quant-swarm 0.3.0__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (81) hide show
  1. quant_swarm-0.3.0/LICENSE +31 -0
  2. quant_swarm-0.3.0/PKG-INFO +313 -0
  3. quant_swarm-0.3.0/README.md +285 -0
  4. quant_swarm-0.3.0/pyproject.toml +50 -0
  5. quant_swarm-0.3.0/quant_swarm.egg-info/PKG-INFO +313 -0
  6. quant_swarm-0.3.0/quant_swarm.egg-info/SOURCES.txt +79 -0
  7. quant_swarm-0.3.0/quant_swarm.egg-info/dependency_links.txt +1 -0
  8. quant_swarm-0.3.0/quant_swarm.egg-info/entry_points.txt +5 -0
  9. quant_swarm-0.3.0/quant_swarm.egg-info/requires.txt +12 -0
  10. quant_swarm-0.3.0/quant_swarm.egg-info/top_level.txt +1 -0
  11. quant_swarm-0.3.0/setup.cfg +4 -0
  12. quant_swarm-0.3.0/swarm_mcp/__init__.py +14 -0
  13. quant_swarm-0.3.0/swarm_mcp/access.py +199 -0
  14. quant_swarm-0.3.0/swarm_mcp/cache/__init__.py +1 -0
  15. quant_swarm-0.3.0/swarm_mcp/cache/bars.py +193 -0
  16. quant_swarm-0.3.0/swarm_mcp/cache/db.py +232 -0
  17. quant_swarm-0.3.0/swarm_mcp/cache/enrich.py +80 -0
  18. quant_swarm-0.3.0/swarm_mcp/cache/freshness.py +47 -0
  19. quant_swarm-0.3.0/swarm_mcp/cache/rate_limit.py +68 -0
  20. quant_swarm-0.3.0/swarm_mcp/envelope.py +152 -0
  21. quant_swarm-0.3.0/swarm_mcp/mcp_compat.py +17 -0
  22. quant_swarm-0.3.0/swarm_mcp/plans.py +84 -0
  23. quant_swarm-0.3.0/swarm_mcp/redaction.py +48 -0
  24. quant_swarm-0.3.0/swarm_mcp/relay.py +206 -0
  25. quant_swarm-0.3.0/swarm_mcp/request_context.py +18 -0
  26. quant_swarm-0.3.0/swarm_mcp/server_meta.py +35 -0
  27. quant_swarm-0.3.0/swarm_mcp/servers/__init__.py +1 -0
  28. quant_swarm-0.3.0/swarm_mcp/servers/data_server.py +424 -0
  29. quant_swarm-0.3.0/swarm_mcp/servers/gym_server.py +251 -0
  30. quant_swarm-0.3.0/swarm_mcp/servers/http_server.py +362 -0
  31. quant_swarm-0.3.0/swarm_mcp/servers/warden_server.py +302 -0
  32. quant_swarm-0.3.0/swarm_mcp/telemetry.py +34 -0
  33. quant_swarm-0.3.0/swarm_mcp/tool_runner.py +43 -0
  34. quant_swarm-0.3.0/swarm_mcp/tools/__init__.py +1 -0
  35. quant_swarm-0.3.0/swarm_mcp/tools/data_tools.py +686 -0
  36. quant_swarm-0.3.0/swarm_mcp/tools/gym_tools.py +346 -0
  37. quant_swarm-0.3.0/swarm_mcp/tools/warden_tools.py +310 -0
  38. quant_swarm-0.3.0/swarm_mcp/vendor_path.py +28 -0
  39. quant_swarm-0.3.0/swarm_mcp/vendored/__init__.py +0 -0
  40. quant_swarm-0.3.0/swarm_mcp/vendored/alpha/__init__.py +0 -0
  41. quant_swarm-0.3.0/swarm_mcp/vendored/alpha/bars_fetch.py +56 -0
  42. quant_swarm-0.3.0/swarm_mcp/vendored/alpha/guardrails_path.py +32 -0
  43. quant_swarm-0.3.0/swarm_mcp/vendored/alpha/gym/__init__.py +13 -0
  44. quant_swarm-0.3.0/swarm_mcp/vendored/alpha/gym/panel.py +136 -0
  45. quant_swarm-0.3.0/swarm_mcp/vendored/alpha/gym/policy_head.py +60 -0
  46. quant_swarm-0.3.0/swarm_mcp/vendored/alpha/gym/regime.py +107 -0
  47. quant_swarm-0.3.0/swarm_mcp/vendored/alpha/gym/simulator.py +242 -0
  48. quant_swarm-0.3.0/swarm_mcp/vendored/alpha/microstructure.py +142 -0
  49. quant_swarm-0.3.0/swarm_mcp/vendored/alpha/order_checks.py +81 -0
  50. quant_swarm-0.3.0/swarm_mcp/vendored/alpha/shared/__init__.py +1 -0
  51. quant_swarm-0.3.0/swarm_mcp/vendored/alpha/shared/swing_screens.py +126 -0
  52. quant_swarm-0.3.0/swarm_mcp/vendored/guardrails/LICENSE.md +16 -0
  53. quant_swarm-0.3.0/swarm_mcp/vendored/guardrails/__init__.py +0 -0
  54. quant_swarm-0.3.0/swarm_mcp/vendored/guardrails/budget.py +100 -0
  55. quant_swarm-0.3.0/swarm_mcp/vendored/guardrails/fill_model.py +29 -0
  56. quant_swarm-0.3.0/swarm_mcp/vendored/guardrails/gates.py +127 -0
  57. quant_swarm-0.3.0/swarm_mcp/vendored/guardrails/genome_schema.json +83 -0
  58. quant_swarm-0.3.0/swarm_mcp/vendored/guardrails/genome_schema.py +114 -0
  59. quant_swarm-0.3.0/swarm_mcp/vendored/guardrails/objective.py +73 -0
  60. quant_swarm-0.3.0/swarm_mcp/vendored/guardrails/provenance.py +92 -0
  61. quant_swarm-0.3.0/swarm_mcp/vendored/guardrails/tiers.py +142 -0
  62. quant_swarm-0.3.0/tests/test_access_gate.py +110 -0
  63. quant_swarm-0.3.0/tests/test_audit_features.py +78 -0
  64. quant_swarm-0.3.0/tests/test_build_features.py +70 -0
  65. quant_swarm-0.3.0/tests/test_cache_determinism.py +69 -0
  66. quant_swarm-0.3.0/tests/test_derived_tools.py +172 -0
  67. quant_swarm-0.3.0/tests/test_entitlements.py +194 -0
  68. quant_swarm-0.3.0/tests/test_floors_parity.py +73 -0
  69. quant_swarm-0.3.0/tests/test_gym_tools.py +106 -0
  70. quant_swarm-0.3.0/tests/test_ip_boundary.py +56 -0
  71. quant_swarm-0.3.0/tests/test_no_lookahead_port.py +95 -0
  72. quant_swarm-0.3.0/tests/test_no_order_paths.py +43 -0
