segurai 0.1.0__tar.gz

segurai-0.1.0/LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        https://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       https://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

segurai-0.1.0/PKG-INFO

@@ -0,0 +1,405 @@
+Metadata-Version: 2.4
+Name: segurai
+Version: 0.1.0
+Summary: Provenance and authority for AI agent chains - who authorized what your agents did
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://github.com/Yaxomiller/segurai
+Project-URL: Repository, https://github.com/Yaxomiller/segurai
+Project-URL: Changelog, https://github.com/Yaxomiller/segurai/blob/master/CHANGELOG.md
+Project-URL: Documentation, https://github.com/Yaxomiller/segurai/blob/master/docs/ARCHITECTURE.md
+Project-URL: Issues, https://github.com/Yaxomiller/segurai/issues
+Keywords: ai-agents,authorization,provenance,audit,opentelemetry,langgraph,mcp
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3 :: Only
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Security
+Classifier: Topic :: System :: Monitoring
+Classifier: Typing :: Typed
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: cryptography>=42.0
+Provides-Extra: otel
+Requires-Dist: opentelemetry-api>=1.30; extra == "otel"
+Requires-Dist: opentelemetry-sdk>=1.30; extra == "otel"
+Provides-Extra: langgraph
+Requires-Dist: langgraph>=1.0; extra == "langgraph"
+Provides-Extra: mcp
+Requires-Dist: mcp>=2.0; extra == "mcp"
+Provides-Extra: collector
+Requires-Dist: opentelemetry-proto>=1.30; extra == "collector"
+Provides-Extra: intent
+Requires-Dist: anthropic>=1.5; extra == "intent"
+Provides-Extra: judge
+Requires-Dist: anthropic>=1.5; extra == "judge"
+Provides-Extra: all
+Requires-Dist: segurai[collector,intent,judge,langgraph,mcp,otel]; extra == "all"
+Provides-Extra: dev
+Requires-Dist: pytest>=8.0; extra == "dev"
+Requires-Dist: ruff>=0.6; extra == "dev"
+Requires-Dist: opentelemetry-exporter-otlp-proto-http>=1.30; extra == "dev"
+Requires-Dist: segurai[all]; extra == "dev"
+Dynamic: license-file
+
+# Segurai
+
+[![CI](https://github.com/Yaxomiller/segurai/actions/workflows/ci.yml/badge.svg)](https://github.com/Yaxomiller/segurai/actions/workflows/ci.yml)
+[![Python 3.10+](https://img.shields.io/badge/python-3.10%2B-blue)](https://www.python.org/downloads/)
+[![License: Apache 2.0](https://img.shields.io/badge/license-Apache%202.0-blue)](LICENSE)
+
+**Provenance and authority for AI agent chains.** Records what a human asked for, records what the agents actually did, and links the two — so "who authorized this?" has an answer.
+
+> A manager types *"clean up our old sales records."* An assistant passes the job to a CRM agent, which deletes 400 customer records. A month later an auditor asks who authorized it. Every log says: *the service account did it.*
+>
+> The trail from the action back to the human is broken. That is what this fixes.
+
+Install it and import it as `segurai`. The signed grant that each agent hop carries is called a **warrant** — a narrow, written-down, traceable permission, like a search warrant — and that word is used throughout the code and the docs.
+
+New here? [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) explains how everything fits together in plain language.
