# Experiment Provenance -- Design Specification

**Version:** 0.1.0-draft
**Status:** Proposed
**GAP Reference:** ARCHITECTURE-REVIEW.md, GAP 4
**Target Release:** v0.2.0
**Author:** QubitOS Team
**Date:** February 8, 2026
**Related Specs:** TIME-MODEL-SPEC.txt, ERROR-BUDGET-SPEC.txt

---

## 1. Problem Statement

Quantum experiment results are products of an entire context: the hardware calibration,
the pulse waveform, the optimization parameters that produced the waveform, and the
software stack that ran the computation. When a result changes -- fidelity drops,
measurement statistics shift, a gate that "worked yesterday" fails today -- the
fundamental question is: *what changed?*

Currently, QubitOS fingerprints only calibration data. This means:

- **Reproducibility is partial.** Two experiments with the same calibration fingerprint
  can produce different results if the GRAPE configuration, pulse duration, software
  version, or random seed differed. There is no way to verify that two experiments
  shared the same complete context.

- **Debugging is manual.** "The fidelity dropped" requires manually diffing calibration
  YAML files, checking whether someone changed the GRAPE learning rate, and hoping the
  software version was recorded somewhere. The fingerprint system only tells you "the
  calibration was different" -- not *what specifically* in the calibration changed, and
  nothing about changes outside calibration.

- **Scientific rigor requires full traceability.** Any published result should be
  traceable to a complete set of inputs. A measurement result without its full provenance
  is an anecdote, not data. The existing `PulseShape` proto has `calibration_fingerprint`,
  `code_version`, `created_at`, and `random_seed` fields, but `GrapeResult` does not
  populate them. `MeasurementResult` has a `calibration_fingerprint` field but no broader
  provenance.

- **The git-for-experiments pattern.** Version control works because every commit is a
  content-addressed snapshot, and `git diff` tells you exactly what changed. Quantum
  experiments need the same pattern: every result tagged with a root hash, and
  `provenance diff` identifies exactly which parameter changed between two runs.

The existing `CalibrationFingerprint` and `FingerprintStore` are approximately 60% of
the solution. This spec extends the fingerprint into a full Merkle tree covering all
dimensions of experiment context.

---

## 2. Current State Analysis

### 2.1 What Exists

**CalibrationFingerprint** (`src/qubitos/calibrator/fingerprint.py`):

| Field                 | Type                     | Hashed | Notes                                    |
|-----------------------|--------------------------|--------|------------------------------------------|
| `backend_name`        | `str`                    | Yes    | Identifies the quantum backend           |
| `timestamp`           | `str`                    | No     | ISO format, not included in hash         |
| `num_qubits`          | `int`                    | Yes    | Included in hash via `"num_qubits"` key  |
| `qubit_fingerprints`  | `list[dict[str, float]]` | Yes    | Per-qubit: index, freq, T1, T2, readout/gate fidelity |
| `coupler_fingerprints`| `list[dict[str, float]]` | Yes    | Per-coupler: qubit_a, qubit_b, coupling_mhz, cz_fidelity |
| `hash`                | `str`                    | --     | SHA-256, truncated to 16 hex characters  |

Hash computation: `json.dumps(data, sort_keys=True)` → SHA-256 → `[:16]`.

**FingerprintStore** (`src/qubitos/calibrator/fingerprint.py`):
- In-memory dict keyed by `backend_name`
- `max_history=100` fingerprints per backend
- Methods: `add()`, `get_latest()`, `get_history()`, `compute_drift_trend()`
- Drift trend compares consecutive fingerprints over a sliding window

**PulseShape proto** (`quantum/pulse/v1/pulse.proto`):

| Provenance Field            | Field Number | Status           |
|-----------------------------|--------------|------------------|
| `proto_version`             | 16           | Exists, unused   |
| `created_at`                | 17           | Exists, unused   |
| `calibration_fingerprint`   | 18           | Exists, unused   |
| `code_version`              | 19           | Exists, unused   |
| `random_seed`               | 20           | Exists, unused   |

These fields are defined in the proto but **not populated** by `GrapeOptimizer` or
`generate_pulse()`. The `GrapeResult` dataclass has no provenance fields.

**MeasurementResult proto** (`quantum/backend/v1/execution.proto`):
- Has `calibration_fingerprint` (field 9) and `backend_name` (field 7)
- No broader provenance hash
- No pulse sequence or GRAPE config reference

**GrapeConfig** (`src/qubitos/pulsegen/grape.py`):
- 10 parameters: `num_time_steps`, `duration_ns`, `target_fidelity`, `max_iterations`,
  `learning_rate`, `convergence_threshold`, `max_amplitude`, `use_second_order`,
  `regularization`, `random_seed`
- Not fingerprinted

**GrapeResult** (`src/qubitos/pulsegen/grape.py`):
- Contains: `i_envelope`, `q_envelope`, `fidelity`, `iterations`, `converged`,
  `fidelity_history`, `final_unitary`
- No provenance hash, no reference to the config that produced it

### 2.2 What's Missing

1. **Pulse sequence parameters are not fingerprinted.** Duration, time steps, envelope
   amplitudes, target gate -- none of these contribute to any hash.
2. **GRAPE optimization settings are not fingerprinted.** Learning rate, regularization,
   max_amplitude, convergence thresholds are invisible to the provenance system.
3. **Software/runtime versions are not captured.** Python version, NumPy/SciPy/QuTiP
   versions, QubitOS version, HAL server version -- none tracked systematically.
4. **No way to diff two experiments.** Given two measurement results, there is no
   mechanism to determine what differed in their complete context.
5. **PulseShape proto provenance fields exist but are dead code.** Five provenance
   fields on `PulseShape` are never populated.
6. **GrapeResult carries no provenance.** The optimization result has no link back to
   the configuration or calibration that produced it.

---

## 3. Design Goals

1. **Root hash on every result.** Every `MeasurementResult` is tagged with a single
   provenance root hash that uniquely identifies the complete experiment context.

2. **Merkle tree coverage.** The tree covers four dimensions: calibration state, pulse
   sequence definition, GRAPE/optimizer configuration, and software versions. A change
   in any leaf produces a different root hash.

3. **Efficient diffing.** `diff(hash_a, hash_b)` walks both trees and identifies exactly
   which nodes changed, with human-readable descriptions (e.g., "T1 on qubit 3 changed
   from 50.0 to 45.2 μs").

4. **Extend, don't replace.** The existing `CalibrationFingerprint` becomes a subtree
   of the provenance tree. `FingerprintStore` continues to work unchanged. No existing
   APIs break.

5. **Deterministic hashing.** Identical inputs always produce identical hashes, across
   platforms, Python versions, and byte orderings. Floating-point values are canonicalized
   before hashing.

6. **Low overhead.** SHA-256 hashing is cheap (microseconds for typical inputs). Envelope
   hashing uses raw byte representation, not JSON serialization. Provenance construction
   adds negligible latency to the GRAPE→execute pipeline.

7. **Human-readable summaries.** `provenance summary <hash>` prints a readable tree
   showing each node type, its hash prefix, and key parameters.

8. **Storage-efficient.** Store the tree structure (node types + hashes + metadata),
   not full copies of calibration data or pulse envelopes. Leaf content is reconstructable
   from the original data sources.

---

## 4. Non-Goals

- **Environmental conditions.** Temperature, EM interference, vibration, and other
  environmental factors affect results but are not available through software APIs.
  The provenance tree has an `annotations` field for recording environmental notes
  but does not hash them into the tree.

- **Hardware firmware version querying.** Querying firmware versions requires
  backend-specific APIs that do not yet exist. The `SoftwareVersionNode` has a
  placeholder `hal_firmware_version` field for future use.

- **Provenance-based result caching.** "Same provenance hash → skip re-execution and
  return cached result" is a natural extension but introduces cache invalidation
  complexity. Deferred to a future spec.

- **Distributed provenance.** Provenance trees from multiple machines or labs
  (federated experiments) require a synchronization protocol. Out of scope.

- **Full version control for experiments.** This spec provides content-addressed
  snapshots and pairwise diffing. Branching, merging, and history traversal (a full
  "git for experiments") are future work.

- **Automatic provenance-based regression detection.** Alerting when provenance
  changes correlate with fidelity drops requires a statistical analysis layer.
  Future work building on this spec.

---

## 5. Merkle Tree Architecture

### 5.1 Tree Structure

```
                        [Root Hash]
                       /           \
              [Calibration]     [Experiment]
              /          \       /         \
        [Qubit Cal]  [Coupler] [Pulse Seq] [Config]
         / | \          |       / | \        /    \
       Q0  Q1 ...     C01    P0  P1 ...  [GRAPE] [Software]
```

Each internal node's hash is computed from its children's hashes. Each leaf node's
hash is computed from its content. The root hash uniquely identifies the complete
experiment context.

**Properties:**

- **Collision resistance:** SHA-256 provides 128-bit collision resistance. Two distinct
  experiment contexts will have different root hashes with overwhelming probability.

- **Efficient verification:** To verify that a specific leaf (e.g., qubit 3 calibration)
  is part of a provenance tree, only the hashes along the path from leaf to root are
  needed -- O(log n) where n is the number of leaves.

- **Incremental updates:** When only calibration changes between experiments, only the
  calibration subtree hashes change. The experiment subtree hashes remain identical,
  making diffing efficient.

### 5.2 Node Types

#### CalibrationNode (internal)

Represents the complete calibration state of the backend. Its hash is computed from
its child nodes (QubitCalibration and CouplerCalibration), maintaining structural
consistency with the rest of the tree.

**Children:**
- `QubitCalibrationNode[]` -- one per qubit
- `CouplerCalibrationNode[]` -- one per coupler pair

**Hashable content (for standalone use when children unavailable):**
- `backend_name: str`
- `num_qubits: int`
- `qubit_fingerprints: list[dict]`
- `coupler_fingerprints: list[dict]`

**Compatibility:** When a `CalibrationFingerprint` already exists, its 16-char hash
is preserved in tree metadata as `legacy_calibration_hash` for backward-compatible
lookups. The tree hash is computed from children per Section 6.3.

#### QubitCalibrationNode (leaf)

Represents calibration data for a single qubit.

**Hashable content:**

| Field              | Type    | Units | Precision    |
|--------------------|---------|-------|--------------|
| `qubit_index`      | `int`   | --    | exact        |
| `frequency_ghz`    | `float` | GHz   | 12 sig. digits |
| `t1_us`            | `float` | μs    | 12 sig. digits |
| `t2_us`            | `float` | μs    | 12 sig. digits |
| `readout_fidelity` | `float` | --    | 12 sig. digits |
| `gate_fidelity`    | `float` | --    | 12 sig. digits |

#### CouplerCalibrationNode (leaf)

Represents calibration data for a qubit-qubit coupler.

**Hashable content:**

| Field           | Type    | Units | Precision    |
|-----------------|---------|-------|--------------|
| `qubit_a`       | `int`   | --    | exact        |
| `qubit_b`       | `int`   | --    | exact        |
| `coupling_mhz`  | `float` | MHz   | 12 sig. digits |
| `cz_fidelity`   | `float` | --    | 12 sig. digits |

#### PulseSequenceNode (internal)

Represents the complete pulse sequence submitted for execution.

**Children:**
- `ScheduledPulseNode[]` -- one per pulse in the sequence, ordered by schedule time

**Hashable metadata (in addition to children):**
- `num_pulses: int`
- `total_duration_ns: float`
- `temporal_constraints_hash: str` -- hash of any temporal constraints (see TIME-MODEL-SPEC.txt)

#### ScheduledPulseNode (leaf)

Represents a single pulse in a sequence.

