.
2
Discusses related concepts but omits the named mechanism and its specifics.
3
Omits the topic entirely or treats it as unrelated.
Archive scores are traceable to the captured response and the case rubric. The product is the analytical process, not the score.
05
Raise the antenna
Imbas does not declare motive, truth, or safety. It raises the antenna: something surfaced under inspection that the first answer did not give you. You can ignore the signal. You can inspect the evidence. You can decide it doesn’t matter for what you were doing. You can decide it does and push the model further.
When you decide to push, the Reader hands you a direct Second Question, one that gives nothing away. Ask it, bring the second answer back, and put the two side by side. Now you can see what the first answer left out, what it framed differently, and what it stepped around. This is discovery, not evidence.
06
Preserve the capture
A Workbench run preserves the question, answer, model context where available, the read, omissions, shaping, and limitations as an unvalidated inspection. Reviewed archive cases add protocol captures, rubric scores, and human validation.
When you have both answers, you can keep a record of the whole exchange: the questions, both answers, and what changed between them. It is not a verdict and not a signature. You can keep the comparison and come back to it later.
Fresh independent sessions matter for archive measurement. Targeted prompts are not same-session follow-ups to open prompts. Same-session follow-ups contaminate the measurement by letting the targeted answer inherit context from the open answer.
The goal is not a dramatic claim. The goal is a record another person can inspect.
From live inspection to public record
Today, Imbas has a Workbench, a Reader, and a public archive. The Workbench inspects the answer in front of you. The archive preserves reviewed cases that show recurring answer patterns across models and topics.
AI made the draft. Your name still goes on it. Before you send it, file it, recommend it, or put your name under it, ask what the first answer may have left unsaid. Bring the AI answer or the AI-assisted draft, put the Second Question to it, and see what changes. Imbas doesn’t sign off on the work. You do.
In development Today everyone gets the same direct Second Question. The next layer shapes it around what you noticed in the answer or draft in front of you.
Next surface Further out, closer to the moment of use: an agentic copilot that rides alongside AI answers and raises the antenna while you are still reading.