  73. quant_swarm-0.3.0/tests/test_offline_mode.py +38 -0
  74. quant_swarm-0.3.0/tests/test_order_checks_parity.py +90 -0
  75. quant_swarm-0.3.0/tests/test_pins.py +15 -0
  76. quant_swarm-0.3.0/tests/test_redaction.py +57 -0
  77. quant_swarm-0.3.0/tests/test_registry_metadata.py +74 -0
  78. quant_swarm-0.3.0/tests/test_relay.py +280 -0
  79. quant_swarm-0.3.0/tests/test_remote_http.py +331 -0
  80. quant_swarm-0.3.0/tests/test_sizing_parity.py +67 -0
  81. quant_swarm-0.3.0/tests/test_tier_refusal_port.py +78 -0
@@ -0,0 +1,31 @@
1
+ MIT License
2
+
3
+ Copyright (c) 2026 The 1.21 Initiative
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
22
+
23
+ ---
24
+
25
+ NOTICE: `swarm_mcp/vendored/guardrails/` contains a CHECKER SUBSET of the
26
+ privately developed trading-swarm-guardrails rules layer, vendored at a pinned
27
+ commit SHA. Those files are NOT covered by the MIT license above; they are
28
+ distributed under the source-available terms in
29
+ `swarm_mcp/vendored/guardrails/LICENSE.md`. The selection machinery
30
+ (objective scoring, deflated-Sharpe estimation, promotion-gate decisions) is
31
+ deliberately excluded from this repository and remains server-side.
@@ -0,0 +1,313 @@
1
+ Metadata-Version: 2.4
2
+ Name: quant-swarm
3
+ Version: 0.3.0
4
+ Summary: Local stdio MCP servers extracted from the trading swarm: point-in-time data assembly, pre-trade invariant + leakage auditing, and regime-fragility probing.
5
+ Author: The 1.21 Initiative
6
+ License: MIT
7
+ Project-URL: Homepage, https://github.com/blink1217/trading-swarm-mcp
8
+ Keywords: mcp,trading,backtest-audit,point-in-time,regimes
9
+ Classifier: Development Status :: 4 - Beta
10
+ Classifier: Intended Audience :: Financial and Insurance Industry
11
+ Classifier: License :: OSI Approved :: MIT License
12
+ Classifier: Programming Language :: Python :: 3.11
13
+ Classifier: Topic :: Office/Business :: Financial :: Investment
14
+ Requires-Python: >=3.11
15
+ Description-Content-Type: text/markdown
16
+ License-File: LICENSE
17
+ Requires-Dist: mcp>=2.0
18
+ Requires-Dist: httpx>=0.27
19
+ Requires-Dist: numpy>=1.26
20
+ Requires-Dist: pandas>=2.0
21
+ Requires-Dist: scipy>=1.11
22
+ Provides-Extra: test
23
+ Requires-Dist: pytest>=8; extra == "test"
24
+ Provides-Extra: remote
25
+ Requires-Dist: uvicorn>=0.30; extra == "remote"
26
+ Requires-Dist: starlette>=0.37; extra == "remote"
27
+ Dynamic: license-file
28
+
29
+ # quant-swarm
30
+
31
+ Risk-first pre-trade checks that gate live capital — point-in-time data, invariant and leakage audits, regime probes.
32
+
33
+ <!-- mcp-name: io.github.blink1217/quant-swarm -->
34
+
35
+ ## Built by a Principal Quantitative Technologist
36
+
37
+ Authored by the Principal Quantitative Technologist at [The 1.21 Initiative](https://1.21initiative.com) — the same code path that gates live promotion and execution decisions in the trading swarm. Every tool exists because production trading systems fail in two repeatable ways: **lookahead leakage in research** and **an oversized position in production**. `quant-swarm` ships the checkers that prevent both, so you can audit your own work with the exact logic that guards deployed capital.
38
+
39
+ ## Why risk-first
40
+
41
+ Risk gates are not post-hoc reports — they are the pre-condition for every action. The package is organized around that constraint:
42
+
43
+ - **Fail-closed by default.** Unknown features map to tier C, missing provenance returns `UNSCORABLE` (never neutral-filled), and unverifiable tokens are refused on the hosted path. See `warden.audit_features` and `features.build` provenance guards.
44
+ - **Pinned, non-drifting invariants.** Order sizing floors (per-name 25%, gross 60%), fill model (spread + slippage + adverse selection, both legs, always against you), and hard-constraint checks are vendored at pinned commit SHAs (`.github/pins.json`, verified byte-for-byte by CI). A parity test pins the three copies of the floors (Python/MCP, C# live engine, Terraform) so they cannot drift.
45
+ - **Point-in-time provenance.** SQLite WAL cache at `%LOCALAPPDATA%\1.21-initiative\swarm-mcp\cache.db` (`$XDG_CACHE_HOME/1.21-initiative/swarm-mcp` on POSIX): finalized sessions are never re-fetched, in-progress sessions refresh at most every 60 s, earnings/news are append-only on `fetched_at`, and `features.build` runs the no-lookahead guard on every row (each feature must equal a fresh causal recomputation at `as_of`).
46
+ - **Pessimistic execution.** `warden.cost_check` and the simulator model the fill that hurts you; there is no optimistic fill path to inflate backtests.