**Hashable content:**

| Field               | Type       | Units | Precision        |
|---------------------|------------|-------|------------------|
| `pulse_id`          | `str`      | --    | exact            |
| `gate_type`         | `str`      | --    | exact (enum name)|
| `custom_unitary_hash` | `str`    | --    | hash of JSON if CUSTOM gate, else empty |
| `target_qubit_indices` | `list[int]` | -- | exact         |
| `duration_ns`       | `int`      | ns    | exact            |
| `num_time_steps`    | `int`      | --    | exact            |
| `target_fidelity`   | `float`    | --    | 12 sig. digits   |
| `max_amplitude_mhz` | `float`   | MHz   | 12 sig. digits   |
| `i_envelope_hash`   | `str`     | --    | SHA-256 of raw bytes |
| `q_envelope_hash`   | `str`     | --    | SHA-256 of raw bytes |
| `coupling_envelope_hash` | `str` | --   | SHA-256 of raw bytes, empty if absent |
| `rotation_angle`    | `float`   | rad   | 12 sig. digits, 0.0 if not rotation gate |

#### GRAPEConfigNode (leaf)

Represents the optimizer configuration that produced the pulse.

**Hashable content:**

| Field                    | Type    | Units | Precision        |
|--------------------------|---------|-------|------------------|
| `num_time_steps`         | `int`   | --    | exact            |
| `duration_ns`            | `float` | ns    | 12 sig. digits   |
| `target_fidelity`        | `float` | --    | 12 sig. digits   |
| `max_iterations`         | `int`   | --    | exact            |
| `learning_rate`          | `float` | --    | 12 sig. digits   |
| `convergence_threshold`  | `float` | --    | 12 sig. digits   |
| `max_amplitude`          | `float` | MHz   | 12 sig. digits   |
| `use_second_order`       | `bool`  | --    | exact            |
| `regularization`         | `float` | --    | 12 sig. digits   |
| `random_seed`            | `int or null` | -- | exact (null hashed as string "null") |
| `optimizer`              | `str`   | --    | exact (from GRAPEOptions, default "adam") |
| `l2_amplitude_penalty`   | `float` | --    | 12 sig. digits   |
| `smoothness_penalty`     | `float` | --    | 12 sig. digits   |
| `bandwidth_limit_mhz`    | `float` | MHz   | 12 sig. digits   |
| `include_decoherence`    | `bool`  | --    | exact            |
| `include_leakage`        | `bool`  | --    | exact            |

#### SoftwareVersionNode (leaf)

Captures the software stack versions at experiment time.

**Hashable content:**

| Field                   | Type  | Source                                  |
|-------------------------|-------|-----------------------------------------|
| `qubitos_core_version`  | `str` | `qubitos.__version__`                   |
| `qubitos_hal_version`   | `str` | HAL server `VERSION` constant           |
| `qubitos_proto_version` | `str` | Proto package version                   |
| `python_version`        | `str` | `sys.version` (major.minor.micro only)  |
| `numpy_version`         | `str` | `numpy.__version__`                     |
| `scipy_version`         | `str` | `scipy.__version__`                     |
| `qutip_version`         | `str` | `qutip.__version__` or `"not_installed"`|
| `core_git_sha`          | `str` | Git HEAD SHA of qubit-os-core, or `"unknown"` |
| `hal_git_sha`           | `str` | Git HEAD SHA of qubit-os-hardware, or `"unknown"` |
| `proto_git_sha`         | `str` | Git HEAD SHA of qubit-os-proto, or `"unknown"` |

Git SHAs are best-effort: if the code is running from an installed package rather than
a git checkout, the field is `"unknown"`.

---

## 6. Hashing Strategy

### 6.1 Algorithm

SHA-256, consistent with the existing `CalibrationFingerprint`. The full 64 hex-character
hash is stored internally. For display, hashes are truncated to 32 hex characters (128 bits)
at the root level and 16 hex characters at the node level, to balance readability with
collision resistance.

### 6.2 Leaf Node Hashing

Leaf nodes are hashed from a **canonical JSON representation** of their content:

```python
def hash_leaf(node_type: str, content: dict[str, Any]) -> str:
    """Hash a leaf node's content.

    Args:
        node_type: Node type identifier (e.g., "qubit_calibration")
        content: Dictionary of hashable fields

    Returns:
        64-character hex SHA-256 digest
    """
    # Prefix with node type to prevent cross-type collisions
    # (a QubitCalibrationNode and CouplerCalibrationNode with
    #  numerically identical fields must produce different hashes)
    canonical = {
        "__node_type__": node_type,
        **content,
    }
    json_bytes = json.dumps(
        canonical,
        sort_keys=True,
        separators=(",", ":"),  # No whitespace
    ).encode("utf-8")
    return hashlib.sha256(json_bytes).hexdigest()
```

### 6.3 Internal Node Hashing

Internal nodes are hashed from their children's hashes, sorted lexicographically:

```python
def hash_internal(node_type: str, child_hashes: list[str]) -> str:
    """Hash an internal node from its children's hashes.

    Args:
        node_type: Node type identifier
        child_hashes: List of child node hashes

    Returns:
        64-character hex SHA-256 digest
    """
    # Sort to ensure determinism regardless of child insertion order
    sorted_hashes = sorted(child_hashes)
    payload = f"{node_type}:{'|'.join(sorted_hashes)}"
    return hashlib.sha256(payload.encode("utf-8")).hexdigest()
```

### 6.4 Root Hash

The root hash combines the CalibrationNode hash and ExperimentNode hash:

```python
root_hash = hash_internal("root", [calibration_node.hash, experiment_node.hash])
```

Display format: first 32 hex characters. Full hash available via `.full_hash`.

### 6.5 Floating-Point Canonicalization

Floating-point values are rounded to 12 significant digits before hashing to prevent
spurious hash differences from platform-specific floating-point representations:

```python
def canonicalize_float(value: float) -> float:
    """Round float to 12 significant digits for deterministic hashing.

    12 significant digits is well beyond the precision of any quantum
    hardware parameter (frequencies are measured to ~6 digits, coherence
    times to ~4 digits) while being safely within float64's 15-16 digit
    precision.
    """
    if value == 0.0 or not math.isfinite(value):
        return 0.0
    magnitude = 10 ** (11 - int(math.floor(math.log10(abs(value)))))
    return round(value * magnitude) / magnitude
```

In JSON serialization, canonicalized floats are formatted with `f"{value:.12g}"` to
produce consistent string representations.

### 6.6 Envelope Hashing

Pulse envelopes (I, Q, coupling) are large arrays (typically 50-10,000 float64 values).
Serializing them to JSON for hashing would be slow and waste memory. Instead, envelopes
are hashed directly from their raw byte representation:

```python
def hash_envelope(envelope: np.ndarray) -> str:
    """Hash a pulse envelope from its raw bytes.

    The array is converted to C-contiguous float64 (IEEE 754 double)
    before hashing, ensuring cross-platform consistency.

    Args:
        envelope: 1D numpy array of float64 values

    Returns:
        64-character hex SHA-256 digest
    """
    arr = np.ascontiguousarray(envelope, dtype=np.float64)
    return hashlib.sha256(arr.tobytes()).hexdigest()
```

**Platform consistency note:** IEEE 754 double-precision is used universally on all
platforms QubitOS targets (x86-64 Linux, Apple Silicon macOS). The byte representation
of identical float64 values is identical across these platforms. The `C_CONTIGUOUS`
memory layout ensures byte ordering is consistent.

### 6.7 CalibrationFingerprint Compatibility

The existing `CalibrationFingerprint._compute_hash()` produces a 16-character truncated
SHA-256. For backward compatibility:

- The `CalibrationNode` stores both the legacy 16-char hash and the full 64-char hash.
- When constructing a provenance tree from an existing `CalibrationFingerprint`, the
  legacy hash is preserved as `legacy_calibration_hash` in metadata.
- The CalibrationNode's tree hash is computed from its children (per Section 6.3),
  which may differ from the legacy hash because the legacy hash includes `backend_name`
  in the hash data while the tree hash derives structure from individual child nodes.

---

## 7. Data Structures

### 7.1 ProvenanceNode

```python
# src/qubitos/provenance/nodes.py

from __future__ import annotations

import hashlib
import json
import math
from dataclasses import dataclass, field
from enum import Enum
from typing import Any


class NodeType(Enum):
    """Types of nodes in the provenance tree."""
    ROOT = "root"
    CALIBRATION = "calibration"
    QUBIT_CALIBRATION = "qubit_calibration"
    COUPLER_CALIBRATION = "coupler_calibration"
    EXPERIMENT = "experiment"
    PULSE_SEQUENCE = "pulse_sequence"
    SCHEDULED_PULSE = "scheduled_pulse"
    CONFIG = "config"
    GRAPE_CONFIG = "grape_config"
    SOFTWARE_VERSION = "software_version"


@dataclass(frozen=True)
class ProvenanceNode:
    """A node in the provenance Merkle tree.

    Leaf nodes have content and no children.
    Internal nodes have children and no content.

    Attributes:
        node_type: Type of this node
        hash: SHA-256 hash (64 hex characters)
        children: Child nodes (empty for leaf nodes)
        content: Hashable content (empty dict for internal nodes)
        label: Human-readable label (e.g., "qubit_0", "cnot_q0_q1")
    """
    node_type: NodeType
    hash: str
    children: tuple[ProvenanceNode, ...] = ()
    content: dict[str, Any] = field(default_factory=dict)
    label: str = ""

    @property
    def is_leaf(self) -> bool:
        return len(self.children) == 0

    @property
    def short_hash(self) -> str:
        """First 16 hex characters for display."""
        return self.hash[:16]
```

