> 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/design-principles.md).

# Design principles

The protocol is built on six beliefs. Each is opinionated. Each is reflected in concrete design decisions throughout the system. Reading them is the fastest way to predict what Glass Box Protocol will and will not do.

## 1. Trust comes from process, not outcomes

A correct prediction may be a guess. A careful reasoning chain may be wrong. Capital allocators need to know how an agent thinks, not just whether it was right last quarter.

[Stair AI Score](/stair-ai-docs/concepts/stair-ai-score.md) evaluates reasoning quality independently of **P\&L** (profit and loss). Mixing the two creates the wrong incentives: penalizing low outcomes rewards risk-aversion, and rewarding high outcomes is fraud-tolerant for cherry-pickers. The full argument lives in the [scoring philosophy](/stair-ai-docs/scoring/philosophy.md).

## 2. Identity is the wrong primitive for agent trust

Wallets can be discarded; reasoning records persist. Glass Box Protocol attaches reputation to verifiable behavior, not to identity. The *BYOI* design accepts any cryptographic identifier the partner already uses, so reputation is portable across infrastructure changes. The `agent_id` is the canonical attribution key on every record and is decoupled from the wallets that may sign anchor transactions on the agent's behalf.

## 3. Verifiable means cryptographically verifiable

Procedural audits can be gamed. Cryptographic commitments cannot.

Every reasoning step is captured under a typed schema and stored append-only, in custody held independently of the agent's developer. Each record includes the cryptographic hash of its predecessor (a hash chain), so tampering is detected by hash comparison, not by Stair AI's judgment. Verification is binary, not adjudicated. Anchoring on a public chain is the upgrade path that closes agent-to-agent verification later, when the trust problem moves beyond a single auditor or risk officer.

## 4. Trust and adoption pull in opposite directions; offer both

Higher trust requires more runtime control. Runtime control raises adoption friction. The two cannot both be maximized in one product; they have to be exposed as separate dials.

Glass Box Protocol exposes [three integration modes](/stair-ai-docs/concepts/integration-modes.md): framework-native, SDK-instrumented, and self-attested. Each carries a different *attestation* label that travels with the score. Consumers know what residual trust assumption they are taking on; agents pick the integration cost that matches their use case.

## 5. Slashing requires market-independence

Penalties must only attach to dimensions an honest agent can pass through any market regime. A bad market cycle should never *slash* an honest agent.

Process dimensions like Internal Consistency and Information Breadth depend on the trace's structure, not the world the trace describes; they can anchor [slashing](/stair-ai-docs/scoring/slashing.md). Outcome dimensions like Confidence Calibration depend on market reality and cannot. The slashing-eligible subset is intentionally narrow.

## 6. Compose with the agent ecosystem; don't replace it

Stair AI is middleware. Agents run in any framework, talk to any LLM, anchor on any chain. The SDK is non-prescriptive about how the agent thinks; the only opinion is on the structure of reasoning records.

This makes the protocol a small commitment for partners to adopt and keeps Stair AI out of competition with the runtimes and providers it integrates with. Composability is the design constraint that determines what gets built and what gets left to the ecosystem.

***

The rest of the documentation makes these principles concrete. [Core concepts](/stair-ai-docs/concepts.md) defines the primitives that operationalize them. [Glass Box Protocol](/stair-ai-docs/protocol.md) describes the architecture that enforces them. [Reasoning Ledger SDK](/stair-ai-docs/sdk.md) is the integration surface that respects them.