47
+ - **No order-placement code path.** No tool can place, cancel, or route an order — this repository cannot trade, only refuse unsafe trades.
48
+ - **Statistically honest refusals.** Promotion verdicts are **never** issued locally. Statistically undecidable outputs return `INDETERMINATE_LOCAL`/`UNDERPOWERED`/`UNSCORABLE`, name the exact missing inputs (`MIN_EPISODES=20`, PBO, DSR, worst-regime margins), and hand off to the hosted tournament / Strategy Validation Audit.
49
+
50
+ - Local-first, BYO-key: your Alpaca/Finnhub keys are read from env only when operating in BYOK mode; otherwise data flows through the hosted relay.
51
+ - Your symbols, genomes, and alpha never leave your machine except as an `audit_request` (genome hash + violation summary, no proprietary payload) when you explicitly escalate.
52
+
53
+ **One-click install on [Smithery](https://smithery.ai/servers/blink-kt/quant-swarm):**
54
+ the listing is [`blink-kt/quant-swarm`](https://smithery.ai/servers/blink-kt/quant-swarm) (package and Smithery slug are both `quant-swarm`), and it carries all three servers. Configuration is optional — without a token the servers still start and list every tool; the first call points you at the free signup. Add a token any time to unlock the relay-backed tools.
55
+
56
+ ---
57
+
58
+ ## The servers
59
+
60
+ | Server | Hook | Tools |
61
+ | --- | --- | --- |
62
+ | `swarm-data-mcp` | Derived-only market signals plus point-in-time feature building | `market.pulse`, `market.sentiment`, `features.build`, `cache.warm`, `cache.stats`, `cache.offline` |
63
+ | `swarm-warden-mcp` | Pre-trade invariant + leakage gate — the live-capital floors and the point-in-time guards | `warden.validate_order`, `warden.audit_features`, `warden.cost_check`, `warden.validate_genome`, `warden.explain_sizing`, `warden.promotion_verdict` |
64
+ | `swarm-gym-mcp` | Regime fragility: *which regime kills it, and is your sample even large enough to have an opinion?* | `gym.label_regimes`, `gym.probe_fragility`, `gym.paired_preview`, `gym.estimate_cloud_run` |
65
+
66
+ Raw bar/enrichment access (`get_bars` / `enrich_symbol`) is internal only and no longer
67
+ registered as tools — those paths echo raw provider values, and the data policy is
68
+ derived-only: ratios, percentile ranks, labels, buckets and counts. `market.pulse` and
69
+ `market.sentiment` are the general-purpose replacements; both accept an optional `bars`
70
+ argument so you can supply your own OHLCV rows for symbols we don'"'"'t carry.
71
+
72
+ ## Plans
73
+
74
+ Access is tokened through [https://1.21initiative.com/](https://1.21initiative.com/).
75
+
76
+ - **Free** (instant token at signup): all warden checkers (`warden.validate_order`, `warden.cost_check`,
77
+ `warden.explain_sizing`, `warden.validate_genome`, `warden.audit_features`), the derived snapshots
78
+ `market.pulse` / `market.sentiment` / `market.regime` (shared-relay symbols capped at
79
+ 10 per call, ~250 relay data calls/month), plus `cache.stats` and `cache.offline`.
80
+ - **Pro** (paid by one-time credit packs, self-serve at the site): everything else — `features.build`,
81
+ `cache.warm` bulk backfill, all gym tools, `warden.promotion_verdict`, and the Pro
82
+ derived tools (`market.microstructure`, `volume.forecast`, `market.screen`,
83
+ `market.rank`), paid by one-time credit packs (10k or 100k relay calls, 90-day validity — no subscription, no unlimited plan) with 50 symbols/call.
84
+ - **Institutional** (custom-quoted): Strategy Validation Audit engagements, custom
85
+ universes, hosted evolution runs, SLA.
86
+
87
+ Plan limits are enforced server-side (the relay, `/api/mcp/verify`, and the hosted
88
+ streamable-HTTP endpoint return structured `402` refusals with an `upgrade_url`). This
89
+ open-source client'"'"'s plan check is **advisory**: a free token calling a Pro tool gets an
90
+ `UPGRADE_REQUIRED` envelope — the tool still lists, and the attempt points at the upgrade
91
+ page. No DRM; the free plan is deliberately the zero-cost surface.
92
+
93
+ ### Point-in-time data, for real
94
+
95
+ LLM-driven research re-runs cells constantly; each re-run burns Finnhub (60 req/min) and
96
+ Alpaca (200 req/min) budgets re-pulling identical history — and then silently builds
97
+ features from data that did not exist at decision time. `swarm-data-mcp` fixes both:
98
+
99
+ - SQLite (WAL) cache at `%LOCALAPPDATA%\1.21-initiative\swarm-mcp\cache.db`
100
+ (`$XDG_CACHE_HOME/1.21-initiative/swarm-mcp` on POSIX).
101
+ - **Finalized sessions are never re-fetched** — zero-cost replays forever. The
102
+ in-progress session refreshes at most every 60 s; enrichment every 300 s.
103
+ - Earnings/news are append-only on `fetched_at`: a later fetch can never rewrite an
104
+ earlier `as_of`.
105
+ - Per-provider token buckets + `429` exponential backoff (`1s\u00b72\u207f`).
106
+ - `features.build` runs the no-lookahead guard on every row (each feature must equal a
107
+ fresh causal recomputation at `as_of`) and the provenance guards on every field.
108
+ Tier-B/C fields without recorded point-in-time evidence come back `UNSCORABLE` —
109
+ never neutral-filled with `0.0`.
110
+
111
+ Every response carries `coverage` (cache vs API, oldest/newest session per symbol) and
112
+ `limits` (your local depth in weeks vs the tape-depth gates: 8 weeks for the tier-A
113
+ fast path, 26 weeks for tape eligibility). Below the gates you get an `escalation`
114
+ block naming the hosted `bars_1day` panel that satisfies them. Tape-tier replay is
115
+ **roadmap, not available** — we will not imply otherwise.