+
+---
+
+## Status
+
+**M0 — chain capture: complete.**
+
+- Signed, attenuation-only delegation (Ed25519), with keys that persist so old chains can be re-verified
+- Tier 1 admission checks on every action, in about a microsecond
+- Observe mode (record and flag) and enforce mode (stop before it happens)
+- Integrations: **LangGraph** (nodes, subgraphs, ToolNode), **MCP** (client and server), **OpenTelemetry** (GenAI semantic conventions)
+- SQLite storage, chain reconstruction, CLI
+- A collector that receives spans from many processes and re-checks every signature
+
+**M1 — intent compiler: built, first live eval pending.**
+
+- One Claude call turns a request into a scope contract (structured output, cached instructions, refusal fallbacks)
+- Contracts are checked against the organisation's catalog and clamped to the person's real permissions
+- Conditions that the catalog's evidence can test are written as Checks automatically
+- A read-only web view: the request and contract beside the chain, findings inline, signatures re-verified on every load
+- 50 labeled eval cases and a grading harness; running them needs an Anthropic API key
+
+**M2 — Tier 1 check suite: complete.**
+
+- Every rule an action breaks is reported: operation, target, per-action scale, session total, irreversibility, and conditions
+- Conditions become testable Checks (e.g. `days_since_activity >= 180`), evaluated against evidence the agent reports; missing evidence is "unverifiable", never a false alarm
+- 20 seeded scenarios: 10/10 violations caught for the right reason, 0/10 false positives, 0.15 ms per checked action (p99)
+
+**M3 — public release: in progress.**
+
+- Zero-integration path: existing OpenTelemetry GenAI spans become an observed chain with no code changes
+- CI on Linux, Windows and macOS across Python 3.10–3.13, plus packaging checks
+- Public repository, CI green on Linux, macOS and Windows across Python 3.10–3.13
+- PyPI release wired up and waiting on one PyPI setting ([docs/RELEASING.md](docs/RELEASING.md)); distribution is the remaining work
+
+**M4 — Tier 2 semantic judge: built, first live eval pending.**
+
+- Reviews only the actions Tier 1 flagged, one narrow question each: within intent, outside intent, or needs a human
+- Verdicts sit beside Tier 1's record — advice, never authority — and agent-written text is treated as untrusted data
+- 50 labeled cases (27 ambiguous, 3 with planted prompt injection); approving a bad action counts as a dangerous miss
+
+**M6 — scope-creep report: built.**
+
+- Per agent and per person, over any window: operations never used, targets never touched, wildcards that could shrink, size limits against the largest real change
+- A proposed tightened scope built only from observed use, with the evidence behind it
+- Grants from warrants or from an identity-provider export; Markdown, HTML, JSON, and a CSV of every unused permission
+
+**M5 — design-partner pilots: kit ready.**
+
+- A hardened container for the collector and web view, and a Compose file that won't start without a token
+- A pilot runbook: day-0 setup, the weekly review, the day-30 wrap-up, data handling ([docs/PILOT_RUNBOOK.md](docs/PILOT_RUNBOOK.md))
+- The pilots themselves need partners — that part is people, not code
+
+## Install
+
+```bash
+pip install segurai                 # core: needs only `cryptography`
+pip install "segurai[langgraph]"    # + LangGraph integration
+pip install "segurai[mcp]"          # + MCP integration
+pip install "segurai[otel]"         # + OpenTelemetry spans
+pip install "segurai[intent]"       # + intent compiler (Anthropic SDK)
+pip install "segurai[judge]"        # + Tier 2 judge (Anthropic SDK)
+pip install "segurai[collector]"    # + collector protobuf support
+```
+
+Working on Segurai itself:
+
+```bash
+git clone https://github.com/Yaxomiller/segurai && cd segurai
+pip install -e ".[dev]"             # everything, plus pytest and ruff
+python -m pytest -q
+```
+
+## Quick start
+
+```python
+import segurai
+from segurai import Scope
+
+segurai.configure(store="segurai.db")      # signing key kept beside it as segurai.key
+
+granted = Scope.of(
+    operations=["read", "update", "delete"],
+    targets=["crm.*"],
+    max_scale=50,
+    constraints=["only records with no recent activity"],
+)
+
+# An agent says only what it narrows; everything else is inherited.