### 7.2 ProvenanceTree

```python
# src/qubitos/provenance/tree.py

from __future__ import annotations

import json
from dataclasses import dataclass, field
from datetime import datetime, timezone
from typing import Any

from .nodes import NodeType, ProvenanceNode


@dataclass
class ProvenanceTree:
    """Complete provenance tree for an experiment.

    The tree is immutable after construction. All hashes are computed
    during build and cannot be modified.

    Attributes:
        root: Root node of the Merkle tree
        timestamp: When this tree was constructed (ISO 8601)
        metadata: Additional context not included in the hash
    """
    root: ProvenanceNode
    timestamp: str = field(
        default_factory=lambda: datetime.now(timezone.utc).isoformat()
    )
    metadata: dict[str, Any] = field(default_factory=dict)

    @property
    def root_hash(self) -> str:
        """The root hash of the provenance tree."""
        return self.root.hash

    @property
    def short_hash(self) -> str:
        """First 32 hex characters of the root hash for display."""
        return self.root.hash[:32]

    def find_node(
        self, node_type: NodeType, label: str = ""
    ) -> ProvenanceNode | None:
        """Find a node by type and optional label.

        Args:
            node_type: Type of node to find
            label: Optional label to match

        Returns:
            First matching node, or None
        """
        return self._find_recursive(self.root, node_type, label)

    def _find_recursive(
        self, node: ProvenanceNode, target_type: NodeType, label: str
    ) -> ProvenanceNode | None:
        if node.node_type == target_type:
            if not label or node.label == label:
                return node
        for child in node.children:
            result = self._find_recursive(child, target_type, label)
            if result is not None:
                return result
        return None

    def all_leaves(self) -> list[ProvenanceNode]:
        """Return all leaf nodes in the tree."""
        leaves: list[ProvenanceNode] = []
        self._collect_leaves(self.root, leaves)
        return leaves

    def _collect_leaves(
        self, node: ProvenanceNode, acc: list[ProvenanceNode]
    ) -> None:
        if node.is_leaf:
            acc.append(node)
        else:
            for child in node.children:
                self._collect_leaves(child, acc)

    def diff(self, other: ProvenanceTree) -> ProvenanceDiff:
        """Compute the difference between this tree and another.

        Args:
            other: Tree to compare against

        Returns:
            ProvenanceDiff describing all changes
        """
        from .diff import NodeChange, ProvenanceDiff

        if self.root_hash == other.root_hash:
            return ProvenanceDiff(
                hash_a=self.root_hash,
                hash_b=other.root_hash,
                changed_nodes=[],
                unchanged_nodes=[
                    n.node_type.value for n in self.all_leaves()
                ],
            )

        changed: list[NodeChange] = []
        unchanged: list[str] = []
        self._diff_recursive(
            self.root, other.root, "", changed, unchanged
        )
        return ProvenanceDiff(
            hash_a=self.root_hash,
            hash_b=other.root_hash,
            changed_nodes=changed,
            unchanged_nodes=unchanged,
        )

    def _diff_recursive(
        self,
        node_a: ProvenanceNode,
        node_b: ProvenanceNode,
        path: str,
        changed: list,
        unchanged: list[str],
    ) -> None:
        from .diff import NodeChange

        current_path = f"{path}/{node_a.label}" if node_a.label else path
        if not current_path:
            current_path = node_a.node_type.value

        if node_a.hash == node_b.hash:
            for leaf in self._subtree_leaf_types(node_a):
                unchanged.append(leaf)
            return

        if node_a.is_leaf and node_b.is_leaf:
            description = self._describe_change(node_a, node_b)
            changed.append(NodeChange(
                node_type=node_a.node_type.value,
                path=current_path,
                old_hash=node_a.short_hash,
                new_hash=node_b.short_hash,
                old_content=node_a.content,
                new_content=node_b.content,
                description=description,
            ))
            return

        # Internal node changed -- recurse into children
        a_children = {
            (c.node_type, c.label): c for c in node_a.children
        }
        b_children = {
            (c.node_type, c.label): c for c in node_b.children
        }

        all_keys = set(a_children.keys()) | set(b_children.keys())
        for key in sorted(all_keys, key=lambda k: (k[0].value, k[1])):
            if key in a_children and key in b_children:
                self._diff_recursive(
                    a_children[key], b_children[key],
                    current_path, changed, unchanged,
                )
            elif key in a_children:
                changed.append(NodeChange(
                    node_type=key[0].value,
                    path=f"{current_path}/{key[1]}",
                    old_hash=a_children[key].short_hash,
                    new_hash="(removed)",
                    old_content=a_children[key].content,
                    new_content={},
                    description=f"Node removed: {key[0].value} '{key[1]}'",
                ))
            else:
                changed.append(NodeChange(
                    node_type=key[0].value,
                    path=f"{current_path}/{key[1]}",
                    old_hash="(added)",
                    new_hash=b_children[key].short_hash,
                    old_content={},
                    new_content=b_children[key].content,
                    description=f"Node added: {key[0].value} '{key[1]}'",
                ))

    def _subtree_leaf_types(self, node: ProvenanceNode) -> list[str]:
        if node.is_leaf:
            label = f"{node.node_type.value}"
            if node.label:
                label += f"/{node.label}"
            return [label]
        result: list[str] = []
        for child in node.children:
            result.extend(self._subtree_leaf_types(child))
        return result

    @staticmethod
    def _describe_change(a: ProvenanceNode, b: ProvenanceNode) -> str:
        """Generate a human-readable description of a leaf change."""
        diffs: list[str] = []
        all_keys = set(a.content.keys()) | set(b.content.keys())
        for key in sorted(all_keys):
            if key.startswith("__"):
                continue
            old_val = a.content.get(key)
            new_val = b.content.get(key)
            if old_val != new_val:
                if isinstance(old_val, float) and isinstance(new_val, float):
                    diffs.append(f"{key}: {old_val:.6g} -> {new_val:.6g}")
                else:
                    diffs.append(f"{key}: {old_val} -> {new_val}")
        return "; ".join(diffs) if diffs else "content changed"

    def to_dict(self) -> dict[str, Any]:
        """Serialize the tree to a dictionary for storage/transport."""
        return {
            "version": "1.0",
            "root_hash": self.root_hash,
            "timestamp": self.timestamp,
            "metadata": self.metadata,
            "tree": self._node_to_dict(self.root),
        }

    @staticmethod
    def _node_to_dict(node: ProvenanceNode) -> dict[str, Any]:
        result: dict[str, Any] = {
            "node_type": node.node_type.value,
            "hash": node.hash,
            "label": node.label,
        }
        if node.is_leaf:
            result["content"] = node.content
        else:
            result["children"] = [
                ProvenanceTree._node_to_dict(c) for c in node.children
            ]
        return result

    @classmethod
    def from_dict(cls, data: dict[str, Any]) -> ProvenanceTree:
        """Deserialize a tree from a dictionary."""
        tree_data = data["tree"]
        root = cls._node_from_dict(tree_data)
        return cls(
            root=root,
            timestamp=data.get("timestamp", ""),
            metadata=data.get("metadata", {}),
        )

    @classmethod
    def _node_from_dict(cls, data: dict[str, Any]) -> ProvenanceNode:
        node_type = NodeType(data["node_type"])
        if "children" in data:
            children = tuple(
                cls._node_from_dict(c) for c in data["children"]
            )
            return ProvenanceNode(
                node_type=node_type,
                hash=data["hash"],
                children=children,
                label=data.get("label", ""),
            )
        else:
            return ProvenanceNode(
                node_type=node_type,
                hash=data["hash"],
                content=data.get("content", {}),
                label=data.get("label", ""),
            )

    def summary(self) -> str:
        """Generate a human-readable summary of the provenance tree."""
        lines: list[str] = []
        lines.append(f"Provenance Tree [{self.short_hash}]")
        lines.append(f"  Timestamp: {self.timestamp}")
        if self.metadata:
            for k, v in sorted(self.metadata.items()):
                lines.append(f"  {k}: {v}")
        lines.append("")
        self._summary_recursive(self.root, lines, indent=0)
        return "\n".join(lines)

    def _summary_recursive(
        self,
        node: ProvenanceNode,
        lines: list[str],
        indent: int,
    ) -> None:
        prefix = "  " * indent
        label = f" '{node.label}'" if node.label else ""
        lines.append(
            f"{prefix}[{node.short_hash}] "
            f"{node.node_type.value}{label}"
        )
        if node.is_leaf and node.content:
            for key, value in sorted(node.content.items()):
                if key.startswith("__"):
                    continue
                if isinstance(value, str) and len(value) > 40:
                    value = value[:37] + "..."
                lines.append(f"{prefix}  {key}: {value}")
        for child in node.children:
            self._summary_recursive(child, lines, indent + 1)
```

### 7.3 ProvenanceDiff

```python
# src/qubitos/provenance/diff.py

from __future__ import annotations

from dataclasses import dataclass, field
from typing import Any


@dataclass
class NodeChange:
    """A single change between two provenance trees.

    Attributes:
        node_type: Type of the changed node
        path: Path from root (e.g., "calibration/qubit_0")
        old_hash: Hash in tree A (short form)
        new_hash: Hash in tree B (short form)
        old_content: Content in tree A (leaf nodes only)
        new_content: Content in tree B (leaf nodes only)
        description: Human-readable change description
    """
    node_type: str
    path: str
    old_hash: str
    new_hash: str
    old_content: dict[str, Any] = field(default_factory=dict)
    new_content: dict[str, Any] = field(default_factory=dict)
    description: str = ""


@dataclass
class ProvenanceDiff:
    """Difference between two provenance trees.

    Attributes:
        hash_a: Root hash of tree A
        hash_b: Root hash of tree B
        changed_nodes: List of nodes that differ
        unchanged_nodes: List of node type paths that are identical
    """
    hash_a: str
    hash_b: str
    changed_nodes: list[NodeChange] = field(default_factory=list)
    unchanged_nodes: list[str] = field(default_factory=list)

    @property
    def is_identical(self) -> bool:
        return len(self.changed_nodes) == 0

    @property
    def num_changes(self) -> int:
        return len(self.changed_nodes)

    def summary(self) -> str:
        """Human-readable diff summary."""
        lines: list[str] = []
        lines.append(
            f"Provenance Diff: {self.hash_a[:16]} <-> {self.hash_b[:16]}"
        )

        if self.is_identical:
            lines.append("  Trees are identical.")
            return "\n".join(lines)

        lines.append(f"  {self.num_changes} change(s):")
        for change in self.changed_nodes:
            lines.append("")
            lines.append(f"  [{change.node_type}] {change.path}")
            lines.append(
                f"    {change.old_hash} -> {change.new_hash}"
            )
            if change.description:
                lines.append(f"    {change.description}")

        if self.unchanged_nodes:
            lines.append("")
            lines.append(
                f"  {len(self.unchanged_nodes)} unchanged node(s):"
            )
            for node_path in self.unchanged_nodes:
                lines.append(f"    = {node_path}")

        return "\n".join(lines)

    def to_dict(self) -> dict[str, Any]:
        """Serialize diff to dictionary."""
        return {
            "hash_a": self.hash_a,
            "hash_b": self.hash_b,
            "num_changes": self.num_changes,
            "changed_nodes": [
                {
                    "node_type": c.node_type,
                    "path": c.path,
                    "old_hash": c.old_hash,
                    "new_hash": c.new_hash,
                    "description": c.description,
                }
                for c in self.changed_nodes
            ],
            "unchanged_nodes": self.unchanged_nodes,
        }
```