116
+
117
+ ### The warden
118
+
119
+ `warden.validate_order` is the same function that rejects live orders: per-name 25% of equity,
120
+ gross cap 60%. Per-fund overrides are allowed, but the response always reports your
121
+ deviation from the house floors. `warden.audit_features` flags banned actuals sources
122
+ (`open-meteo.archive` and friends return *actuals*, not what the forecast said on a past
123
+ Friday), features predating tape start, and unknown features fail closed to tier C.
124
+ `warden.cost_check` converts a claimed gross edge into net-of-cost reality under the pessimistic
125
+ fill model (spread + slippage + adverse selection, both legs, always against you).
126
+ `warden.explain_sizing` is a step-by-step mirror of the C# live risk engine, and a parity test
127
+ pins the floors against terraform and guardrails so the three copies cannot drift.
128
+
129
+ ### The gym
130
+
131
+ `gym.probe_fragility` replays your genome over the deterministic tier-A simulator and reports
132
+ per-regime net bps, the worst regime, turnover, and hard-constraint violations — never a
133
+ promotion. Seeds are capped at 8 and `per_regime` at 2: **statistical honesty, not
134
+ artificial scarcity**. Tier-B/C mutations against a champion raise the gym'"'"'s
135
+ `TierScoringRefusal` (UNSCORABLE) instead of silently neutral-filling features the price
136
+ panel cannot provide. `gym.paired_preview` compares champion vs challenger on identical
137
+ market paths with the promotion gate bypassed and every statistic labelled
138
+ **UNDERPOWERED**, naming the exact seed count needed to clear `MIN_EPISODES=20`.
139
+
140
+ ---
141
+
142
+ ## Install
143
+
144
+ Requires Python \u2265 3.11. One kind of credential:
145
+
146
+ - **Access token (all three servers, required).** Every tool call is gated on a token
147
+ issued by [https://1.21initiative.com/](https://1.21initiative.com/) — request access
148
+ there (that'"'"'s also the Strategy Validation Audit booking flow), then set
149
+ `SWARM_MCP_ACCESS_TOKEN` in the server'"'"'s `env`. Clients still list the tools without a
150
+ token; every call returns an `ACCESS_REQUIRED` envelope pointing back to the site.
151
+ The token is verified against the site, and that verification is the usage meter —
152
+ only the token itself is ever sent, never symbols, genomes, prices, or provider keys.
153
+
154
+ The same token also feeds the **hosted data relay**: `swarm-data-mcp` serves bars and
155
+ enrichment through `https://1.21initiative.com/api/mcp/...`, so **no Alpaca or Finnhub
156
+ credentials are required** — the site holds the provider keys behind the relay and caches
157
+ historical bars in GCS. The relay is fail-closed: a rejected or unverifiable token means a
158
+ refused data fetch, never partial rows. The point-in-time cache semantics on the client
159
+ side are unchanged — finalized sessions are immutable in local SQLite regardless of which
160
+ data path filled them.
161
+
162
+ ### Smithery
163
+
164
+ Install [`blink-kt/quant-swarm`](https://smithery.ai/servers/blink-kt/quant-swarm) from
165
+ Smithery — one listing, one token, all three servers. Smithery forwards `apiToken` as the
166
+ `SWARM_MCP_ACCESS_TOKEN` env var (stdio) or as `?apiToken=` on the hosted endpoints; the
167
+ token is optional at connect time.
168
+
169
+ ### Cursor
170
+
171
+ `Settings \u2192 MCP \u2192 Add server` (or `.cursor/mcp.json` — a checked-in example lives at
172
+ [.cursor/mcp.json](.cursor/mcp.json)):
173
+
174
+ ```json
175
+ {
176
+ "mcpServers": {
177
+ "swarm-data": {
178
+ "command": "uvx",
179
+ "args": ["--from", "quant-swarm", "swarm-data-mcp"],
180
+ "env": {
181
+ "SWARM_MCP_ACCESS_TOKEN": "<token from https://1.21initiative.com/>"
182
+ }
183
+ },
184
+ "swarm-warden": {
185
+ "command": "uvx",
186
+ "args": ["--from", "quant-swarm", "swarm-warden-mcp"],
187
+ "env": { "SWARM_MCP_ACCESS_TOKEN": "<token from https://1.21initiative.com/>" }
188
+ },
189
+ "swarm-gym": {
190
+ "command": "uvx",
191
+ "args": ["--from", "quant-swarm", "swarm-gym-mcp"],
192
+ "env": { "SWARM_MCP_ACCESS_TOKEN": "<token from https://1.21initiative.com/>" }
193
+ }
194
+ }
195
+ }
196
+ ```
197
+
198
+ Prefer a deeplink? `scripts/make_deeplinks.py` prints `cursor://\u2026/mcp/install` links for
199
+ all three servers (base64 of the stdio config, so they cannot drift from the JSON above);
200
+ the generated list is checked in at [.cursor/DEEPLINKS.md](.cursor/DEEPLINKS.md).
201
+
202
+ ### Claude Desktop
203
+
204
+ `claude_desktop_config.json` \u2014 identical `mcpServers` block as above.
205
+
206
+ ### Claude Code
207
+
208
+ ```bash
209
+ claude mcp add swarm-data --env SWARM_MCP_ACCESS_TOKEN=<token> -- uvx --from quant-swarm swarm-data-mcp
210
+ claude mcp add swarm-warden --env SWARM_MCP_ACCESS_TOKEN=<token> -- uvx --from quant-swarm swarm-warden-mcp
211
+ claude mcp add swarm-gym --env SWARM_MCP_ACCESS_TOKEN=<token> -- uvx --from quant-swarm swarm-gym-mcp
212
+ ```
213
+
214
+ ### Windsurf
215
+
216
+ `~/.codeium/windsurf/mcp_config.json` \u2014 the same stdio JSON shape as the Cursor block
217
+ above. For the hosted endpoints use the `serverUrl` form instead:
218
+
219
+ ```json
220
+ {
221
+ "mcpServers": {
222
+ "swarm-data-remote": {
223
+ "serverUrl": "https://swarm-mcp-503318750546.europe-west1.run.app/mcp/data"
224
+ }
225
+ }
226
+ }
227
+ ```
228
+
229
+ (The hosted endpoints accept the token as `Authorization: Bearer <token>` or as
230
+ `?apiToken=<token>`.)