+@segurai.agent("crm_agent", scope=Scope.of(targets=["crm.opportunities"]))
+def crm_agent():
+    segurai.act("read", "crm.opportunities", scale=412)
+    segurai.act("delete", "crm.opportunities", scale=400, irreversible=True)
+
+@segurai.agent("assistant")
+def assistant():
+    crm_agent()
+
+with segurai.session(
+    principal="sarah@company.com",
+    request="clean up our old sales records",
+    scope=granted,
+) as s:
+    assistant()
+
+print(segurai.render(s.session_id))
+```
+
+```
+Delegation chain
+  └─ entrypoint  [wrt_7a07522323244bb1]
+        ops={delete,read,update} targets={crm.*} max_scale=50
+     └─ assistant  [wrt_3f22fd11d57f4d91]
+           ops={delete,read,update} targets={crm.*} max_scale=50
+        └─ crm_agent  [wrt_ba50b4faee524582]
+              ops={delete,read,update} targets={crm.opportunities} max_scale=50
+              [ok     ] read crm.opportunities scale=412
+              [FAIL   ] delete crm.opportunities scale=400
+                        -> scale 400 exceeds limit 50
+                        -> irreversible action not permitted by scope
+
+Summary: 3 warrants, 2 actions, 1 violation(s)
+  ! crm_agent: delete crm.opportunities (scale=400) - scale 400 exceeds limit 50; irreversible action not permitted by scope [went ahead]
+```
+
+`[FAIL   ]` means the action was outside the warrant and went ahead (observe mode). With `segurai.configure(mode="enforce")` it is stopped before it happens and shows as `[BLOCKED]`.
+
+## Guarding tools
+
+Describe a tool once with an `ActionSpec` — what operation it performs, on which resource, at what scale — and every call is checked *before* the tool body runs:
+
+```python
+from segurai import ActionSpec, count
+
+@segurai.tool(ActionSpec("delete", "crm.{table}", scale=count("ids"), irreversible=True))
+def delete_records(table: str, ids: list[int]) -> int:
+    ...
+```
+
+## Conditions
+
+A plain-English condition can't be checked by a program; a `Check` can. The agent reports evidence about the records it touches, and every item must satisfy the condition:
+
+```python
+from segurai import Check
+
+granted = Scope.of(
+    operations=["read", "archive"], targets=["crm.*"], max_scale=50, max_total=50,
+    checks=[Check(field="days_since_activity", op=">=", value=180,
+                  condition="only records with no recent activity")],
+)
+
+segurai.act("archive", "crm.opportunities", scale=3, days_since_activity=[200, 12, 400])
+# -> condition not met: only records with no recent activity (days_since_activity = 12, required >= 180)
+```
+
+`max_total` caps the records changed across the whole session, so 400 one-record deletions can't slip under a per-action limit of 50. An action that reports no evidence for a check is marked unverifiable (`?` in the tree) rather than treated as a violation. Run the 20-scenario suite with `segurai eval-tier1`.
+
+## LangGraph
+
+```python
+from segurai.instrument.langgraph import ToolGuard, agent_node
+
+guard = ToolGuard({
+    "delete_records": ActionSpec("delete", "crm.{table}", scale=count("ids"), irreversible=True),
+})
+
+crm = StateGraph(MessagesState)
+crm.add_node("model", call_model)
+crm.add_node("tools", ToolNode(tools, wrap_tool_call=guard.wrap, awrap_tool_call=guard.awrap))
+...
+
+app = StateGraph(MessagesState)
+app.add_node("crm_agent", agent_node("crm_agent", crm.compile(),
+                                     scope=Scope.of(targets=["crm.opportunities"])))
+
+with segurai.session(principal="sarah@company.com", request=request, scope=granted):
+    app.invoke({"messages": [HumanMessage(request)]})
+```
+
+Works for plain, async and parallel nodes, and for compiled subgraphs. In enforce mode a blocked tool call comes back to the model as an error message saying why, so the agent can change course instead of the graph crashing. Full example: [examples/langgraph_demo.py](examples/langgraph_demo.py).
+
+## MCP
+
+```python
+from segurai.instrument.mcp import WarrantedClient, bind_from_meta
+
+# Agent side: calls are checked before they are sent, and carry the warrant.
+crm = WarrantedClient(mcp_client, server="crm", rules={...})
+await crm.call_tool("delete_records", {"table": "opportunities", "ids": ids})
+
+# Server side: verify the incoming warrant and continue the same chain.
+@server.tool()
+async def delete_records(table: str, ids: list[int], ctx: Context) -> str:
+    with bind_from_meta(ctx, holder="crm_server"):
+        segurai.act("delete", f"crm.{table}", scale=len(ids), irreversible=True)
+        ...
+```
+
+Tools without a rule are described from the server's own `readOnlyHint` / `destructiveHint` annotations. Full example: [examples/mcp_demo.py](examples/mcp_demo.py).
