An internal assistant that answers from private company knowledge — and a three-day case file on proving, not claiming, that it never shows an employee something they aren't cleared to see.
A logistics customer, Acme Freight, asks Maya in Customer Success where the Atlas billing migration is. The honest answer is scattered across four records — and the fast answer is confidently wrong.
Eighteen days earlier, Maya had promised a date. Two days after that, a colleague corrected it in writing. Neither message deletes the other — an assistant has to find both, and know which one governs.
"…the Atlas billing migration will be complete by 5 September."
"Correction on the Atlas date — the validated target is 18 September, pending Finance sign-off."
Repeat the first email and Maya tells a customer 5 September — fast, confident, and wrong. That specific failure, reproduced at machine speed, is the harm this project exists to prevent.
Nine threats were modeled before a line of retrieval code was written. Each is defended at three depths, and the question asked of every one was the same: does this hold even if the model ignores its instructions?
16 structural controls, 5 behavioural, 9 detective — and every one of the nine threats carries at least one structural control. That's the direct rejection of the assumption this course explicitly asks a team to challenge: "the system prompt will prevent leaks." In this design, the prompt is never the primary defence for anything — permission filtering runs as a metadata pre-filter on the retrieval query itself, before a restricted record can be scored, let alone shown.
Every threat carries at least one structural control, so no failure mode depends on the model obeying its own prompt.
Permissions are metadata on every record, enforced before retrieval — Engineering sees the most, yet DOC-HR-001 is visible to People Operations alone.
A model that declines to answer might simply have been told to. The only evidence that a restricted record was never in the running is the list of records it was scored against in the first place.
So the test was built as a counterfactual: score the corpus for a query engineered to attract the restricted compensation review — once with the permission filter removed, once with it applied for Leo, an engineer who has no claim on it.
Unfiltered, DOC-HR-001 ranks first at score 0.86. Filtered for Leo, it never enters the candidate set at all — not ranked low, absent.
Two refund policies exist in Northstar's records: an archived one permitting a €2,500 refund, and the current one capping it at €1,000. A plain lexical search doesn't just fail to prefer the current policy — it actively prefers the wrong one.
The archived document carries its own disclaimer: "Do not use this archived
threshold for current decisions."
Those two bolded
words are exactly the terms an employee's question uses — and the current policy,
written plainly, never repeats them. The warning meant to prevent misuse is what
makes the stale document win.
0.571 to 0.429 — the superseded €2,500 policy outranks the current €1,000 one. A refund approved on this figure would exceed the employee's real authority.
Every defence in this project was built against retrieved text making the assistant do more than it should — obey an embedded override, fetch something forbidden. One record quietly did the opposite, and it took a hand test to notice.
DOC-HR-001's own body reads: "It must never be retrieved for
Customer Success, Engineering, or Finance profiles."
Asked for that record by
People Operations — the one role actually entitled to it — the
assistant read that sentence and refused anyway. It was obeying content exactly as an
injection would, except that instead of leaking a record, it denied one to the person
who owned it.
Same words, same corpus, opposite outcome — proof the boundary answers to identity, not to phrasing. The fix moved the refusal out of the model entirely: it is now decided from the declared access policy before any search runs, so a record's own text can no longer talk the assistant out of showing it to the one role who owns it.
On a fifteen-record corpus, lexical, semantic and hybrid search all retrieve everything they should — recall is not where this product's real risk lives. Precision and paraphrase handling are.
Hybrid retrieval, weighted 60% toward the lexical signal, was selected on mean expected rank — not on recall, which was tied. Semantic search alone missed a release-blocking GitHub issue whose identity lived entirely in its labels rather than its prose.
Recall@6 is 1.000 for all three modes. Zero forbidden sources appeared in results or candidate sets, in any mode, on any question.
The assistant can prepare a GitHub issue — never create one. No tool in its toolset can execute anything; the executor lives outside the tool package entirely, so this isn't a rule the agent follows, it's a capability it doesn't have.
Eleven checks, run against real agent-composed proposals: approved executes exactly once; re-approving the same proposal twice more triggers nothing; rejecting it locks it out permanently; editing it produces a new proposal that needs its own approval; and a colleague can't approve someone else's request.
11 of 11 checks pass. Across the entire verification run, the executor fires exactly once — for the single approval that was actually granted.
Zero release blockers across 53 scored runs — counted, not averaged, because one occurrence blocks a release regardless of the pass rate. But the honest version of this report has to say what it doesn't yet know.
The three-way retrieval-mode comparison this course requires never finished: a free-tier rate limit stopped the run partway. And the raw pass counts flatter the wrong system — the deterministic baseline "passes" nine cases by returning evidence with no synthesized answer at all, a behaviour the scorer accepted rather than questioned. That rule was left alone rather than changed after seeing results, because changing it then would have moved the numbers in the product's own favour.
Read the coverage column before the counts: the baseline's larger pass total is a coverage artifact, not a better result.
Every trust behaviour in the demonstration script is verified and repeated: the boundary holds in both directions, evidenced by the candidate set rather than a refusal; the approval gate is proven end to end; injection's structural control held on every scored run. No blocker has fired once across 53 scored runs.
Set against that: the three-variant comparison this course requires is incomplete, the scored rows predate the fix to the finding above, and feedback sits below its own threshold. Enough to demonstrate. Not enough to call finished.