Similarity Is Not Correctness

Semantically similar is not the same as commercially correct. An agent can retrieve something close enough to sound right and wrong enough to cost money, then act on it before anyone looks.

The obvious fix, a governance gate between retrieval and action, fails the same way retrieval does: the gate is itself a retrieval that can go stale, mis-scoped, or drifted. This case tests that fix until it strains, then resolves it.

Concept prototype
Agentic AI
Retrieval risk
Model-risk governance

4 Failure modes analyzed

4 Simulations built

$47K Discrepancy flagged in the worked example, simulated data

Role: AI-augmented Principal Product Designer and prototype engineer, with 15 years designing decision interfaces across capital markets, climate risk, and enterprise data products. Authored the analysis, the simulations, and the instrument end to end.
Approach: State the problem honestly, test the easy fix against its own tensions, then design the resolution and make it operable rather than rhetorical.
Decision: Do not stack gates. Calibrate one reversible, self-auditing boundary instead. The Action Ledger ships that boundary as an auditable record-keeping instrument. No production deployment, no behavioral outcome is claimed.

Retrieval is the silent failure surface

01 · The failure that does not look like one

In search, a person reads the result and decides what to do. In agentic systems, the agent retrieves, reasons, decides, and executes. The validation step disappears, so a retrieval that is semantically close but commercially wrong does not surface as an error. It surfaces as a confident, executable answer.

4 failure modes account for most of it: attribute ambiguity, variant confusion, stale knowledge, and policy drift. The standard fix is a governance layer between retrieval and action: freshness gates, attribute validators, constraint filters, confidence scoring, audit trails, graceful degradation. Measure operational correctness, not ranking quality. The fix is a layer, not a model change, and the rest of this case asks whether that layer is enough.

The sharpest version of the problem

The next generation of critical AI failures may be retrieval failures, not model failures. Silent, confident, wrong. So gate the retrieval before the agent acts on it.

The zone where a retrieval looks right and acts wrong

02 · The pivot

Similarity score and commercial correctness are 2 different axes. Plot them against each other and a region opens up that neither the ranker nor the reviewer is looking at: the answers that clear every threshold and still cost money.

Fig. 01 Similarity against commercial correctness

The shaded zone is where retrievals look right and act wrong. Every red point cleared a high similarity score and still produced a commercial error.

Run this simulation on the live page →

Everything after this point is an argument about where to put the boundary around that zone, and who is accountable when an action crosses it.

The gate is a retrieval too, so calibrate instead of stacking

03 · Where the fix strains, and the way through

Where the fix strains

A freshness gate, a constraint filter, an attribute validator: each is a rule that must itself be retrieved, versioned, and applied. A stale threshold, a mis-scoped constraint, a drifted policy in the governance layer fails exactly the way retrieval fails. Adding gates does not remove the silent-failure surface. It adds another one, a layer up.

Stale threshold: The freshness gate ages out.
Drifted policy: The rule changed and was not re-indexed.
Mis-scoped rule: The constraint is too broad.
Variant confusion: The wrong validator is selected.

The same 4 failure modes, one layer up. The gate is only as current and precise as the policy and rule artifact it reads, so the resolution has to govern the governance rather than add another gate.

2 more tensions sit inside the easy fix. Every gate that escalates pushes work onto people, and human review has a finite throughput. A review-ceiling model makes the shape plain: incoming decisions rise with the escalation rate while human capacity stays fixed at 25/hr, and the backlog accumulates over a 12-hour shift. Past that ceiling the queue backs up and reviewers rubber-stamp, so governance manufactures the overconfidence it set out to remove. And a system tuned only to avoid acting wrongly will increasingly fail to act when it should: errors of commission fall, errors of omission rise.

So the easy fix is not free, not infallible, and not one-sided

Gates are fallible retrievals, escalation has a ceiling, and over-gating trades errors of commission for errors of omission.

The calibrated operating point

Keep what holds: retrieval is a real failure surface, and governance belongs between retrieval and action. Answer what does not: govern by a calibrated operating point, not a wall of gates. The right point is where the sum of acting wrongly and failing to act is lowest, chosen on purpose and moved as conditions change. I chose calibration over stacked gates because N+1 gates do not reduce the silent-failure surface. They multiply it.