+
+## OpenTelemetry
+
+```python
+from segurai.instrument import otel
+otel.enable()
+```
+
+Sessions, agent hops and tool calls become spans named per the OpenTelemetry GenAI conventions (`invoke_workflow`, `invoke_agent`, `execute_tool`), carrying the full signed warrant as `warrant.*` attributes. The chain shows up in whatever tracing backend a team already uses. Full example: [examples/otel_demo.py](examples/otel_demo.py).
+
+## Scopes from requests
+
+Instead of writing the scope by hand, let the intent compiler read the request. One Claude call produces a structured contract, which is checked against your catalog of real operations and resources and clamped to the person's actual permissions — so a request can never grant more than the person already has.
+
+```python
+from segurai.intent import Catalog, IntentCompiler
+
+segurai.configure(store="segurai.db", intent=IntentCompiler(Catalog.load("catalog.json")))
+
+with segurai.session(principal="sarah@company.com",
+                     request="clean up our old sales records",
+                     ceiling=sarahs_permissions) as s:
+    ...
+```
+
+Needs `pip install -e ".[intent]"` and `ANTHROPIC_API_KEY`. Try one request with `segurai intent "clean up our old sales records" --catalog evals/intent/catalog.json`; measure the compiler on 50 labeled cases with `segurai eval-intent` (see [evals/intent/](evals/intent/README.md)).
+
+## Collector
+
+Agents on many machines send spans to one place with the standard OpenTelemetry OTLP exporter:
+
+```bash
+segurai collect --db central.db --key org.pub --port 4318
+```
+
+The collector rebuilds every chain and re-checks every signature with the organisation's public key, refusing anything forged.
+
+**Already using OpenTelemetry? Try it with no code changes.** Point your existing exporter at the collector. Standard GenAI spans (`invoke_agent`, `execute_tool`) from any framework are rebuilt into an *observed* chain: which agent called which tool, and for whom. Observed chains are clearly labeled — nothing in them was authorized, signed or checked — so they show where warrants would go before you adopt them.
+
+## Scope-creep report
+
+Which permissions did we give agents that they never used?
+
+```bash
+segurai report --days 30                                   # Markdown to the terminal
+segurai report --days 30 --format csv --out unused.csv     # one row per unused permission
+segurai report --grants okta-export.json --format html --out report.html
+```
+
+For every agent, and for each person's sessions as a whole, it lists operations never used, targets never touched, wildcards that were only used for a few resources, and size limits against the largest change actually made. It then proposes a tightened scope built only from observed use (it never adds anything), with the evidence behind it. Pass `--grants` with standing permissions exported from an identity provider ([format](examples/grants.example.json)) to compare those too, including against observed chains from plain OpenTelemetry. Rarely-run work may need permissions a short window doesn't show, so every report says so and rates its confidence by how much activity it saw.
+
+## Semantic judge (Tier 2)
+
+Tier 1 is exact but can only compare. For the actions it flags, the judge asks Claude one narrow question: given what this person asked for, is this action within intent, outside it, or does a person need to decide?
+
+```bash
+segurai judge <session_id> --yes      # judge the session's flagged actions
+```
+
+Verdicts are stored beside Tier 1's findings and shown in `segurai show` and the web view; they never change Tier 1's record. Everything the judge reads from agents and tools is treated as untrusted data, so "this was pre-approved" written by an agent grants nothing. Measure it with `segurai eval-judge` on 50 labeled cases (see [evals/tier2/](evals/tier2/README.md)).
+
+## Web view
+
+```bash
+segurai ui --db segurai.db        # then open http://127.0.0.1:8700/
+```
+
+One page per session: the request and its contract beside the delegation chain, every action's findings inline, and the signatures re-checked each time the page loads. It is plain HTML with no scripts; everything from a request is escaped, and pages carry a strict Content-Security-Policy. With `SEGURAI_COLLECTOR_TOKEN` set, the browser asks for the token as a password. The collector serves the same pages at `/`.