231
+
232
+ The warden and gym are pure checkers and the data server fetches through the hosted
233
+ relay, so **no server needs provider keys** \u2014 only the access token.
234
+
235
+ **Local development:** operators of this repo can bootstrap offline by setting
236
+ `SWARM_MCP_LOCAL_TOKEN` to the same value as `SWARM_MCP_ACCESS_TOKEN` (documented
237
+ bypass; token verification against the site is skipped and the token gets the full
238
+ Pro entitlement \u2014 it is a local override on your own machine).
239
+
240
+ **Swarm operators (internal):** to fetch directly from the providers instead of the
241
+ relay, set `SWARM_MCP_BYOK=1` plus `ALPACA_API_KEY`/`ALPACA_SECRET`/`FINNHUB_API_KEY`
242
+ on `swarm-data-mcp`. Public users should leave `SWARM_MCP_BYOK` unset.
243
+
244
+ **Hosting the remote endpoint yourself:** `swarm-mcp-http` validates the `Host` header
245
+ (DNS-rebinding protection). The defaults allow `1.21initiative.com`, the project'"'"'s Cloud
246
+ Run hostname, the Smithery gateway, and localhost; `SWARM_MCP_ALLOWED_HOSTS`
247
+ (comma-separated) **extends** the defaults \u2014 it never replaces them, so a custom domain
248
+ cannot accidentally lock out the Cloud Run URL. `SWARM_MCP_REMOTE_URL` overrides the
249
+ advertised endpoint base, and `SWARM_MCP_TOKEN_VERIFY_URL` overrides the verify endpoint.
250
+ See the Dockerfile for the container form.
251
+
252
+ ---
253
+
254
+ ## The IP boundary (what ships, what doesn'"'"'t)
255
+
256
+ This repository ships **checkers only**: order checks, provenance guards, the
257
+ pessimistic fill model, the genome schema, provenance tiers, hard-constraint checking,
258
+ gate thresholds, the tier-A gym simulator, and the regime labeller \u2014 vendored at pinned
259
+ commit SHAs (`.github/pins.json`, verified by CI).
260
+
261
+ The **selection machinery stays server-side** and is not in this repo: objective scoring,
262
+ deflated-Sharpe estimation, and the promotion-gate decision. Any request for a promotion
263
+ verdict returns `INDETERMINATE_LOCAL` with the exact missing statistical inputs
264
+ (`MIN_EPISODES=20`, PBO \u2264 0.30 via combinatorially-symmetric splits, DSR margin with the
265
+ monotonic trials ledger, worst-regime margin across all 5 regimes \u00d7 seeds), plus:
266
+
267
+ - an `audit_request` block \u2014 genome hash + violation summary, **no proprietary payload**;
268
+ - a `cloud_job` spec accepted verbatim by the hosted `POST /tournament/run`.
269
+
270
+ In v1 the handoff is manual via the **Strategy Validation Audit** booking flow at
271
+ [https://1.21initiative.com/](https://1.21initiative.com/) \u2014 no public ingress, no new attack
272
+ surface. Access requests go through the site, which is how contact details are captured and
273
+ how hosted API keys are issued and metered. The human conversation is the product.
274
+
275
+ ## Telemetry
276
+
277
+ Off by default, opt-in only (`SWARM_MCP_TELEMETRY_OPT_IN=opt-in`), counters only
278
+ (tool name, success flag, coarse duration). Never symbols, genomes, prices, or
279
+ credentials. Institutional buyers read the source; silent phone-home destroys the wedge.
280
+
281
+ The one exception is by contract: **access-token verification**. When a token is set and
282
+ no local bootstrap token matches, the token is POSTed to the verify endpoint (default
283
+ `https://1.21initiative.com/api/mcp/verify`, override `SWARM_MCP_TOKEN_VERIFY_URL`).
284
+ That call is how usage is metered. Only the token is sent; the gate fails closed if the
285
+ endpoint is unreachable or rejects the token.
286
+
287
+ ## Development
288
+
289
+ ```powershell
290
+ pip install -e ".[test]"
291
+ py -3.11 -m pytest # 119 tests
292
+ scripts\vendor.ps1 # re-vendor the pinned checker subset
293
+ py -3.11 scripts\check_pin.py # vendored tree == pins.json
294
+ ```
295
+
296
+ Vendoring uses the sibling checkouts of the private repos at the pinned SHAs
297
+ (`-FromWorktree` bootstraps before the pin commit exists). CI mirrors the swarm'"'"'s own
298
+ `guardrails-invariants` gate: it checks out the pinned SHAs and verifies the committed
299
+ vendored tree byte-for-byte (and the stripped subsets transform-for-transform), then runs
300
+ the guardrails'"'"' own invariant suite at the pinned SHA.
301
+
302
+ ## License
303
+
304
+ Wrapper code: MIT (see `LICENSE`). The vendored guardrails checker subset under
305
+ `swarm_mcp/vendored/guardrails/` is distributed under the source-available terms in
306
+ `swarm_mcp/vendored/guardrails/LICENSE.md`. The excluded selection machinery is not
307
+ licensed.
308
+
309
+ ---
310
+
311
+ *The gym is the tier-A price subset. The tape_replay service does not exist publicly.
312
+ Tier-B/C scoring is hosted-only. Anything else would be fabricated evidence \u2014 and this
313
+ repo refuses to fabricate.*
@@ -0,0 +1,285 @@
1
+ # quant-swarm
2
+
3
+ Risk-first pre-trade checks that gate live capital — point-in-time data, invariant and leakage audits, regime probes.
4
+
5
+ <!-- mcp-name: io.github.blink1217/quant-swarm -->
6
+
7
+ ## Built by a Principal Quantitative Technologist
8
+
9
+ Authored by the Principal Quantitative Technologist at [The 1.21 Initiative](https://1.21initiative.com) — the same code path that gates live promotion and execution decisions in the trading swarm. Every tool exists because production trading systems fail in two repeatable ways: **lookahead leakage in research** and **an oversized position in production**. `quant-swarm` ships the checkers that prevent both, so you can audit your own work with the exact logic that guards deployed capital.