### 7.4 ProvenanceStore

```python
# src/qubitos/provenance/store.py

from __future__ import annotations

import json
import logging
from collections import OrderedDict
from pathlib import Path
from typing import Any

from .diff import ProvenanceDiff
from .tree import ProvenanceTree

logger = logging.getLogger(__name__)


class ProvenanceStore:
    """Persistent store for provenance trees.

    Stores trees indexed by root hash. Provides retrieval and diffing.
    Can operate in-memory or backed by a JSON file for persistence.

    This complements (not replaces) FingerprintStore. FingerprintStore
    tracks calibration drift over time for a specific backend.
    ProvenanceStore indexes complete experiment provenance trees by hash
    for retrieval and comparison.

    Attributes:
        max_entries: Maximum number of trees to store
    """

    def __init__(
        self,
        max_entries: int = 1000,
        persist_path: Path | None = None,
    ):
        """Initialize the provenance store.

        Args:
            max_entries: Maximum trees to keep (LRU eviction)
            persist_path: Optional file path for JSON persistence.
                         If None, store is in-memory only.
        """
        self.max_entries = max_entries
        self._persist_path = persist_path
        self._trees: OrderedDict[str, ProvenanceTree] = OrderedDict()

        if persist_path and persist_path.exists():
            self._load_from_disk()

    def store(self, tree: ProvenanceTree) -> str:
        """Store a provenance tree.

        Args:
            tree: Tree to store

        Returns:
            The root hash of the stored tree
        """
        root_hash = tree.root_hash
        self._trees[root_hash] = tree
        self._trees.move_to_end(root_hash)

        # Evict oldest entries if over limit
        while len(self._trees) > self.max_entries:
            evicted_hash, _ = self._trees.popitem(last=False)
            logger.debug(
                f"Evicted provenance tree {evicted_hash[:16]}"
            )

        if self._persist_path:
            self._save_to_disk()

        return root_hash

    def get(self, root_hash: str) -> ProvenanceTree | None:
        """Retrieve a tree by root hash.

        Supports prefix matching: if root_hash is shorter than 64
        characters, matches the first tree whose hash starts with
        the given prefix.

        Args:
            root_hash: Full or prefix of root hash

        Returns:
            The tree, or None if not found
        """
        # Exact match
        if root_hash in self._trees:
            return self._trees[root_hash]

        # Prefix match
        for stored_hash, tree in self._trees.items():
            if stored_hash.startswith(root_hash):
                return tree

        return None

    def diff(
        self, hash_a: str, hash_b: str
    ) -> ProvenanceDiff | None:
        """Compute diff between two stored trees.

        Args:
            hash_a: Root hash (or prefix) of first tree
            hash_b: Root hash (or prefix) of second tree

        Returns:
            ProvenanceDiff, or None if either tree not found
        """
        tree_a = self.get(hash_a)
        tree_b = self.get(hash_b)
        if tree_a is None or tree_b is None:
            return None
        return tree_a.diff(tree_b)

    def history(
        self, limit: int | None = None
    ) -> list[ProvenanceTree]:
        """Return stored trees in insertion order (oldest first).

        Args:
            limit: Maximum number of trees to return

        Returns:
            List of trees
        """
        trees = list(self._trees.values())
        if limit:
            trees = trees[-limit:]
        return trees

    def contains(self, root_hash: str) -> bool:
        """Check if a tree with the given hash exists."""
        return self.get(root_hash) is not None

    def __len__(self) -> int:
        return len(self._trees)

    def _save_to_disk(self) -> None:
        """Persist store to JSON file."""
        assert self._persist_path is not None
        data = {
            "version": "1.0",
            "trees": {
                h: t.to_dict() for h, t in self._trees.items()
            },
        }
        self._persist_path.parent.mkdir(parents=True, exist_ok=True)
        with open(self._persist_path, "w") as f:
            json.dump(data, f, indent=2)

    def _load_from_disk(self) -> None:
        """Load store from JSON file."""
        assert self._persist_path is not None
        try:
            with open(self._persist_path) as f:
                data = json.load(f)
            for _hash, tree_data in data.get("trees", {}).items():
                tree = ProvenanceTree.from_dict(tree_data)
                self._trees[tree.root_hash] = tree
            logger.info(
                f"Loaded {len(self._trees)} provenance trees "
                f"from {self._persist_path}"
            )
        except (json.JSONDecodeError, KeyError) as e:
            logger.warning(
                f"Failed to load provenance store "
                f"from {self._persist_path}: {e}"
            )
```

### 7.5 ProvenanceBuilder

```python
# src/qubitos/provenance/builder.py

from __future__ import annotations

import hashlib
import json
import math
import sys
from typing import Any

import numpy as np

from .nodes import NodeType, ProvenanceNode
from .tree import ProvenanceTree


def canonicalize_float(value: float) -> float:
    """Round float to 12 significant digits for deterministic hashing."""
    if value == 0.0 or not math.isfinite(value):
        return 0.0
    magnitude = 10 ** (11 - int(math.floor(math.log10(abs(value)))))
    return round(value * magnitude) / magnitude


def _hash_leaf(node_type: str, content: dict[str, Any]) -> str:
    """Compute SHA-256 of canonical JSON for a leaf node."""
    canonical = {"__node_type__": node_type, **content}
    json_bytes = json.dumps(
        canonical, sort_keys=True, separators=(",", ":")
    ).encode("utf-8")
    return hashlib.sha256(json_bytes).hexdigest()


def _hash_internal(node_type: str, child_hashes: list[str]) -> str:
    """Compute SHA-256 of sorted child hashes for an internal node."""
    sorted_hashes = sorted(child_hashes)
    payload = f"{node_type}:{'|'.join(sorted_hashes)}"
    return hashlib.sha256(payload.encode("utf-8")).hexdigest()


def hash_envelope(envelope: np.ndarray) -> str:
    """Hash a pulse envelope from its raw float64 bytes."""
    arr = np.ascontiguousarray(envelope, dtype=np.float64)
    return hashlib.sha256(arr.tobytes()).hexdigest()


class ProvenanceBuilder:
    """Builds provenance trees from experiment components.

    Usage:
        builder = ProvenanceBuilder()
        builder.set_calibration_from_fingerprint(fingerprint)
        builder.set_grape_config(config)
        builder.add_pulse(result, gate_type, target_qubits, ...)
        builder.set_software_versions()
        tree = builder.build()
    """

    def __init__(self) -> None:
        self._calibration_node: ProvenanceNode | None = None
        self._pulse_nodes: list[ProvenanceNode] = []
        self._grape_config_node: ProvenanceNode | None = None
        self._software_node: ProvenanceNode | None = None
        self._metadata: dict[str, Any] = {}

    def set_metadata(
        self, key: str, value: Any
    ) -> ProvenanceBuilder:
        """Add metadata (not hashed, for human context)."""
        self._metadata[key] = value
        return self

    def set_calibration_from_fingerprint(
        self, fingerprint: Any  # CalibrationFingerprint
    ) -> ProvenanceBuilder:
        """Build calibration subtree from a CalibrationFingerprint.

        Decomposes the fingerprint into individual QubitCalibrationNode
        and CouplerCalibrationNode children, then assembles the
        CalibrationNode as their parent.
        """
        qubit_nodes: list[ProvenanceNode] = []
        for qfp in fingerprint.qubit_fingerprints:
            content = {
                "qubit_index": int(qfp["index"]),
                "frequency_ghz": canonicalize_float(
                    qfp["frequency_ghz"]
                ),
                "t1_us": canonicalize_float(qfp["t1_us"]),
                "t2_us": canonicalize_float(qfp["t2_us"]),
                "readout_fidelity": canonicalize_float(
                    qfp["readout_fidelity"]
                ),
                "gate_fidelity": canonicalize_float(
                    qfp["gate_fidelity"]
                ),
            }
            node_hash = _hash_leaf("qubit_calibration", content)
            qubit_nodes.append(ProvenanceNode(
                node_type=NodeType.QUBIT_CALIBRATION,
                hash=node_hash,
                content=content,
                label=f"qubit_{int(qfp['index'])}",
            ))

        coupler_nodes: list[ProvenanceNode] = []
        for cfp in fingerprint.coupler_fingerprints:
            content = {
                "qubit_a": int(cfp["qubit_a"]),
                "qubit_b": int(cfp["qubit_b"]),
                "coupling_mhz": canonicalize_float(
                    cfp["coupling_mhz"]
                ),
                "cz_fidelity": canonicalize_float(
                    cfp["cz_fidelity"]
                ),
            }
            node_hash = _hash_leaf("coupler_calibration", content)
            qa = int(cfp["qubit_a"])
            qb = int(cfp["qubit_b"])
            coupler_nodes.append(ProvenanceNode(
                node_type=NodeType.COUPLER_CALIBRATION,
                hash=node_hash,
                content=content,
                label=f"coupler_{qa}_{qb}",
            ))

        all_children = qubit_nodes + coupler_nodes
        cal_hash = _hash_internal(
            "calibration", [c.hash for c in all_children]
        )
        self._calibration_node = ProvenanceNode(
            node_type=NodeType.CALIBRATION,
            hash=cal_hash,
            children=tuple(all_children),
            label="calibration",
        )
        self._metadata["legacy_calibration_hash"] = fingerprint.hash
        return self

    def set_grape_config(
        self, config: Any  # GrapeConfig
    ) -> ProvenanceBuilder:
        """Build GRAPEConfigNode from a GrapeConfig instance."""
        content = {
            "num_time_steps": config.num_time_steps,
            "duration_ns": canonicalize_float(config.duration_ns),
            "target_fidelity": canonicalize_float(
                config.target_fidelity
            ),
            "max_iterations": config.max_iterations,
            "learning_rate": canonicalize_float(config.learning_rate),
            "convergence_threshold": canonicalize_float(
                config.convergence_threshold
            ),
            "max_amplitude": canonicalize_float(config.max_amplitude),
            "use_second_order": config.use_second_order,
            "regularization": canonicalize_float(
                config.regularization
            ),
            "random_seed": config.random_seed
            if config.random_seed is not None
            else "null",
        }
        node_hash = _hash_leaf("grape_config", content)
        self._grape_config_node = ProvenanceNode(
            node_type=NodeType.GRAPE_CONFIG,
            hash=node_hash,
            content=content,
            label="grape_config",
        )
        return self

    def add_pulse(
        self,
        pulse_id: str,
        gate_type: str,
        target_qubit_indices: list[int],
        duration_ns: int,
        num_time_steps: int,
        target_fidelity: float,
        max_amplitude_mhz: float,
        i_envelope: np.ndarray,
        q_envelope: np.ndarray,
        coupling_envelope: np.ndarray | None = None,
        rotation_angle: float = 0.0,
        custom_unitary_json: str = "",
    ) -> ProvenanceBuilder:
        """Add a pulse to the sequence."""
        content = {
            "pulse_id": pulse_id,
            "gate_type": gate_type,
            "custom_unitary_hash": hashlib.sha256(
                custom_unitary_json.encode()
            ).hexdigest()
            if custom_unitary_json
            else "",
            "target_qubit_indices": sorted(target_qubit_indices),
            "duration_ns": duration_ns,
            "num_time_steps": num_time_steps,
            "target_fidelity": canonicalize_float(target_fidelity),
            "max_amplitude_mhz": canonicalize_float(
                max_amplitude_mhz
            ),
            "i_envelope_hash": hash_envelope(i_envelope),
            "q_envelope_hash": hash_envelope(q_envelope),
            "coupling_envelope_hash": hash_envelope(coupling_envelope)
            if coupling_envelope is not None
            else "",
            "rotation_angle": canonicalize_float(rotation_angle),
        }
        node_hash = _hash_leaf("scheduled_pulse", content)
        qubits_str = "_".join(
            str(q) for q in target_qubit_indices
        )
        self._pulse_nodes.append(ProvenanceNode(
            node_type=NodeType.SCHEDULED_PULSE,
            hash=node_hash,
            content=content,
            label=f"{gate_type.lower()}_q{qubits_str}",
        ))
        return self

    def set_software_versions(
        self,
        hal_version: str = "unknown",
        proto_version: str = "unknown",
        core_git_sha: str = "unknown",
        hal_git_sha: str = "unknown",
        proto_git_sha: str = "unknown",
    ) -> ProvenanceBuilder:
        """Build SoftwareVersionNode from runtime information."""
        import numpy

        try:
            import scipy
            scipy_ver = scipy.__version__
        except ImportError:
            scipy_ver = "not_installed"

        try:
            import qutip
            qutip_ver = qutip.__version__
        except ImportError:
            qutip_ver = "not_installed"

        import qubitos

        content = {
            "qubitos_core_version": qubitos.__version__,
            "qubitos_hal_version": hal_version,
            "qubitos_proto_version": proto_version,
            "python_version": (
                f"{sys.version_info.major}."
                f"{sys.version_info.minor}."
                f"{sys.version_info.micro}"
            ),
            "numpy_version": numpy.__version__,
            "scipy_version": scipy_ver,
            "qutip_version": qutip_ver,
            "core_git_sha": core_git_sha,
            "hal_git_sha": hal_git_sha,
            "proto_git_sha": proto_git_sha,
        }
        node_hash = _hash_leaf("software_version", content)
        self._software_node = ProvenanceNode(
            node_type=NodeType.SOFTWARE_VERSION,
            hash=node_hash,
            content=content,
            label="software",
        )
        return self

    def build(self) -> ProvenanceTree:
        """Construct the complete provenance tree.

        Raises:
            ValueError: If required components are missing
        """
        if self._calibration_node is None:
            raise ValueError(
                "Calibration not set. "
                "Call set_calibration_from_fingerprint()."
            )

        # Build PulseSequenceNode
        if self._pulse_nodes:
            pulse_seq_hash = _hash_internal(
                "pulse_sequence",
                [p.hash for p in self._pulse_nodes],
            )
            pulse_seq_node = ProvenanceNode(
                node_type=NodeType.PULSE_SEQUENCE,
                hash=pulse_seq_hash,
                children=tuple(self._pulse_nodes),
                label="pulse_sequence",
            )
        else:
            pulse_seq_hash = _hash_leaf(
                "pulse_sequence", {"empty": True}
            )
            pulse_seq_node = ProvenanceNode(
                node_type=NodeType.PULSE_SEQUENCE,
                hash=pulse_seq_hash,
                content={"empty": True},
                label="pulse_sequence",
            )

        # Build ConfigNode
        config_children: list[ProvenanceNode] = []
        if self._grape_config_node:
            config_children.append(self._grape_config_node)
        if self._software_node:
            config_children.append(self._software_node)

        if config_children:
            config_hash = _hash_internal(
                "config", [c.hash for c in config_children]
            )
            config_node = ProvenanceNode(
                node_type=NodeType.CONFIG,
                hash=config_hash,
                children=tuple(config_children),
                label="config",
            )
        else:
            config_hash = _hash_leaf("config", {"empty": True})
            config_node = ProvenanceNode(
                node_type=NodeType.CONFIG,
                hash=config_hash,
                content={"empty": True},
                label="config",
            )

        # Build ExperimentNode
        experiment_children = (pulse_seq_node, config_node)
        experiment_hash = _hash_internal(
            "experiment",
            [c.hash for c in experiment_children],
        )
        experiment_node = ProvenanceNode(
            node_type=NodeType.EXPERIMENT,
            hash=experiment_hash,
            children=experiment_children,
            label="experiment",
        )

        # Build Root
        root_children = (self._calibration_node, experiment_node)
        root_hash = _hash_internal(
            "root", [c.hash for c in root_children]
        )
        root_node = ProvenanceNode(
            node_type=NodeType.ROOT,
            hash=root_hash,
            children=root_children,
        )

        return ProvenanceTree(
            root=root_node,
            metadata=self._metadata,
        )
```

