Blog · ← All posts
Ask the Tape: Interrogating an Agent Ledger in Plain Language Without an LLM Guessing
By the Ingenious AI team · Published: November 14, 2026 · Reviewed: November 14, 2026 · 6 min read
The short answer: Ask the Tape is Agent Flight Recorder's ledger interrogation feature: you ask a question about an agent's recorded history in plain language — every refund over a threshold, the first time a tool was used, what happened after a given event — and it answers from the ledger with citations to the exact records (sequence numbers and hashes) behind the answer. It is deliberately deterministic, with no language model generating the answer, so a cited fact is a found fact: the system cannot hallucinate a record that is not in the chain.
Replay a recorded agent incident free → Step through a real recorded run in the replay view, or paste ledger records into the public verifier and get a VERIFIED/BROKEN verdict in seconds — no signup.
The value of an audit ledger is capped by how fast a human can get answers out of it. SQL helps engineers; it does not help the support lead at midnight, the auditor in a sampling session, or the founder answering a customer's pointed question on a call. The natural interface is a question in plain language — and the dangerous implementation is the fashionable one, where a language model reads some records and composes a plausible answer.
Plausible is the problem. An answer about evidence must be exact, and it must show its work. Ask the Tape takes the other path: a deterministic interrogation layer over the ledger that parses the question's intent, runs it against the records, and answers with citations. This guide explains what it answers, how citation discipline works, and where its limits honestly lie.
What is ledger interrogation? question-answering over an audit ledger in which every factual claim in the answer is produced by querying the records and is accompanied by a citation — the sequence number and hash of the record it came from. In Ask the Tape the process is deterministic (pattern-matched intents over structured record fields), so the same question over the same ledger always yields the same answer, and any claim can be re-checked against the chain by the reader.
The question shapes it answers well
Interrogation shines on the questions incident work actually asks. Enumerations: 'show every refund the support agent issued over $200 in March.' Firsts and lasts: 'when did this agent first call the export tool?' Sequences: 'what did the agent do in the ten steps after record 4,102?' Counts and sums: 'how many actions touched customer data last week, and what did they total?' Provenance: 'which records cite this approval as their cause?' Each maps to structured fields the ledger already holds — tool, action type, amounts in summaries, timestamps, causal links — so each is answerable by query, not by interpretation.
The phrasing can be natural without being magical: the system recognizes the intent patterns it supports and says so when a question falls outside them, rather than improvising. A tool that declines a question it cannot answer exactly is more useful than one that answers everything approximately, especially when the answers go to auditors.
Citations are the product
Every answer arrives wearing its evidence: matching records listed with sequence numbers, timestamps, and hashes, so the reader can jump to any cited record, see it in the replay, and confirm it sits in a verified chain. The citation converts the answer from a claim into a shortcut — the system's job becomes finding, and the reader's job stays verifying, which is the correct division of labor for evidence.
Citations also bound errors honestly. If a question's intent was misread, the cited records will visibly not match the question asked, and the mistake is caught by the reader in seconds. An uncited prose answer hides the same mistake inside fluent text, which is how interrogation tools manufacture false confidence.
Why no LLM — a deliberate architecture
It would be easy to wire a model to summarize the ledger, and AFR refuses to, for a reason worth stating as a principle: the component that reports evidence must not be a component that can confabulate. A language model answering 'did the agent email the customer?' can produce a confident yes assembled from vibes and a nearby record; the deterministic layer produces yes-with-records or an explicit no-match, and the difference is the difference between an exhibit and a rumor.
The model-free design also makes answers reproducible — a property auditors quietly require. Re-ask the same question over the same ledger next quarter and the answer is byte-identical, because nothing in the path samples, imagines, or forgets. Where genuine summarization is wanted, it belongs downstream of the citations: a human (or a model, clearly labeled) may narrate over cited records, but the evidentiary layer underneath stays arithmetic.
Limits, stated plainly
Deterministic interrogation answers questions about what is in the records, in the shapes it knows. It will not judge intent ('was the agent trying to hide this?'), will not assess quality ('were these refunds justified?'), and will not reach outside the ledger (the customer's feelings are not in the chain). Those are human questions — the tool's role is to hand the human the exact records from which judgment proceeds, fast enough that judgment happens while it still matters.
Field quality bounds it too: an interrogation over argument summaries can only find amounts that were summarized. This is one more reason the recording discipline matters — masked but meaningful summaries are what make the ledger answer questions instead of merely storing them.
Five actions you must take today
Interrogation value is created at recording time and harvested at question time. Set up both ends:
- Write summaries a query can use. Amounts, counterparties (masked), and object IDs belong in argument summaries in consistent formats — interrogation finds only what records state clearly.
- Teach the team the question shapes. Enumerations, firsts, sequences, totals, provenance — a one-page cheat sheet turns midnight fumbling into midnight answers.
- Require citations in every answer you forward. If a number leaves the building, its records travel with it; uncited figures get sent back, kindly and always.
- Use interrogation in audit sessions live. Answering a sampler's question in the room, from the tape, with citations, ends more audit threads than any architecture slide.
- Log the questions you could not answer. Each miss is a recording improvement ticket: a field, a summary habit, or an intent worth adding.
Worked example: an auditor's question, answered in the room
During a fictional enterprise security review, Northbeam Outfitters' assessor asked the question designed to end meetings: 'Show me every time your support agent exported customer data in the last 90 days, and who approved it.' The team asked the tape. The answer came back with 14 matching records across 61,204 in the window — each an export action, each cited by sequence and hash, 13 carrying approval references and one, from the agent's first week, without.
The exception was real and already known: it sat in the incident file with its remediation, because the team's own monthly review had run the same query. The assessor's follow-up — verifying three cited records against the chain — took minutes on the public verifier.
The review's written comment used the phrase teams should frame: 'Evidence requests were answered from the system of record, during the session, with independently verifiable citations.' That sentence is what interrogation is for.
FAQ
What kinds of questions can Ask the Tape not answer?
Questions requiring judgment, external facts, or unstated fields: whether an action was wise, what a customer intended, or totals over values that were never put in the records' summaries. It also declines intent patterns it does not support rather than guessing. The honest-decline behavior is a feature — in evidence work, 'I can't answer that from the records' is a valid and valuable output.
Can answers be exported for an audit file?
Yes — the answer with its cited records is exactly the unit an audit file wants: the question as asked, the matching records with sequence numbers and hashes, and the chain-verification result for the period. File them together and the auditor inherits a checkable artifact instead of a screenshot of a dashboard.
Does interrogation work across multiple agents?
Questions can span the fleet where identity and causal links connect the records — provenance questions ('what did this output cause?') follow caused_by links across agents. Per-agent scoping remains the default for a simple reason: attribution stays crisp when the answer names the agent for every record it cites.
Why not add an LLM summary on top for executives?
You can — as a labeled narration layer, downstream of the citations, never as the source of a number. The rule is mechanical: every figure in the narrative must trace to a cited record the reader can open. If a summary sentence cannot carry its citation, it does not ship. That single rule lets you have readable reports and untainted evidence in the same document.
How does interrogation handle masked data?
It queries the summaries as recorded — masked identifiers remain searchable by their masked form and by the object IDs and digests around them, while the sensitive raw values never enter the ledger to be queried in the first place. An auditor can establish that the same masked account appears in 14 records without the ledger ever exposing the underlying number.
Start logging your agents free → The free tier gives you the full hash-chained ledger for 1 agent and 1,000 actions a month, with chain verification and replay included.