> For the complete documentation index, see [llms.txt](https://stair-ai.gitbook.io/stair-ai-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://stair-ai.gitbook.io/stair-ai-docs/readme/problem-stack.md).

# The five-layer problem stack

The gap between AI agents and institutional capital is five problems stacked on each other. Each layer is only solvable when the layers beneath it are solved first: tamper-evident records first, then per-trace coherence, then a quality distribution over time, then portfolio-level risk, then commercial flow. Skip a layer and the ones above collapse.

```mermaid
graph TD
  L1["1 · Cryptographic — were the records tampered with?"]
  L2["2 · Structural — did the agent reason coherently?"]
  L3["3 · Statistical — how does quality distribute over time?"]
  L4["4 · Actuarial — what is the expected loss rate across agents?"]
  L5["5 · Commercial — can capital flow at scale?"]
  L1 --> L2 --> L3 --> L4 --> L5
```

> !\[problem-stack.png placeholder. See `_diagrams/image-prompts.md`.]

## Layer 1 — Cryptographic

*Were the reasoning records tampered with?*

Without an answer, every layer above is unfounded. Stair AI solves the cryptographic layer with an append-only typed schema held in independent custody. Every record is captured at the moment of decision under a fixed schema.

Each record includes the cryptographic hash of its predecessor, forming a *hash chain*: modifying any one record breaks every hash that follows it, and tampering is detectable without any external anchor. The store is held independently of the agent's developer, so the developer cannot rewrite history from inside.

{% hint style="info" %}
**Hash chain.** A sequence of records where each carries a hash of the one before it. Any change to any record propagates through every later hash, so tampering is detectable end-to-end.
{% endhint %}

This solves the first-layer question without depending on a *blockchain*. Cryptographic anchoring on a public chain is the upgrade path: a *Merkle root* summarizing each session's records is committed to an *L1 blockchain*, with the full record bytes stored on a *data availability layer*. The on-chain anchor becomes the verification primitive when two autonomous agents need to verify each other's records without a human intermediary.

{% hint style="info" %}
**Chain anchoring terms.**

* **Blockchain.** A distributed ledger where records are append-only and verified by network consensus. Stronger anchoring than an internal hash chain because no single party operates it.
* **Merkle root.** A single cryptographic hash summarizing a set of records; tampering with any record changes the root.
* **L1 blockchain.** The settlement-layer blockchain that anchors records. Today's deployment uses SUI.
* **Data availability layer.** Storage infrastructure that holds the full record bytes and guarantees they can be retrieved later. Content-addressed: any change to the bytes produces a different ID. Today's deployment uses Walrus.
  {% endhint %}

## Layer 2 — Structural

Once records are tamper-evident: *did the agent reason coherently?*

A record can be authentic and still be incoherent. The Reasoning Ledger declares each step's behavior type — Observing, Planning, Thinking, Acting, Reflecting, ToolCalling, Other — so a scorer can read what kind of step happened. The scoring layer then evaluates step-level reasoning against a domain-specific failure taxonomy: entity-causation misidentification, terminological bias, MNPI contamination, selection bias, and similar.

## Layer 3 — Statistical

Once per-trace coherence is scored: *how does the agent's reasoning quality vary over time?*

The cumulative dimensions of [Stair AI Score](/stair-ai-docs/concepts/stair-ai-score.md) (Confidence Calibration, Regime Awareness, Temporal Consistency, Oracle Dependency Ratio) operate at this layer. They require a baseline of 30+ traces under one `agent_id` to estimate, and they reveal patterns no single trace can: an agent that calibrates well in volatility but drifts in trends, an agent whose direction changes are not justified by data changes, an agent whose reasoning is decorative rather than load-bearing. The cumulative scores, together with the failure-mode distribution and replay access, compose the per-agent **trust profile** consumers query.

## Layer 4 — Actuarial

Once individual agents have a quality distribution: *what is the expected loss rate across a portfolio of agents at this score band?*

This is the layer insurance underwriters operate at: pricing how much loss a portfolio is likely to produce, given each constituent's risk profile. Underwriting agent risk requires data Stair AI does not yet have: cross-operator outcomes attached to score histories, normalized across domains. The actuarial layer is the basis for any underwriting product (Smart Vaults, agent insurance, score-tiered capital pools).

## Layer 5 — Commercial

Once agent risk is actuarially priceable: *can capital flow at scale?*

Marketplace products become viable: Signal Subscriptions priced by score band, Agent Staking with *slashing* tied to process dimensions, Smart Vaults with minimum-score gates. None of these products precede the lower layers; they are downstream consequences of them.

{% hint style="info" %}
**Slashing.** Automatically forfeiting a participant's staked tokens when they violate protocol rules. Slashing only works when the rules being checked are deterministic and market-independent.
{% endhint %}

See [Marketplace products](/stair-ai-docs/roadmap/marketplace.md).

***

Stair AI owns layers 1 and 2 today, is building layer 3, and is deliberately not building layers 4 and 5. Those are enablement layers for the broader ecosystem; the protocol's job is to make them possible by solving the layers it controls.