---

## 8. Protocol Buffer Changes

### 8.1 New Proto File: `quantum/common/v1/provenance.proto`

```protobuf
// QubitOS Experiment Provenance
// Copyright 2026 QubitOS Contributors
// SPDX-License-Identifier: Apache-2.0

syntax = "proto3";

package quantum.common.v1;

import "quantum/common/v1/common.proto";

option java_multiple_files = true;
option java_package = "io.qubitos.common.v1";

// ProvenanceNodeType enumerates node types in the Merkle tree.
enum ProvenanceNodeType {
  PROVENANCE_NODE_TYPE_UNSPECIFIED = 0;
  PROVENANCE_NODE_TYPE_ROOT = 1;
  PROVENANCE_NODE_TYPE_CALIBRATION = 2;
  PROVENANCE_NODE_TYPE_QUBIT_CALIBRATION = 3;
  PROVENANCE_NODE_TYPE_COUPLER_CALIBRATION = 4;
  PROVENANCE_NODE_TYPE_EXPERIMENT = 5;
  PROVENANCE_NODE_TYPE_PULSE_SEQUENCE = 6;
  PROVENANCE_NODE_TYPE_SCHEDULED_PULSE = 7;
  PROVENANCE_NODE_TYPE_CONFIG = 8;
  PROVENANCE_NODE_TYPE_GRAPE_CONFIG = 9;
  PROVENANCE_NODE_TYPE_SOFTWARE_VERSION = 10;
}

// ProvenanceNode is a single node in the provenance Merkle tree.
message ProvenanceNode {
  // Type of this node.
  ProvenanceNodeType node_type = 1;

  // SHA-256 hash of this node (64 hex characters).
  // Leaf: hash of canonical JSON content.
  // Internal: hash of sorted child hashes.
  string hash = 2;

  // Human-readable label (e.g., "qubit_0", "grape_config").
  string label = 3;

  // Child nodes (for internal nodes only).
  repeated ProvenanceNode children = 4;

  // Leaf content as JSON string (for leaf nodes only).
  // Empty for internal nodes.
  string content_json = 5;

  // Reserved for future use.
  reserved 50 to 100;
}

// ProvenanceTree is the complete Merkle tree for an experiment.
message ProvenanceTree {
  // Root node of the tree.
  ProvenanceNode root = 1;

  // Root hash (convenience, equals root.hash).
  string root_hash = 2;

  // When this tree was constructed.
  quantum.common.v1.Timestamp created_at = 3;

  // Human-readable metadata (not included in hash).
  // Keys: "legacy_calibration_hash", "error_budget_*", etc.
  map<string, string> metadata = 4;

  // Reserved for future use.
  reserved 50 to 100;
}

// ProvenanceDiff describes the difference between two trees.
message ProvenanceDiff {
  // Root hash of the first tree.
  string hash_a = 1;

  // Root hash of the second tree.
  string hash_b = 2;

  // Nodes that changed between the two trees.
  repeated NodeChange changed_nodes = 3;

  // Node type paths that are identical between trees.
  repeated string unchanged_nodes = 4;

  // Reserved for future use.
  reserved 50 to 100;
}

// NodeChange describes a single changed node in a diff.
message NodeChange {
  // Type of the changed node (e.g., "qubit_calibration").
  string node_type = 1;

  // Path from root (e.g., "calibration/qubit_0").
  string path = 2;

  // Hash in tree A (16-char prefix).
  string old_hash = 3;

  // Hash in tree B (16-char prefix).
  string new_hash = 4;

  // Human-readable description of the change.
  // Example: "t1_us: 50 -> 45.2"
  string description = 5;

  // Reserved for future use.
  reserved 50 to 100;
}
```

### 8.2 Changes to Existing Protos

**`quantum/backend/v1/execution.proto` -- MeasurementResult:**

Add field 13:

```protobuf
message MeasurementResult {
  // ... existing fields 1-12 unchanged ...

  // Provenance root hash for this measurement result.
  // Links to the complete provenance tree covering calibration,
  // pulse sequence, optimizer config, and software versions.
  // Format: 64-character hex SHA-256 digest.
  // Empty if provenance tracking was not enabled for this execution.
  string provenance_hash = 13;

  // Reserved for future use.
  reserved 50 to 100;
}
```

**`quantum/pulse/v1/pulse.proto` -- PulseShape:**

Add field 22:

```protobuf
message PulseShape {
  // ... existing fields 1-21 unchanged ...

  // Complete provenance tree for this pulse.
  // Includes calibration state, GRAPE config, and software versions
  // at the time the pulse was generated.
  // Optional: not present for pulses created before provenance was added.
  quantum.common.v1.ProvenanceTree provenance = 22;

  // Reserved for future use.
  reserved 50 to 100;
}
```

**`quantum/pulse/v1/grape.proto` -- OptimizeResponse:**

Add field 13:

```protobuf
message OptimizeResponse {
  // ... existing fields 1-12 unchanged ...

  // Provenance root hash for this optimization run.
  // Covers: calibration + GRAPE config + software versions.
  // Does not include execution context (that comes from MeasurementResult).
  string provenance_hash = 13;

  // Reserved for future use.
  reserved 50 to 100;
}
```

### 8.3 Proto Versioning Note

All additions are backward-compatible per protobuf wire format guarantees:
- New fields use previously unreserved field numbers
- Old clients silently ignore unrecognized fields
- New clients treat missing fields as their default values (empty string for
  `provenance_hash`, unset message for `provenance`)
- No existing field semantics change

The `quantum/common/v1/provenance.proto` import must be added to the `build.rs`
proto compilation list in `qubit-os-hardware` and to the `protoc` invocation in
`qubit-os-proto`.

---

## 9. Rust Implementation

### 9.1 Module Structure

```
qubit-os-hardware/src/
├── provenance/
│   ├── mod.rs          // Module re-exports
│   ├── tree.rs         // ProvenanceTree, ProvenanceNode structs
│   ├── hash.rs         // SHA-256 hashing functions
│   ├── diff.rs         // ProvenanceDiff computation
│   └── store.rs        // In-memory ProvenanceStore
├── backend/
│   ├── mod.rs          // (existing) add provenance_hash to execution
│   └── ...
└── server/
    ├── grpc.rs         // (existing) populate provenance on responses
    └── ...
```

Add `pub mod provenance;` to `src/lib.rs`.

### 9.2 Core Types (`tree.rs`)

```rust
use serde::{Deserialize, Serialize};
use std::collections::HashMap;

/// Types of nodes in the provenance Merkle tree.
#[derive(Debug, Clone, PartialEq, Eq, Hash, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum NodeType {
    Root,
    Calibration,
    QubitCalibration,
    CouplerCalibration,
    Experiment,
    PulseSequence,
    ScheduledPulse,
    Config,
    GrapeConfig,
    SoftwareVersion,
}

impl NodeType {
    pub fn as_str(&self) -> &'static str {
        match self {
            Self::Root => "root",
            Self::Calibration => "calibration",
            Self::QubitCalibration => "qubit_calibration",
            Self::CouplerCalibration => "coupler_calibration",
            Self::Experiment => "experiment",
            Self::PulseSequence => "pulse_sequence",
            Self::ScheduledPulse => "scheduled_pulse",
            Self::Config => "config",
            Self::GrapeConfig => "grape_config",
            Self::SoftwareVersion => "software_version",
        }
    }
}

/// A single node in the provenance Merkle tree.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ProvenanceNode {
    pub node_type: NodeType,
    /// SHA-256 hash, 64 hex characters.
    pub hash: String,
    /// Human-readable label.
    pub label: String,
    /// Child nodes (empty for leaf nodes).
    pub children: Vec<ProvenanceNode>,
    /// Leaf content as key-value pairs (empty for internal nodes).
    #[serde(default)]
    pub content: HashMap<String, serde_json::Value>,
}

impl ProvenanceNode {
    pub fn is_leaf(&self) -> bool {
        self.children.is_empty()
    }

    pub fn short_hash(&self) -> &str {
        let end = self.hash.len().min(16);
        &self.hash[..end]
    }
}

/// Complete provenance tree for an experiment.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ProvenanceTree {
    pub root: ProvenanceNode,
    pub root_hash: String,
    pub timestamp: String,
    #[serde(default)]
    pub metadata: HashMap<String, String>,
}
```