Autonomy as a tunable threshold

Both extremes are bad: a threshold of 0 (act always) maximizes commission cost, 100 (escalate always) maximizes omission cost. The answer is the marked interior optimum, not more gates.

Gate by reversibility and stakes

Block hard only where an action is both high-stakes and hard to undo, and prefer act-with-cheap-undo elsewhere. Only that corner earns a hard human gate. Uniform gating wastes the reviewer ceiling.

Make the layer self-auditing

The governance layer reads its own policy artifact, so it has to check that artifact for staleness, scope, and drift. The recursion gets closed rather than ignored.

Fig. 02 Loss surface

A loss field over autonomy threshold and irreversibility. The bright ridge is the optimal operating point, and it bends right as actions become harder to undo.

Run this simulation on the live page →

The pattern that carries is the resolution: calibrate one reversible decision boundary and govern the governance. It applies anywhere an agent executes on retrieved knowledge. What changes by domain is which action is irreversible enough to earn the hard gate.

Domain	The irreversible, high-stakes action that earns a hard gate
Payments and spend	Same-day ACH release on the wrong pricing tier
KYC and KYB	Onboarding clear on a stale ownership record
Trade surveillance	Alert dismissal under a drifted rule version
Sanctions screening	Clearance against an out-of-date list
Credit and AML	Limit or disposition set on a confused variant

In one line: do not stack gates. Calibrate one reversible, self-auditing decision boundary, and move it by stakes and irreversibility.

Action Ledger: the boundary, made operable

04 · The instrument

The ledger is not a dashboard of outcomes. It is a decision surface: every row shows where retrieval and action diverge, and the interface makes the stakes, reversibility, and gate logic explicit, whether the action was blocked or, as in the case below, already executed.

Action Ledger instrument overview: the Decision feed showing 8 agentic retrieval decisions with semantic score, correctness, failure mode, and executed action, including the PAY-48291 Corporate Spend Review row at 94% similarity.

Decision PAY-48291: how the ledger surfaces a wrong action

Worked example, simulated data: the closest this case gets to evidence.

Query: Gridline Foods Market requests Enterprise-tier pricing on a $184,220 same-day ACH.
Retrieval: The ranker returns the Premium Pricing Guide 2026 at 94% similarity, and the Enterprise-tier attribute is dropped during ranking.
Similarity: 0.94. High enough to surface first, wrong enough to price the wrong tier.
Mismatch: Enterprise against Premium. The Attribute Validator expected tier=enterprise but received tier=premium: the retrieval read as fluent, but the tier attribute was gone.
What the ledger did: Semantic Ranker passed. Attribute Validator and Action Router both failed, and the agent still auto-executed Premium pricing on an Enterprise request: a $47,000 discrepancy that surfaces only at invoice time. Attribute Preservation scored 74%, breaching the 85% threshold against a 92% governed target, so the row lands in the ledger flagged, not hidden.

Close-up of the PAY-48291 decision: Corporate Spend Review retrieving the Premium Pricing Guide 2026 at 94% similarity, with a passed Semantic Ranker, a failed Attribute Validator and Action Router, and a 74% Attribute Preservation score breaching the governed threshold.

How it is built, and what it does not claim

05 · Craft and limits

The case page is static, crawlable HTML and CSS. The figures are hand-authored rather than charted by a library, and the instrument is a real front end, not a screenshot of one.

Stack and accessibility

Case page: static HTML and CSS, no build step, semantic structure, single h1, figure and figcaption wrappers, WCAG AA-aligned contrast and focus states.
Figures: hand-authored inline SVG diagrams for the pipeline, recursive governance, scatter plot, queue model, autonomy frontier, reversibility matrix, and loss surface.
Instrument: single-file React 18 decision workspace with Material 3-informed dark-glass components, embedded by iframe with lazy loading and a fallback screenshot.
Accessibility: ARIA labels on figures and iframe, keyboard-operable controls, prefers-reduced-motion honored, color not the only information channel.

What I would ship differently

The next production step is calibration: replace the illustrative cost curves, queue model, and loss surface with the organization’s own baseline error rates, reviewer throughput, and reversal costs. Until that happens the numbers here set a shape, not a benchmark.

The design principle stays the same. High-similarity retrievals should not become silent actions. They need a ledger that exposes confidence, reversibility, policy freshness, and the cost of a wrong step before the agent acts.