+
+## Deploying a pilot
+
+```bash
+export SEGURAI_COLLECTOR_TOKEN="$(python -c 'import secrets; print(secrets.token_urlsafe(32))')"
+docker compose -f deploy/compose.yaml up -d --build
+```
+
+The container runs as an unprivileged user on a read-only filesystem, keeps everything in one volume, listens only on localhost (put your TLS proxy in front), and refuses traffic from other machines without a token. Settings come from `SEGURAI_DB`, `SEGURAI_KEY`, `SEGURAI_HOST`, `SEGURAI_PORT` and `SEGURAI_COLLECTOR_TOKEN`. The [pilot runbook](docs/PILOT_RUNBOOK.md) covers the rest: keys, connecting agents, the weekly review and the final report.
+
+## CLI
+
+```bash
+segurai sessions              # list recent sessions
+segurai show <session_id>     # render the delegation tree (--json, --ascii)
+segurai verify <session_id>   # re-check every signature, link and narrowing
+segurai creep <session_id>    # granted vs exercised scope
+segurai report --days 30      # scope-creep report across sessions
+segurai keygen                # create a signing key
+segurai collect               # receive spans from agents (OTLP/HTTP)
+segurai ui                    # browse sessions in a web page
+segurai backup --out copy.db  # consistent copy of the database, safe while running
+segurai intent "<request>"    # compile a request into a scope contract
+segurai eval-intent ...       # grade the intent compiler on labeled cases
+segurai eval-tier1            # run the Tier 1 scenario suite (no API calls)
+segurai judge <session_id>    # Tier 2: judge a session's flagged actions
+segurai eval-judge            # grade the judge on labeled cases
+```
+
+`--db` picks the database (default `segurai.db`); the key is found beside it.
+
+## Design
+
+**Warrants attenuate, never widen.** Each hop derives a child warrant from its parent. A child may narrow operations, targets, scale, or reversibility — never broaden them. Attempting to widen raises rather than silently clamping, because a caller trying to widen has a bug worth seeing. The signature covers the parent link and depth, so a chain cannot be re-parented or reordered after the fact.
+
+If that invariant held at every hop, then whatever happened at the leaf was inside what the root principal granted. That is the whole claim.
+
+**Patterns are compared as patterns.** A child asking for `crm.*` under a parent limited to `crm.?` must be refused, even though plain glob matching would read the child's `*` as an ordinary character and say yes. Containment is checked token by token and errs toward refusing: it may reject a scope that was technically safe, but never accepts one that isn't. Matching is case-sensitive on every platform.
+
+**Unspecified fields inherit.** A declared child scope says what it *narrows*; anything it doesn't mention comes from the parent. Without this, narrowing one field would mean retyping every other — which trains people to paste a wide scope everywhere just to silence the error, producing exactly the over-permissioning this is meant to expose. Leaving a field out (`targets=None`) inherits it; an explicit empty list (`targets=[]`) means *none*.
+
+**Scale bounds mutations, not reads.** An agent scanning 412 rows to find deletion candidates is doing its job; deleting 400 of them is the concern. Conflating them makes every table scan a violation and teaches people to ignore alerts.
+
+**Unverifiable is not a violation.** When an action doesn't report the evidence a condition needs, the finding says so plainly. Treating it as a violation would bury real alerts in noise; treating it as a pass would be untrue.
+
+**An action nobody can describe is not one anybody authorized.** A tool call whose arguments don't fit its `ActionSpec`, or a tool nobody wrote a spec for, is recorded as a violation rather than waved through.
+
+**Observe by default.** Violations are recorded, nothing is blocked. Nobody lets an unproven library block production traffic — but they will let it watch, and watching produces the evidence that earns the right to enforce later.
+
+**Observability never breaks the agent.** A failing listener or exporter is logged and ignored.
+
+## What's next
+
+| Milestone | |
+|---|---|
+| M1 | Intent compiler — built; first live eval run pending |
+| M2 | Tier 1 check suite — complete: 10/10 caught, 0 false positives |
+| M3 | Public release; ingest existing OpenTelemetry spans with no code changes |
+| M4 | Tier 2 semantic judge — built; first live eval pending |
+| M5 | Design-partner pilots — deployment kit and runbook ready |
+| M6 | Scope-creep report — built |
+
+Known limits are listed honestly at the end of [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md#honest-limits-what-is-not-done-yet).
+
+## Tests
+
+```bash
+python -m pytest -q
+```
+
+## License
+
+Apache-2.0