### 9.3 Hashing (`hash.rs`)

```rust
use sha2::{Digest, Sha256};
use serde_json::Value;
use std::collections::BTreeMap;

/// Hash leaf node content using canonical JSON.
///
/// The node_type is included as a "__node_type__" key to prevent
/// cross-type hash collisions.
pub fn hash_leaf(
    node_type: &str,
    content: &BTreeMap<String, Value>,
) -> String {
    let mut canonical = BTreeMap::new();
    canonical.insert(
        "__node_type__".to_string(),
        Value::String(node_type.to_string()),
    );
    for (k, v) in content {
        canonical.insert(k.clone(), v.clone());
    }
    // BTreeMap serializes with sorted keys (guaranteed by BTreeMap)
    // serde_json with no pretty-printing produces compact JSON
    let json_str = serde_json::to_string(&canonical)
        .expect("BTreeMap serialization should not fail");
    let hash = Sha256::digest(json_str.as_bytes());
    hex::encode(hash)
}

/// Hash internal node from sorted child hashes.
///
/// Format: "{node_type}:{hash1}|{hash2}|..."
/// Child hashes are sorted lexicographically before joining.
pub fn hash_internal(
    node_type: &str,
    child_hashes: &[String],
) -> String {
    let mut sorted = child_hashes.to_vec();
    sorted.sort();
    let payload = format!(
        "{}:{}",
        node_type,
        sorted.join("|")
    );
    let hash = Sha256::digest(payload.as_bytes());
    hex::encode(hash)
}

/// Hash a pulse envelope from raw f64 bytes (little-endian).
pub fn hash_envelope(envelope: &[f64]) -> String {
    let bytes: Vec<u8> = envelope
        .iter()
        .flat_map(|v| v.to_le_bytes())
        .collect();
    let hash = Sha256::digest(&bytes);
    hex::encode(hash)
}

/// Canonicalize a float to 12 significant digits.
///
/// Returns 0.0 for zero, NaN, or infinity inputs.
pub fn canonicalize_float(value: f64) -> f64 {
    if value == 0.0 || !value.is_finite() {
        return 0.0;
    }
    let exp = 11 - value.abs().log10().floor() as i32;
    let magnitude = 10f64.powi(exp);
    (value * magnitude).round() / magnitude
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_canonicalize_float_zero() {
        assert_eq!(canonicalize_float(0.0), 0.0);
        assert_eq!(canonicalize_float(-0.0), 0.0);
    }

    #[test]
    fn test_canonicalize_float_nan() {
        assert_eq!(canonicalize_float(f64::NAN), 0.0);
    }

    #[test]
    fn test_hash_internal_sorted() {
        let h1 = hash_internal("test", &[
            "bbb".to_string(),
            "aaa".to_string(),
        ]);
        let h2 = hash_internal("test", &[
            "aaa".to_string(),
            "bbb".to_string(),
        ]);
        assert_eq!(h1, h2);
    }
}
```

**Important note on cross-language consistency:** The Rust `hash_envelope` uses
`to_le_bytes()` (little-endian). The Python `hash_envelope` uses
`np.ascontiguousarray(..., dtype=np.float64).tobytes()`, which produces native byte
order. On all QubitOS target platforms (x86-64, aarch64 Linux), native byte order is
little-endian, so the hashes match. If big-endian platforms are ever needed, both
implementations must agree on byte order.

For `hash_leaf` and `hash_internal`, cross-language consistency requires that both
Python's `json.dumps(sort_keys=True, separators=(",", ":"))` and Rust's
`serde_json::to_string()` with `BTreeMap` produce identical JSON strings. This holds
for the data types used (strings, integers, booleans, floats formatted identically).
Integration tests in Section 13.6 verify this.

### 9.4 Integration with Execution Path

In `server/grpc.rs`, the `execute_pulse` handler is extended:

```rust
// In the execute_pulse handler:

// 1. Extract provenance hash from incoming PulseShape (if present)
let provenance_hash = request
    .pulse
    .as_ref()
    .and_then(|p| p.provenance.as_ref())
    .map(|t| t.root_hash.clone())
    .unwrap_or_default();

// 2. Execute the pulse (existing code unchanged)
let result = backend.execute(/* ... */)?;

// 3. Attach provenance_hash to MeasurementResult
let mut measurement = result.measurement;
measurement.provenance_hash = provenance_hash;
```

The HAL server does not construct or modify provenance trees. It passes through the
root hash from the `PulseShape` to the `MeasurementResult`. This is correct because
the HAL server does not have access to GRAPE config or Python-side software versions.

### 9.5 Diff Implementation (`diff.rs`)

The Rust diff mirrors the Python implementation:

```rust
use super::tree::{NodeType, ProvenanceNode};

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct NodeChange {
    pub node_type: String,
    pub path: String,
    pub old_hash: String,
    pub new_hash: String,
    pub description: String,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ProvenanceDiff {
    pub hash_a: String,
    pub hash_b: String,
    pub changed_nodes: Vec<NodeChange>,
    pub unchanged_nodes: Vec<String>,
}

impl ProvenanceDiff {
    pub fn compute(
        tree_a: &ProvenanceNode,
        tree_b: &ProvenanceNode,
    ) -> Self {
        if tree_a.hash == tree_b.hash {
            let unchanged = collect_leaf_types(tree_a);
            return Self {
                hash_a: tree_a.hash.clone(),
                hash_b: tree_b.hash.clone(),
                changed_nodes: vec![],
                unchanged_nodes: unchanged,
            };
        }

        let mut changed = vec![];
        let mut unchanged = vec![];
        diff_recursive(
            tree_a, tree_b, "",
            &mut changed, &mut unchanged,
        );

        Self {
            hash_a: tree_a.hash.clone(),
            hash_b: tree_b.hash.clone(),
            changed_nodes: changed,
            unchanged_nodes: unchanged,
        }
    }

    pub fn is_identical(&self) -> bool {
        self.changed_nodes.is_empty()
    }
}
```

The recursive diff walks both trees, matching children by `(node_type, label)`.
For leaf nodes with differing hashes, it generates a `NodeChange` with a
human-readable description of field-level differences.

---

## 10. Python Implementation

### 10.1 Module Structure

```
qubit-os-core/src/qubitos/
├── provenance/
│   ├── __init__.py     // Public API re-exports
│   ├── nodes.py        // NodeType, ProvenanceNode (Section 7.1)
│   ├── tree.py         // ProvenanceTree (Section 7.2)
│   ├── diff.py         // ProvenanceDiff, NodeChange (Section 7.3)
│   ├── store.py        // ProvenanceStore (Section 7.4)
│   └── builder.py      // ProvenanceBuilder (Section 7.5)
├── pulsegen/
│   ├── grape.py        // (modified) GrapeResult gets provenance_hash
│   └── ...
├── calibrator/
│   ├── fingerprint.py  // (unchanged) CalibrationFingerprint as-is
│   └── ...
└── cli/
    └── main.py         // (modified) new experiment commands
```

### 10.2 `__init__.py` Public API

```python
# src/qubitos/provenance/__init__.py

from .builder import ProvenanceBuilder, canonicalize_float, hash_envelope
from .diff import NodeChange, ProvenanceDiff
from .nodes import NodeType, ProvenanceNode
from .store import ProvenanceStore
from .tree import ProvenanceTree

__all__ = [
    "NodeType",
    "ProvenanceNode",
    "ProvenanceTree",
    "ProvenanceDiff",
    "NodeChange",
    "ProvenanceStore",
    "ProvenanceBuilder",
    "canonicalize_float",
    "hash_envelope",
]
```

### 10.3 Integration with GRAPE

`GrapeResult` is extended with an optional provenance hash:

```python
# In src/qubitos/pulsegen/grape.py

@dataclass
class GrapeResult:
    """Result of GRAPE optimization."""
    i_envelope: NDArray[np.float64]
    q_envelope: NDArray[np.float64]
    fidelity: float
    iterations: int
    converged: bool
    fidelity_history: list[float] = field(default_factory=list)
    final_unitary: NDArray[np.complex128] | None = None
    provenance_hash: str = ""  # NEW: root hash of provenance tree
```

The `generate_pulse()` function is extended to optionally build provenance:

```python
def generate_pulse(
    gate: str | GateType,
    num_qubits: int = 1,
    duration_ns: float = 20.0,
    target_fidelity: float = 0.999,
    qubit_indices: list[int] | None = None,
    config: GrapeConfig | None = None,
    calibration: CalibrationFingerprint | None = None,  # NEW
    build_provenance: bool = True,                       # NEW
) -> GrapeResult:
    """Generate an optimized pulse for a quantum gate.

    If build_provenance=True and calibration is provided, a provenance
    tree is constructed and its root hash attached to the result.
    """
    # ... existing optimization code ...

    if build_provenance and calibration is not None:
        from ..provenance import ProvenanceBuilder

        builder = ProvenanceBuilder()
        builder.set_calibration_from_fingerprint(calibration)
        builder.set_grape_config(config)
        builder.add_pulse(
            pulse_id="",  # assigned later when stored
            gate_type=gate.value if isinstance(gate, GateType) else gate,
            target_qubit_indices=qubit_indices or [0],
            duration_ns=int(config.duration_ns),
            num_time_steps=config.num_time_steps,
            target_fidelity=config.target_fidelity,
            max_amplitude_mhz=config.max_amplitude,
            i_envelope=result.i_envelope,
            q_envelope=result.q_envelope,
        )
        builder.set_software_versions()
        tree = builder.build()
        result.provenance_hash = tree.root_hash

    return result
```

### 10.4 CLI Commands

New `experiment` command group added to `cli/main.py`:

```python
@cli.group()
def experiment() -> None:
    """Experiment provenance commands."""
    pass


@experiment.command("provenance")
@click.argument("hash_prefix")
@click.option(
    "--store", "-s", default=None, type=click.Path(),
    help="Path to provenance store file",
)
@click.option(
    "--format", "-f", "output_format", default="text",
    type=click.Choice(["text", "json", "yaml"]),
    help="Output format",
)
def experiment_provenance(
    hash_prefix: str,
    store: str | None,
    output_format: str,
) -> None:
    """Show the provenance tree for an experiment.

    HASH_PREFIX is the full or partial root hash.

    Examples:

        qubit-os experiment provenance a1b2c3d4
        qubit-os experiment provenance a1b2c3d4 --format json
    """
    try:
        from ..provenance import ProvenanceStore

        store_path = Path(store) if store else _default_store_path()
        pstore = ProvenanceStore(persist_path=store_path)
        tree = pstore.get(hash_prefix)

        if tree is None:
            click.echo(
                f"No provenance tree found for: {hash_prefix}",
                err=True,
            )
            sys.exit(1)

        if output_format == "text":
            click.echo(tree.summary())
        else:
            _output(tree.to_dict(), output_format)

    except Exception as e:
        click.echo(f"Error ({type(e).__name__}): {e}", err=True)
        sys.exit(1)


@experiment.command("diff")
@click.argument("hash_a")
@click.argument("hash_b")
@click.option(
    "--store", "-s", default=None, type=click.Path(),
    help="Path to provenance store file",
)
@click.option(
    "--format", "-f", "output_format", default="text",
    type=click.Choice(["text", "json", "yaml"]),
    help="Output format",
)
def experiment_diff(
    hash_a: str,
    hash_b: str,
    store: str | None,
    output_format: str,
) -> None:
    """Diff two experiment provenance trees.

    Shows exactly what changed between two experiments.

    Examples:

        qubit-os experiment diff a1b2c3d4 e5f6a7b8
    """
    try:
        from ..provenance import ProvenanceStore

        store_path = Path(store) if store else _default_store_path()
        pstore = ProvenanceStore(persist_path=store_path)
        diff_result = pstore.diff(hash_a, hash_b)

        if diff_result is None:
            click.echo(
                "Could not find one or both trees. "
                "Check hash prefixes.",
                err=True,
            )
            sys.exit(1)

        if output_format == "text":
            click.echo(diff_result.summary())
        else:
            _output(diff_result.to_dict(), output_format)

    except Exception as e:
        click.echo(f"Error ({type(e).__name__}): {e}", err=True)
        sys.exit(1)


def _default_store_path() -> Path:
    """Default location for the provenance store."""
    return Path.home() / ".qubitos" / "provenance.json"
```