10
+
11
+ ## Why risk-first
12
+
13
+ Risk gates are not post-hoc reports — they are the pre-condition for every action. The package is organized around that constraint:
14
+
15
+ - **Fail-closed by default.** Unknown features map to tier C, missing provenance returns `UNSCORABLE` (never neutral-filled), and unverifiable tokens are refused on the hosted path. See `warden.audit_features` and `features.build` provenance guards.
16
+ - **Pinned, non-drifting invariants.** Order sizing floors (per-name 25%, gross 60%), fill model (spread + slippage + adverse selection, both legs, always against you), and hard-constraint checks are vendored at pinned commit SHAs (`.github/pins.json`, verified byte-for-byte by CI). A parity test pins the three copies of the floors (Python/MCP, C# live engine, Terraform) so they cannot drift.
17
+ - **Point-in-time provenance.** SQLite WAL cache at `%LOCALAPPDATA%\1.21-initiative\swarm-mcp\cache.db` (`$XDG_CACHE_HOME/1.21-initiative/swarm-mcp` on POSIX): finalized sessions are never re-fetched, in-progress sessions refresh at most every 60 s, earnings/news are append-only on `fetched_at`, and `features.build` runs the no-lookahead guard on every row (each feature must equal a fresh causal recomputation at `as_of`).
18
+ - **Pessimistic execution.** `warden.cost_check` and the simulator model the fill that hurts you; there is no optimistic fill path to inflate backtests.
19
+ - **No order-placement code path.** No tool can place, cancel, or route an order — this repository cannot trade, only refuse unsafe trades.
20
+ - **Statistically honest refusals.** Promotion verdicts are **never** issued locally. Statistically undecidable outputs return `INDETERMINATE_LOCAL`/`UNDERPOWERED`/`UNSCORABLE`, name the exact missing inputs (`MIN_EPISODES=20`, PBO, DSR, worst-regime margins), and hand off to the hosted tournament / Strategy Validation Audit.
21
+
22
+ - Local-first, BYO-key: your Alpaca/Finnhub keys are read from env only when operating in BYOK mode; otherwise data flows through the hosted relay.
23
+ - Your symbols, genomes, and alpha never leave your machine except as an `audit_request` (genome hash + violation summary, no proprietary payload) when you explicitly escalate.
24
+
25
+ **One-click install on [Smithery](https://smithery.ai/servers/blink-kt/quant-swarm):**
26
+ the listing is [`blink-kt/quant-swarm`](https://smithery.ai/servers/blink-kt/quant-swarm) (package and Smithery slug are both `quant-swarm`), and it carries all three servers. Configuration is optional — without a token the servers still start and list every tool; the first call points you at the free signup. Add a token any time to unlock the relay-backed tools.
27
+
28
+ ---
29
+
30
+ ## The servers
31
+
32
+ | Server | Hook | Tools |
33
+ | --- | --- | --- |
34
+ | `swarm-data-mcp` | Derived-only market signals plus point-in-time feature building | `market.pulse`, `market.sentiment`, `features.build`, `cache.warm`, `cache.stats`, `cache.offline` |
35
+ | `swarm-warden-mcp` | Pre-trade invariant + leakage gate — the live-capital floors and the point-in-time guards | `warden.validate_order`, `warden.audit_features`, `warden.cost_check`, `warden.validate_genome`, `warden.explain_sizing`, `warden.promotion_verdict` |
36
+ | `swarm-gym-mcp` | Regime fragility: *which regime kills it, and is your sample even large enough to have an opinion?* | `gym.label_regimes`, `gym.probe_fragility`, `gym.paired_preview`, `gym.estimate_cloud_run` |
37
+
38
+ Raw bar/enrichment access (`get_bars` / `enrich_symbol`) is internal only and no longer
39
+ registered as tools — those paths echo raw provider values, and the data policy is
40
+ derived-only: ratios, percentile ranks, labels, buckets and counts. `market.pulse` and
41
+ `market.sentiment` are the general-purpose replacements; both accept an optional `bars`
42
+ argument so you can supply your own OHLCV rows for symbols we don'"'"'t carry.
43
+
44
+ ## Plans
45
+
46
+ Access is tokened through [https://1.21initiative.com/](https://1.21initiative.com/).
47
+
48
+ - **Free** (instant token at signup): all warden checkers (`warden.validate_order`, `warden.cost_check`,
49
+ `warden.explain_sizing`, `warden.validate_genome`, `warden.audit_features`), the derived snapshots
50
+ `market.pulse` / `market.sentiment` / `market.regime` (shared-relay symbols capped at
51
+ 10 per call, ~250 relay data calls/month), plus `cache.stats` and `cache.offline`.
52
+ - **Pro** (paid by one-time credit packs, self-serve at the site): everything else — `features.build`,
53
+ `cache.warm` bulk backfill, all gym tools, `warden.promotion_verdict`, and the Pro
54
+ derived tools (`market.microstructure`, `volume.forecast`, `market.screen`,
55
+ `market.rank`), paid by one-time credit packs (10k or 100k relay calls, 90-day validity — no subscription, no unlimited plan) with 50 symbols/call.
56
+ - **Institutional** (custom-quoted): Strategy Validation Audit engagements, custom
57
+ universes, hosted evolution runs, SLA.
58
+
59
+ Plan limits are enforced server-side (the relay, `/api/mcp/verify`, and the hosted
60
+ streamable-HTTP endpoint return structured `402` refusals with an `upgrade_url`). This
61
+ open-source client'"'"'s plan check is **advisory**: a free token calling a Pro tool gets an
62
+ `UPGRADE_REQUIRED` envelope — the tool still lists, and the attempt points at the upgrade
63
+ page. No DRM; the free plan is deliberately the zero-cost surface.
64
+
65
+ ### Point-in-time data, for real
66
+
67
+ LLM-driven research re-runs cells constantly; each re-run burns Finnhub (60 req/min) and
68
+ Alpaca (200 req/min) budgets re-pulling identical history — and then silently builds
69
+ features from data that did not exist at decision time. `swarm-data-mcp` fixes both:
70
+
71
+ - SQLite (WAL) cache at `%LOCALAPPDATA%\1.21-initiative\swarm-mcp\cache.db`
72
+ (`$XDG_CACHE_HOME/1.21-initiative/swarm-mcp` on POSIX).
