> ## Documentation Index
> Fetch the complete documentation index at: https://docs.usetissue.xyz/llms.txt
> Use this file to discover all available pages before exploring further.

# Autonomous Operation & Continuous Improvement

> What Tissue actually does without a human in the loop, and what it deliberately keeps a human in the loop for.

<Warning>
  This page exists to be precise about a real design decision: Tissue does **not**
  self-modify its own policy, and it does **not** feed operator-facing analysis back into
  its own strategy. Both were deliberately not built, for reasons explained below — not
  because they were out of time, but because they would break guarantees this project is
  built around.
</Warning>

## What runs autonomously

Once started with valid credentials, the live daemon requires no human intervention to:

* Ingest both TxLINE SSE streams, reconnect on disconnect, and renew credentials on
  expiry
* Verify every message's source proof before admission
* Reprice, classify, risk-check, and quote on every admitted message
* Halt itself on any of the conditions in [Core Concepts → Discipline
  Engine](/core-concepts#the-discipline-engine)
* Recover deterministically after a crash — on restart, it verifies the persisted
  ledger prefix against a fresh replay of the authoritative corpus and repairs only the
  deterministic missing suffix, never trusting a locally cached receipt as source
  authority
* Submit pre-match and periodic checkpoint anchoring transactions
* Record signed policy snapshots when the effective policy changes across a restart

None of this requires a human to click anything. It is not, however, self-*improving*
in the sense of changing its own behavior based on results — that boundary is
intentional.

## Why there is no autonomous policy mutation

`policy.toml` is the single source of truth for every tunable constant in the pricing
and risk logic. That is a real, checkable property of this codebase: any judge can
`grep` the code and confirm no magic numbers exist outside it. A nightly job that
silently rewrote thresholds based on overnight results would erase that property —
every constant would no longer be traceable to a specific, explainable human decision
in git history.

## The honest alternative: human-reviewed tuning suggestions

Instead, `evaluate:tuning` runs a real evaluation:

```bash theme={null}
pnpm --filter @tissue/daemon evaluate:tuning
```

It evaluates a small grid of candidate values for a real policy parameter
(`strategy.edge_threshold_bps`) against the **calibration split only** — it never
touches the holdout fixtures, preserving the same calibration/holdout discipline used
throughout evaluation — and prints ranked suggestions with their evidence:

```json theme={null}
{
  "suggestions": [{
    "parameter": "strategy.edge_threshold_bps",
    "baselineValue": 150,
    "suggestedValue": 90,
    "confident": false,
    "reason": "Only 1 calibration fixture(s) — fewer than 3. This grid is presented for transparency only; it is NOT a statistically valid comparison..."
  }]
}
```

It **never writes `policy.toml`**. A human reads the output, verifies it against the
holdout split (`evaluate:calibration`), and manually edits the config — exactly like
any other change to this codebase, reviewable in a diff.

<Note>
  With only two real captured fixtures so far, the tool honestly reports itself as
  **underpowered** rather than presenting a confident recommendation. That flag is the
  point — it exists specifically to stop this tool from overfitting itself to noise.
</Note>

## Why the analyst layer never feeds back into strategy

The analyst service is a read-only forensics layer — it answers questions about what
the desk already did, using an LLM over a read-only SQLite projection of the ledger. It
opens its database connection with `readOnly: true` at the SQLite layer, has no write,
post, or execute tool of any kind, and this isolation is a tested guarantee (see
[Verifiability](/verifiability)).

Feeding the analyst's output — or a "memory" built from it — back into pricing or
strategy would mean an LLM's narration could eventually influence a real decision. That
is exactly the boundary this project refuses to cross. The analyst can find patterns
(see [`find_similar_decisions`](/reference#find_similar_decisions)) and explain them to
a human, but it cannot act, and it cannot change what Tissue will do next.

## The real proof-of-learning mechanism: the Strategy Arena

Rather than an opaque "the desk is improving" claim, Tissue's actual answer to "is this
working?" is the Strategy Arena and its regime ablation matrix (see
[Architecture](/architecture#the-strategy-arena-ab-comparison)): a real, computed,
re-runnable head-to-head between the full desk and a neutralized baseline, plus a
per-regime breakdown of which heuristic is actually contributing. This is graded with
the same CLV/Brier grader used everywhere else, on demand, over the same real captured
corpus — not asserted, not simulated.