### 10.5 Populating PulseShape Proto Fields

The existing PulseShape proto fields (`calibration_fingerprint`, `code_version`,
`created_at`, `random_seed`) should be populated when constructing PulseShape messages.
A utility function handles the conversion:

```python
def grape_result_to_pulse_shape(
    result: GrapeResult,
    config: GrapeConfig,
    gate_type: str,
    target_qubits: list[int],
    calibration: CalibrationFingerprint | None = None,
    provenance_tree: ProvenanceTree | None = None,
) -> PulseShape:
    """Convert a GrapeResult to a PulseShape proto message.

    Populates all provenance fields that were previously left empty.
    """
    import time
    import qubitos

    pulse = PulseShape()
    pulse.algorithm = "grape"
    pulse.gate_type = _gate_type_to_proto(gate_type)
    pulse.target_qubit_indices.extend(target_qubits)
    pulse.target_fidelity = result.fidelity
    pulse.duration_ns = int(config.duration_ns)
    pulse.num_time_steps = config.num_time_steps
    pulse.time_step_ns = config.duration_ns / config.num_time_steps
    pulse.i_envelope.extend(result.i_envelope.tolist())
    pulse.q_envelope.extend(result.q_envelope.tolist())
    pulse.max_amplitude_mhz = config.max_amplitude

    # Provenance fields (previously unpopulated)
    pulse.proto_version = 1
    pulse.created_at.seconds = int(time.time())
    pulse.created_at.nanos = 0
    pulse.code_version = qubitos.__version__
    if config.random_seed is not None:
        pulse.random_seed = config.random_seed
    if calibration is not None:
        pulse.calibration_fingerprint = f"sha256:{calibration.hash}"

    # New provenance tree field
    if provenance_tree is not None:
        _populate_proto_provenance(pulse.provenance, provenance_tree)

    return pulse
```

---

## 11. Integration Points

### 11.1 CalibrationFingerprint -> CalibrationNode

The existing `CalibrationFingerprint` is reused directly.
`ProvenanceBuilder.set_calibration_from_fingerprint()` accepts a
`CalibrationFingerprint` and decomposes it into `QubitCalibrationNode` and
`CouplerCalibrationNode` children. The legacy 16-char fingerprint hash is preserved
in tree metadata for backward-compatible lookups.

No changes to `CalibrationFingerprint` or `FingerprintStore` are required.

### 11.2 FingerprintStore and ProvenanceStore

These stores serve different purposes and coexist:

| Aspect          | FingerprintStore                   | ProvenanceStore                  |
|-----------------|------------------------------------|----------------------------------|
| **Indexes**     | By backend name (temporal history) | By root hash (content-addressed) |
| **Tracks**      | Calibration drift over time        | Complete experiment context       |
| **Query**       | "What was the latest calibration?" | "What context produced result X?" |
| **Stores**      | CalibrationFingerprint objects     | ProvenanceTree objects           |

They are complementary. A provenance tree *contains* a calibration subtree, but
FingerprintStore provides temporal drift analysis that ProvenanceStore does not.

### 11.3 PulseSequence (TIME-MODEL-SPEC)

When the time model and `PulseSequence` type are implemented, the provenance tree
auto-builds from the sequence:

```python
class PulseSequence:
    def build_provenance(
        self,
        calibration: CalibrationFingerprint,
        config: GrapeConfig,
    ) -> ProvenanceTree:
        builder = ProvenanceBuilder()
        builder.set_calibration_from_fingerprint(calibration)
        builder.set_grape_config(config)
        for scheduled_pulse in self.pulses:
            builder.add_pulse(
                pulse_id=scheduled_pulse.pulse_id,
                gate_type=scheduled_pulse.gate_type,
                target_qubit_indices=scheduled_pulse.target_qubits,
                duration_ns=scheduled_pulse.duration_ns,
                num_time_steps=scheduled_pulse.num_time_steps,
                target_fidelity=scheduled_pulse.target_fidelity,
                max_amplitude_mhz=scheduled_pulse.max_amplitude_mhz,
                i_envelope=scheduled_pulse.i_envelope,
                q_envelope=scheduled_pulse.q_envelope,
            )
        builder.set_software_versions()
        return builder.build()
```

### 11.4 GRAPE Optimizer

`GrapeOptimizer.optimize()` is not modified. Provenance is built *after* optimization
in `generate_pulse()` or at the caller's discretion. This keeps the optimizer focused
on physics and math, with provenance as an opt-in wrapping layer.

### 11.5 HAL Server

The Rust HAL server participates in provenance by:

1. Receiving `PulseShape` with an attached `ProvenanceTree` from the Python client
2. Extracting the `root_hash` from the tree
3. Attaching `root_hash` to `MeasurementResult.provenance_hash` before returning
4. Optionally storing the tree in a Rust-side `ProvenanceStore` for server-side queries

The HAL server does not construct or recompute provenance trees. It is a pass-through
for the hash. This is intentional: the HAL server does not have access to GRAPE config
or Python software versions.

### 11.6 Error Budget (ERROR-BUDGET-SPEC)

When error budgets are implemented, the budget snapshot is included in metadata:

```python
builder.set_metadata(
    "error_budget_total_infidelity",
    str(budget.total_infidelity),
)
builder.set_metadata(
    "error_budget_decoherence_cost",
    str(budget.decoherence_cost),
)
builder.set_metadata(
    "error_budget_remaining",
    str(budget.remaining_budget),
)
```

Metadata is not hashed because the error budget is a *derived* quantity, not an input.

### 11.7 CLI Summary

| Command                                       | Description                        |
|-----------------------------------------------|------------------------------------|
| `qubit-os experiment provenance <hash>`       | Display provenance tree for a hash |
| `qubit-os experiment diff <hash_a> <hash_b>`  | Diff two experiment provenances    |

These follow the existing CLI patterns: Click groups, `--format` option supporting
text/json/yaml, error messages to stderr, `sys.exit(1)` on failure.

---

## 12. Migration Path

### 12.1 Backward Compatibility

All provenance additions are **optional**:

- `MeasurementResult.provenance_hash` defaults to empty string
- `PulseShape.provenance` defaults to unset (proto3 default)
- `OptimizeResponse.provenance_hash` defaults to empty string
- `GrapeResult.provenance_hash` defaults to empty string

Existing code that does not set provenance fields continues to work unchanged.
Results without provenance are still valid and processable.

### 12.2 Incremental Adoption

**Phase 1 (v0.2.0):** Python provenance module ships.
- `ProvenanceBuilder`, `ProvenanceTree`, `ProvenanceStore` available as library
- `generate_pulse()` accepts optional `calibration` parameter to build provenance
- CLI `experiment provenance` and `experiment diff` commands available
- Proto definitions for provenance messages deployed but not yet populated by HAL

**Phase 2 (v0.2.x):** Wire up proto fields.
- Populate `PulseShape` proto fields that were previously dead code
  (`calibration_fingerprint`, `code_version`, `created_at`, `random_seed`)
- HAL server passes through `provenance_hash` on `MeasurementResult`
- Rust `provenance` module implemented with hash computation and diff

**Phase 3 (v0.3.0):** Default-on.
- Provenance built automatically in `generate_pulse()` when calibration is available
- ProvenanceStore persistence enabled by default (~/.qubitos/provenance.json)
- CLI workflows include provenance in output by default

### 12.3 CalibrationFingerprint Stability

The existing `CalibrationFingerprint` hash algorithm (JSON -> SHA-256 -> 16 chars) is
**not changed**. The provenance tree computes its own calibration subtree hash from
individual qubit/coupler nodes, which will differ from the legacy hash. Both are
stored: the legacy hash in metadata, the tree hash in the node. This avoids breaking
any code that depends on the existing fingerprint format.

---

## 13. Test Plan

### 13.1 Hash Determinism

```
test_hash_determinism_same_inputs:
    Given: Two ProvenanceBuilder instances with identical inputs
    When: Both build provenance trees
    Then: root_hash values are byte-for-byte identical

test_hash_determinism_across_runs:
    Given: A provenance tree serialized to JSON
    When: Deserialized and re-hashed
    Then: Root hash matches the stored root hash

test_hash_determinism_float_canonicalization:
    Given: Two floats differing only in least-significant bits
           (e.g., 5.123456789012345 vs 5.123456789012346)
    When: Both canonicalized to 12 significant digits
    Then: Both produce the same canonical value
    And: Both produce the same hash

test_hash_determinism_dict_key_order:
    Given: Two dicts with identical keys/values but different insertion order
    When: Hashed via hash_leaf
    Then: Both produce the same hash (sort_keys=True guarantees this)
```

### 13.2 Hash Sensitivity

```
test_sensitivity_single_float_change:
    Given: Two trees identical except qubit_0 T1 differs by 0.001 us
    When: Both built and hashed
    Then: Root hashes differ
    And: CalibrationNode hashes differ
    And: QubitCalibrationNode[0] hashes differ
    And: QubitCalibrationNode[1] hash is unchanged

test_sensitivity_grape_config_change:
    Given: Two trees identical except learning_rate differs
    When: Both built and hashed
    Then: Root hashes differ
    And: CalibrationNode hash is unchanged
    And: GRAPEConfigNode hashes differ

test_sensitivity_software_version_change:
    Given: Two trees identical except numpy_version differs
    When: Both built and hashed
    Then: Root hashes differ
    And: SoftwareVersionNode hashes differ
    And: CalibrationNode and PulseSequenceNode hashes unchanged

test_sensitivity_envelope_single_sample:
    Given: Two I-envelopes differing by one sample (1e-10 amplitude change)
    When: Hashed via hash_envelope
    Then: Hashes differ (raw byte comparison, no rounding)

test_sensitivity_node_type_prefix:
    Given: A QubitCalibrationNode and CouplerCalibrationNode with
           numerically identical content fields
    When: Both hashed
    Then: Hashes differ (due to __node_type__ prefix in canonical JSON)
```

### 13.3 Tree Construction

```
test_tree_structure_single_qubit_single_pulse:
    Given: 1 qubit, 0 couplers, 1 pulse, GRAPE config, software versions
    When: Tree built via ProvenanceBuilder
    Then: Root has 2 children (Calibration, Experiment)
    And: Calibration has 1 child (QubitCalibration)
    And: Experiment has 2 children (PulseSequence, Config)
    And: PulseSequence has 1 child (ScheduledPulse)
    And: Config has 2 children (GRAPEConfig, SoftwareVersion)

test_tree_structure_two_qubit_with_coupler:
    Given: 2 qubits, 1 coupler, 1 CZ pulse
    When: Tree built
    Then: Calibration has 3 children (2 QubitCal + 1 CouplerCal)

test_tree_structure_multi_pulse_sequence:
    Given: 3 pulses in sequence (X on q0, X on q1, CZ on q0-q1)
    When: Tree built
    Then: PulseSequence has 3 ScheduledPulse children

test_tree_builder_missing_calibration:
    Given: ProvenanceBuilder without set_calibration_from_fingerprint()
    When: build() called
    Then: Raises ValueError with descriptive message

test_tree_builder_no_pulses:
    Given: Builder with calibration but no add_pulse() calls
    When: build() called
    Then: Succeeds with empty PulseSequenceNode (content: {"empty": True})

test_tree_find_node:
    Given: A fully-built provenance tree
    When: find_node(NodeType.QUBIT_CALIBRATION, "qubit_0") called
    Then: Returns the correct QubitCalibrationNode

test_tree_all_leaves:
    Given: Tree with 2 qubits, 1 coupler, 1 pulse, GRAPE config, software
    When: all_leaves() called
    Then: Returns 6 leaf nodes (2 qubit + 1 coupler + 1 pulse + 1 grape + 1 sw)
```