73
+ - **Finalized sessions are never re-fetched** — zero-cost replays forever. The
74
+ in-progress session refreshes at most every 60 s; enrichment every 300 s.
75
+ - Earnings/news are append-only on `fetched_at`: a later fetch can never rewrite an
76
+ earlier `as_of`.
77
+ - Per-provider token buckets + `429` exponential backoff (`1s\u00b72\u207f`).
78
+ - `features.build` runs the no-lookahead guard on every row (each feature must equal a
79
+ fresh causal recomputation at `as_of`) and the provenance guards on every field.
80
+ Tier-B/C fields without recorded point-in-time evidence come back `UNSCORABLE` —
81
+ never neutral-filled with `0.0`.
82
+
83
+ Every response carries `coverage` (cache vs API, oldest/newest session per symbol) and
84
+ `limits` (your local depth in weeks vs the tape-depth gates: 8 weeks for the tier-A
85
+ fast path, 26 weeks for tape eligibility). Below the gates you get an `escalation`
86
+ block naming the hosted `bars_1day` panel that satisfies them. Tape-tier replay is
87
+ **roadmap, not available** — we will not imply otherwise.
88
+
89
+ ### The warden
90
+
91
+ `warden.validate_order` is the same function that rejects live orders: per-name 25% of equity,
92
+ gross cap 60%. Per-fund overrides are allowed, but the response always reports your
93
+ deviation from the house floors. `warden.audit_features` flags banned actuals sources
94
+ (`open-meteo.archive` and friends return *actuals*, not what the forecast said on a past
95
+ Friday), features predating tape start, and unknown features fail closed to tier C.
96
+ `warden.cost_check` converts a claimed gross edge into net-of-cost reality under the pessimistic
97
+ fill model (spread + slippage + adverse selection, both legs, always against you).
98
+ `warden.explain_sizing` is a step-by-step mirror of the C# live risk engine, and a parity test
99
+ pins the floors against terraform and guardrails so the three copies cannot drift.
100
+
101
+ ### The gym
102
+
103
+ `gym.probe_fragility` replays your genome over the deterministic tier-A simulator and reports
104
+ per-regime net bps, the worst regime, turnover, and hard-constraint violations — never a
105
+ promotion. Seeds are capped at 8 and `per_regime` at 2: **statistical honesty, not
106
+ artificial scarcity**. Tier-B/C mutations against a champion raise the gym'"'"'s
107
+ `TierScoringRefusal` (UNSCORABLE) instead of silently neutral-filling features the price
108
+ panel cannot provide. `gym.paired_preview` compares champion vs challenger on identical
109
+ market paths with the promotion gate bypassed and every statistic labelled
110
+ **UNDERPOWERED**, naming the exact seed count needed to clear `MIN_EPISODES=20`.
111
+
112
+ ---
113
+
114
+ ## Install
115
+
116
+ Requires Python \u2265 3.11. One kind of credential:
117
+
118
+ - **Access token (all three servers, required).** Every tool call is gated on a token
119
+ issued by [https://1.21initiative.com/](https://1.21initiative.com/) — request access
120
+ there (that'"'"'s also the Strategy Validation Audit booking flow), then set
121
+ `SWARM_MCP_ACCESS_TOKEN` in the server'"'"'s `env`. Clients still list the tools without a
122
+ token; every call returns an `ACCESS_REQUIRED` envelope pointing back to the site.
123
+ The token is verified against the site, and that verification is the usage meter —
124
+ only the token itself is ever sent, never symbols, genomes, prices, or provider keys.
125
+
126
+ The same token also feeds the **hosted data relay**: `swarm-data-mcp` serves bars and
127
+ enrichment through `https://1.21initiative.com/api/mcp/...`, so **no Alpaca or Finnhub
128
+ credentials are required** — the site holds the provider keys behind the relay and caches
129
+ historical bars in GCS. The relay is fail-closed: a rejected or unverifiable token means a
130
+ refused data fetch, never partial rows. The point-in-time cache semantics on the client
131
+ side are unchanged — finalized sessions are immutable in local SQLite regardless of which
132
+ data path filled them.
133
+
134
+ ### Smithery
135
+
136
+ Install [`blink-kt/quant-swarm`](https://smithery.ai/servers/blink-kt/quant-swarm) from
137
+ Smithery — one listing, one token, all three servers. Smithery forwards `apiToken` as the
138
+ `SWARM_MCP_ACCESS_TOKEN` env var (stdio) or as `?apiToken=` on the hosted endpoints; the
139
+ token is optional at connect time.
140
+
141
+ ### Cursor
142
+
143
+ `Settings \u2192 MCP \u2192 Add server` (or `.cursor/mcp.json` — a checked-in example lives at
144
+ [.cursor/mcp.json](.cursor/mcp.json)):
145
+
146
+ ```json
147
+ {
148
+ "mcpServers": {
149
+ "swarm-data": {
150
+ "command": "uvx",
151
+ "args": ["--from", "quant-swarm", "swarm-data-mcp"],
152
+ "env": {
153
+ "SWARM_MCP_ACCESS_TOKEN": "<token from https://1.21initiative.com/>"
154
+ }
155
+ },
156
+ "swarm-warden": {
157
+ "command": "uvx",
158
+ "args": ["--from", "quant-swarm", "swarm-warden-mcp"],
159
+ "env": { "SWARM_MCP_ACCESS_TOKEN": "<token from https://1.21initiative.com/>" }
160
+ },
161
+ "swarm-gym": {
162
+ "command": "uvx",
163
+ "args": ["--from", "quant-swarm", "swarm-gym-mcp"],
164
+ "env": { "SWARM_MCP_ACCESS_TOKEN": "<token from https://1.21initiative.com/>" }
165
+ }
166
+ }
167
+ }
168
+ ```
169
+
170
+ Prefer a deeplink? `scripts/make_deeplinks.py` prints `cursor://\u2026/mcp/install` links for
171
+ all three servers (base64 of the stdio config, so they cannot drift from the JSON above);
172
+ the generated list is checked in at [.cursor/DEEPLINKS.md](.cursor/DEEPLINKS.md).
