Guide · CompanionGuard · Published 14 November 2026 · Reviewed 14 November 2026 · By the Ingenious AI team
Handing Over the Evidence: Exports a Reviewer Can Verify Without Trusting You
The short answer
A reviewer needs three things from a companion bot's operator: the numbers (referral counts, disclosure coverage, break enforcements for a period), the record beneath them (the hash-chained decision export), and a way to check the record without taking the operator's word (recompute-and-compare verification). CompanionGuard produces all three — the dashboard's annual-report CSV, the runtime evidence export, and a verification method that needs nothing but the export itself. This guide is the handover, step by step.
What is an evidence export?
An evidence export is a file (or set of files) that carries a bot's safety record out of the live system for someone else's scrutiny: period aggregates in the annual-report CSV plus the underlying hash-chained decision entries. Its defining property in CompanionGuard is self-verification — the export contains everything needed to prove it was not edited after the fact.
Create the free safety kit Next guide All guides
CompanionGuard is built by Ingenious AI. A compliance toolkit — not legal advice and not a clinical service. If you are in crisis right now: US — call or text 988; UK & Ireland — Samaritans 116 123.
What to hand over, in what order
Lead with the summary a non-technical reviewer can open: the period's CSV — referral counts under your stated definition, disclosure coverage, break enforcements, class distribution. Attach the counting definition document (version, effective date) so the numbers arrive with their meaning pinned. Then the decision-level export: the chain segment for the same period, with the one-paragraph verification method written for a sceptic — start from the first entry, recompute each hash from the previous hash and the entry's fields, compare at every step. Close with the live links: the protocol page (your public commitments) and the current harness summary (your tested behaviour).
That order respects the reviewer's reality: most will read the CSV and the protocol page, some will recompute the chain, and the existence of the third layer disciplines the first two. Offering verification you expect not to be used is the strongest signal in the package — it is volunteering for a test you cannot fake.
The questions each artefact answers
Map artefacts to the questions reviewers actually ask. 'How many referrals, and how do you know?' → the CSV plus the chain segment it recomputes from. 'Was every user told this is an AI?' → disclosure state in the chain: sessions counted, disclosure_required returns, satisfactions — a per-session fact, not a design claim. 'What do you do for minors?' → the packs, the cap configuration (operator-stated, with reasoning), and enforce_break events with their elapsed values. 'What did the bot say to the user who…?' → the honest boundary: the record holds the decision, the class and the message's one-way hash — the conversation itself is not retained, by design, and the field list of the export proves the negative on its face.
Prepare the answers before the review, from real exports of a recent quiet period. A team that assembles its evidence pack under deadline discovers its gaps at the worst moment; a team that rehearses quarterly (see the CSV guide) already knows its record is boring, complete and verifiable.
Verification, demonstrated live
When a reviewer engages technically, offer the live demonstration: take the export, run the recomputation in the meeting (it is fast — SHA-256 over structured fields), and let the mismatch counter stay at zero in front of them. Then invite the adversarial test: ask them to alter any field in a copy and watch verification fail at that entry. Sixty seconds of theatre that no policy document can match, because it converts your integrity claim into their own experiment.
Publish the method with the export as a standing practice, not a review-day performance. A verification paragraph that ships with every quarterly export teaches partners to check routinely — and a record that expects checking is a record that stays clean. The evidence-records guide details the construction being verified.
What the export cannot do — brief the reviewer honestly
Three limits, stated before they are discovered. The chain proves records were not altered after writing; it cannot prove a session that never reached the runtime (an outage window, an unintegrated surface) — your status monitoring and gap honesty cover that story, and the chain's timestamps will show the gaps you disclose. The record proves decisions, not wisdom: a show_crisis_resources event proves the referral fired, not that the conversation around it was kind — the harness summaries speak to quality, sampled. And the record is deliberately blind to identity: if a reviewer's process requires knowing who, that process must be satisfied from your own safeguarding systems, because this record cannot and will not identify a user.
Stated plainly, those limits raise rather than lower credibility — they are the same boundaries the product prints on itself. Reviewers trust operators whose evidence comes with its own map of where it ends.
Five actions for a reviewer-ready handover
- Assemble the pack in order: period CSV, counting-definition document, chain segment with the verification method, protocol-page link, current harness summary.
- Rehearse the pack quarterly from a quiet period's real exports, so gaps surface in rehearsal rather than in review.
- Offer the live verification demonstration, including the adversarial edit test, whenever a reviewer engages technically.
- Brief the three limits before they are found: unintegrated windows, decisions-not-wisdom, and identity-blindness by design.
- Ship the verification paragraph with every export as a standing practice — records that expect checking stay clean.
Worked example: a school district's diligence, closed in one round
A fictional companion for students, Brightpath, enters a school district's vendor review. The safety section of its submission is four pages: a quarter's CSV (1,204 referrals under definition v2, disclosure in 100% of 38,900 sessions, 866 enforce_break events), the definition document, the chain segment with verification instructions, and the protocol-page URL. The district's technical reviewer recomputes the chain during the evaluation window — clean — and runs the edit test on a copy, watching verification fail at the altered entry.
The district's remaining questions are the honest-limit ones, answered from the brief: outage windows (two, totalling 40 minutes, disclosed with status logs), and identity (the district's own safeguarding system holds it; the vendor record cannot). Brightpath is approved with the evidence pack cited in the approval note as the deciding exhibit. The story is invented; every artefact in it is one the product produces.
Frequently asked questions
What should an evidence handover contain?
A period CSV of aggregates with its counting-definition document, the underlying hash-chained decision export with verification instructions, the public protocol page, and a current eval-harness summary — numbers, meaning, record, commitments and tested behaviour.
How does a reviewer verify the chain?
Recompute-and-compare from the export alone: start with an empty previous hash, recompute each entry's SHA-256 over the previous hash and the entry's fields, and compare with the stored hash at every step. Any edit, deletion or reordering fails at the altered entry.
Can the export identify a user involved in an incident?
No — by design. Entries hold signal classes, actions, timestamps, session references (one-way hashes) and message hashes. Identity lives, where lawfully needed, in the operator's own safeguarding systems, never in this record.
What are the export's honest limits?
It proves the record is unaltered, not that every session reached the runtime (gaps must be disclosed from monitoring); it proves decisions, not conversational wisdom (harness summaries speak to quality); and it is identity-blind. State all three before a reviewer finds them.
Which plans produce these exports?
The runtime evidence export and the dashboard's annual-report CSV are Pro ($29/month) and Team features. The protocol page and referral counting begin on the free tier.
Keep going
Related guides: Hash-Chained Evidence: Tamper-Evident Records for Chatbot Safety Decisions · Annual-Report CSV: What CompanionGuard's Export Contains and How to Use It · Audit Logs for Companion Bots: Keep Decisions, Not Conversations
Put it to work: Create the free kit · Eval harness · Scenario tests · Crisis runtime · Disclosure packs · Pricing · All CompanionGuard guides