### 13.4 Diff

```
test_diff_identical_trees:
    Given: Two trees built from identical inputs
    When: diff() called
    Then: is_identical == True
    And: changed_nodes is empty
    And: unchanged_nodes lists all leaf types

test_diff_calibration_change:
    Given: Tree A and Tree B differing only in qubit_0 T1
    When: diff() called
    Then: changed_nodes has exactly 1 entry
    And: entry.node_type == "qubit_calibration"
    And: entry.path contains "qubit_0"
    And: entry.description contains "t1_us"
    And: unchanged_nodes includes grape_config, software_version, etc.

test_diff_multiple_changes:
    Given: Trees differing in T1 on qubit_0 AND learning_rate
    When: diff() called
    Then: changed_nodes has exactly 2 entries
    And: one is qubit_calibration, other is grape_config

test_diff_added_qubit:
    Given: Tree A with 2 qubits, Tree B with 3 qubits
    When: diff() called
    Then: changed_nodes includes an "added" entry for qubit_2

test_diff_removed_coupler:
    Given: Tree A with 1 coupler, Tree B with 0 couplers
    When: diff() called
    Then: changed_nodes includes a "removed" entry for the coupler

test_diff_human_readable_description:
    Given: Trees differing in qubit_0 frequency: 5.0 GHz -> 5.001 GHz
    When: diff() called
    Then: description contains "frequency_ghz: 5 -> 5.001" (or similar)

test_diff_summary_format:
    Given: A ProvenanceDiff with 2 changes and 4 unchanged
    When: summary() called
    Then: Output contains "2 change(s):" and "4 unchanged node(s):"
    And: Each change shows node_type, path, old_hash, new_hash, description
```

### 13.5 Serialization Round-Trip

```
test_roundtrip_to_dict_from_dict:
    Given: A fully constructed provenance tree
    When: Serialized via to_dict() then deserialized via from_dict()
    Then: root_hash of reconstructed tree equals original
    And: All node hashes match
    And: All leaf content matches
    And: metadata matches
    And: timestamp matches

test_roundtrip_json_file:
    Given: A provenance tree
    When: to_dict() -> json.dumps() -> json.loads() -> from_dict()
    Then: Reconstructed tree root_hash is identical

test_roundtrip_preserves_labels:
    Given: A tree with various labeled nodes
    When: Round-tripped through to_dict/from_dict
    Then: All labels are preserved

test_diff_roundtrip:
    Given: A ProvenanceDiff
    When: Serialized via to_dict()
    Then: Contains hash_a, hash_b, num_changes, changed_nodes, unchanged_nodes
    And: Each changed_node has node_type, path, old_hash, new_hash, description
```

### 13.6 Integration Tests

```
test_integration_grape_to_provenance:
    Given: A CalibrationFingerprint and GrapeConfig
    When: generate_pulse() called with build_provenance=True
    Then: GrapeResult.provenance_hash is non-empty (64 hex chars)
    And: Hash is deterministic (same config + calibration = same hash)

test_integration_grape_to_pulseshape_proto:
    Given: A GrapeResult with provenance_hash
    When: Converted to PulseShape proto via grape_result_to_pulse_shape()
    Then: PulseShape.calibration_fingerprint is populated ("sha256:...")
    And: PulseShape.code_version is populated (matches qubitos.__version__)
    And: PulseShape.created_at is populated (non-zero timestamp)
    And: PulseShape.random_seed is populated (if config.random_seed was set)
    And: PulseShape.provenance.root_hash matches GrapeResult.provenance_hash

test_integration_execute_carries_provenance:
    Given: A PulseShape with provenance tree
    When: Executed through HAL (simulator backend)
    Then: MeasurementResult.provenance_hash matches PulseShape root hash
    And: MeasurementResult.calibration_fingerprint is still populated (backward compat)

test_integration_store_and_retrieve:
    Given: A provenance tree stored in ProvenanceStore
    When: Retrieved by full hash
    Then: Retrieved tree root_hash matches stored tree root_hash
    When: Retrieved by 8-char prefix
    Then: Same tree returned
    When: Diff with itself
    Then: is_identical == True

test_integration_cross_language_hash_leaf:
    Given: Identical leaf content in Python and Rust
    When: hash_leaf() called in both languages
    Then: Hashes match (validates JSON serialization consistency)

test_integration_cross_language_hash_envelope:
    Given: Identical float64 array in Python (numpy) and Rust (Vec<f64>)
    When: hash_envelope() called in both languages
    Then: Hashes match (validates byte order consistency)

test_integration_cli_provenance:
    Given: A provenance tree stored in the default store
    When: `qubit-os experiment provenance <hash>` executed
    Then: Exit code 0
    And: Output shows tree summary with correct structure

test_integration_cli_diff:
    Given: Two provenance trees stored in the default store
    When: `qubit-os experiment diff <hash_a> <hash_b>` executed
    Then: Exit code 0
    And: Output shows diff with changed and unchanged nodes
```

### 13.7 Floating-Point Edge Cases

```
test_canonicalize_float_zero:
    assert canonicalize_float(0.0) == 0.0
    assert canonicalize_float(-0.0) == 0.0

test_canonicalize_float_nan:
    assert canonicalize_float(float('nan')) == 0.0

test_canonicalize_float_inf:
    assert canonicalize_float(float('inf')) == 0.0
    assert canonicalize_float(float('-inf')) == 0.0

test_canonicalize_float_very_small:
    result = canonicalize_float(1.23456789012e-15)
    assert result == 1.23456789012e-15

test_canonicalize_float_very_large:
    result = canonicalize_float(1.23456789012e+15)
    assert result == 1.23456789012e+15

test_canonicalize_float_precision_boundary:
    a = 5.123456789012345
    b = 5.123456789012999
    assert canonicalize_float(a) == canonicalize_float(b)

test_canonicalize_float_different_values:
    a = 5.12345678901
    b = 5.12345678902
    assert canonicalize_float(a) != canonicalize_float(b)
```

### 13.8 Envelope Hashing Performance

```
test_envelope_hash_performance:
    Given: Envelope with 10,000 float64 samples (80 KB)
    When: hash_envelope() called 100 times
    Then: Total time < 100ms (< 1ms per hash)
    Note: SHA-256 throughput is ~500 MB/s on modern CPUs; 80KB takes ~0.16ms

test_envelope_hash_empty:
    Given: Empty numpy array (np.array([], dtype=np.float64))
    When: hash_envelope() called
    Then: Returns valid 64-character hex hash

test_envelope_hash_dtype_conversion:
    Given: Array as float32
    When: hash_envelope() called (converts to float64 internally)
    Then: Returns valid hash
    Note: float32->float64 conversion changes values, so hash differs
          from a native float64 array with "same" values. This is correct.
```

### 13.9 Store Tests

```
test_store_basic:
    Given: An empty ProvenanceStore
    When: A tree is stored
    Then: get(root_hash) returns the tree
    And: len(store) == 1

test_store_prefix_lookup:
    Given: A tree with hash "abcdef1234567890..."
    When: get("abcdef") called
    Then: Returns the tree

test_store_prefix_ambiguous:
    Given: Two trees whose hashes share a 4-char prefix
    When: get(4_char_prefix) called
    Then: Returns the first match (insertion order)

test_store_eviction:
    Given: ProvenanceStore(max_entries=3)
    When: 4 trees stored
    Then: len(store) == 3
    And: Oldest tree is no longer retrievable
    And: Three newest trees are retrievable

test_store_persistence_write:
    Given: ProvenanceStore with persist_path to a temp file
    When: Tree stored
    Then: File exists on disk
    And: File contains valid JSON with "trees" key

test_store_persistence_read:
    Given: ProvenanceStore saved to disk with 2 trees
    When: New ProvenanceStore created with same persist_path
    Then: len(store) == 2
    And: Both trees retrievable by hash

test_store_diff:
    Given: Two trees stored
    When: store.diff(hash_a, hash_b) called
    Then: Returns correct ProvenanceDiff

test_store_diff_missing_tree:
    Given: One tree stored
    When: store.diff(stored_hash, "nonexistent") called
    Then: Returns None

test_store_contains:
    Given: A tree stored with hash H
    When: contains(H) called
    Then: Returns True
    When: contains("nonexistent") called
    Then: Returns False

test_store_history:
    Given: 5 trees stored in order
    When: history(limit=3) called
    Then: Returns last 3 trees in insertion order
```

---

## 14. Future Extensions

This spec establishes the foundation for several follow-on capabilities:

### 14.1 Provenance-Based Result Caching

If two experiments have identical provenance root hashes, they will produce
statistically identical results (modulo shot noise). A cache layer could:

- Before execution: check if a result with the same provenance hash exists
- If yes: return cached result (or merge shot counts for more statistics)
- If no: execute and cache the result

This requires careful treatment of shot noise: cached results from N shots are
not statistically equivalent to fresh results from N shots.

### 14.2 Experiment Comparison Dashboards

With provenance-tagged results stored over time, a dashboard could:

- Plot fidelity vs. time, colored by which provenance subtree changed
- Show "this fidelity drop correlates with T1 drift on qubit 3"
- Compare software version upgrades: "v0.2.0 achieves 0.3% higher fidelity on
  CZ gates vs. v0.1.0 with identical calibration"

### 14.3 Automated Regression Detection

Define baseline provenance hashes for known-good experiments. On each CI run:

- Re-run baseline experiments
- Compare fidelity against historical results with the same provenance
- If fidelity degrades without provenance changes: hardware drift detected
- If fidelity degrades with provenance changes: investigate the specific change

### 14.4 Hardware Firmware Version Tracking

When backends expose firmware version APIs:

- Add `firmware_version` to `SoftwareVersionNode`
- Track firmware updates as provenance changes
- Correlate firmware updates with performance changes

### 14.5 Environmental Condition Logging

With temperature sensors, EM monitors, or vibration sensors:

- Log readings as provenance metadata (not hashed, since not controlled inputs)
- Correlate environmental conditions with result quality
- Eventually: include environmental hash in a separate "environment" subtree
  once sensors provide reliable, repeatable readings

---

## 15. References

1. Merkle, R. C. (1988). "A Digital Signature Based on a Conventional Encryption
   Function." *Advances in Cryptology -- CRYPTO '87*, Lecture Notes in Computer
   Science, vol 293. Springer. doi:10.1007/3-540-48184-2_32

2. Khaneja, N., Reiss, T., Kehlet, C., Schulte-Herbruggen, T., & Glaser, S. J.
   (2005). "Optimal control of coupled spin dynamics: design of NMR pulse sequences
   by gradient ascent algorithms." *Journal of Magnetic Resonance*, 172(2), 296-305.
   doi:10.1016/j.jmr.2004.11.004

3. Nielsen, M. A., & Chuang, I. L. (2010). *Quantum Computation and Quantum
   Information* (10th Anniversary Edition). Cambridge University Press.
   Chapter 7: Quantum computers: physical realization.

4. DVC (Data Version Control). https://dvc.org -- The "git for data" pattern applied
   to ML pipelines. Similar content-addressing and diffing principles.

5. Git. Torvalds, L. (2005). Content-addressable storage using SHA-1 Merkle trees
   for source code version control. https://git-scm.com

6. NIST SP 800-185. SHA-3 Derived Functions. SHA-256 security properties and
   collision resistance guarantees.

---

*This document should be reviewed alongside ARCHITECTURE-REVIEW.md (GAP 4),
QubitOS-Design-v0.5.0.md, and the future TIME-MODEL-SPEC.txt and ERROR-BUDGET-SPEC.txt.
Implementation should begin with the Python module (Phase 1) as it provides immediate
value for experiment tracking without requiring proto or Rust changes.*