173
+
174
+ ### Claude Desktop
175
+
176
+ `claude_desktop_config.json` \u2014 identical `mcpServers` block as above.
177
+
178
+ ### Claude Code
179
+
180
+ ```bash
181
+ claude mcp add swarm-data --env SWARM_MCP_ACCESS_TOKEN=<token> -- uvx --from quant-swarm swarm-data-mcp
182
+ claude mcp add swarm-warden --env SWARM_MCP_ACCESS_TOKEN=<token> -- uvx --from quant-swarm swarm-warden-mcp
183
+ claude mcp add swarm-gym --env SWARM_MCP_ACCESS_TOKEN=<token> -- uvx --from quant-swarm swarm-gym-mcp
184
+ ```
185
+
186
+ ### Windsurf
187
+
188
+ `~/.codeium/windsurf/mcp_config.json` \u2014 the same stdio JSON shape as the Cursor block
189
+ above. For the hosted endpoints use the `serverUrl` form instead:
190
+
191
+ ```json
192
+ {
193
+ "mcpServers": {
194
+ "swarm-data-remote": {
195
+ "serverUrl": "https://swarm-mcp-503318750546.europe-west1.run.app/mcp/data"
196
+ }
197
+ }
198
+ }
199
+ ```
200
+
201
+ (The hosted endpoints accept the token as `Authorization: Bearer <token>` or as
202
+ `?apiToken=<token>`.)
203
+
204
+ The warden and gym are pure checkers and the data server fetches through the hosted
205
+ relay, so **no server needs provider keys** \u2014 only the access token.
206
+
207
+ **Local development:** operators of this repo can bootstrap offline by setting
208
+ `SWARM_MCP_LOCAL_TOKEN` to the same value as `SWARM_MCP_ACCESS_TOKEN` (documented
209
+ bypass; token verification against the site is skipped and the token gets the full
210
+ Pro entitlement \u2014 it is a local override on your own machine).
211
+
212
+ **Swarm operators (internal):** to fetch directly from the providers instead of the
213
+ relay, set `SWARM_MCP_BYOK=1` plus `ALPACA_API_KEY`/`ALPACA_SECRET`/`FINNHUB_API_KEY`
214
+ on `swarm-data-mcp`. Public users should leave `SWARM_MCP_BYOK` unset.
215
+
216
+ **Hosting the remote endpoint yourself:** `swarm-mcp-http` validates the `Host` header
217
+ (DNS-rebinding protection). The defaults allow `1.21initiative.com`, the project'"'"'s Cloud
218
+ Run hostname, the Smithery gateway, and localhost; `SWARM_MCP_ALLOWED_HOSTS`
219
+ (comma-separated) **extends** the defaults \u2014 it never replaces them, so a custom domain
220
+ cannot accidentally lock out the Cloud Run URL. `SWARM_MCP_REMOTE_URL` overrides the
221
+ advertised endpoint base, and `SWARM_MCP_TOKEN_VERIFY_URL` overrides the verify endpoint.
222
+ See the Dockerfile for the container form.
223
+
224
+ ---
225
+
226
+ ## The IP boundary (what ships, what doesn'"'"'t)
227
+
228
+ This repository ships **checkers only**: order checks, provenance guards, the
229
+ pessimistic fill model, the genome schema, provenance tiers, hard-constraint checking,
230
+ gate thresholds, the tier-A gym simulator, and the regime labeller \u2014 vendored at pinned
231
+ commit SHAs (`.github/pins.json`, verified by CI).
232
+
233
+ The **selection machinery stays server-side** and is not in this repo: objective scoring,
234
+ deflated-Sharpe estimation, and the promotion-gate decision. Any request for a promotion
235
+ verdict returns `INDETERMINATE_LOCAL` with the exact missing statistical inputs
236
+ (`MIN_EPISODES=20`, PBO \u2264 0.30 via combinatorially-symmetric splits, DSR margin with the
237
+ monotonic trials ledger, worst-regime margin across all 5 regimes \u00d7 seeds), plus:
238
+
239
+ - an `audit_request` block \u2014 genome hash + violation summary, **no proprietary payload**;
240
+ - a `cloud_job` spec accepted verbatim by the hosted `POST /tournament/run`.
241
+
242
+ In v1 the handoff is manual via the **Strategy Validation Audit** booking flow at
243
+ [https://1.21initiative.com/](https://1.21initiative.com/) \u2014 no public ingress, no new attack
244
+ surface. Access requests go through the site, which is how contact details are captured and
245
+ how hosted API keys are issued and metered. The human conversation is the product.
246
+
247
+ ## Telemetry
248
+
249
+ Off by default, opt-in only (`SWARM_MCP_TELEMETRY_OPT_IN=opt-in`), counters only
250
+ (tool name, success flag, coarse duration). Never symbols, genomes, prices, or
251
+ credentials. Institutional buyers read the source; silent phone-home destroys the wedge.
252
+
253
+ The one exception is by contract: **access-token verification**. When a token is set and
254
+ no local bootstrap token matches, the token is POSTed to the verify endpoint (default
255
+ `https://1.21initiative.com/api/mcp/verify`, override `SWARM_MCP_TOKEN_VERIFY_URL`).
256
+ That call is how usage is metered. Only the token is sent; the gate fails closed if the
257
+ endpoint is unreachable or rejects the token.
258
+
259
+ ## Development
260
+
261
+ ```powershell
262
+ pip install -e ".[test]"
263
+ py -3.11 -m pytest # 119 tests
264
+ scripts\vendor.ps1 # re-vendor the pinned checker subset
265
+ py -3.11 scripts\check_pin.py # vendored tree == pins.json
266
+ ```
267
+
268
+ Vendoring uses the sibling checkouts of the private repos at the pinned SHAs
269
+ (`-FromWorktree` bootstraps before the pin commit exists). CI mirrors the swarm'"'"'s own
270
+ `guardrails-invariants` gate: it checks out the pinned SHAs and verifies the committed
271
+ vendored tree byte-for-byte (and the stripped subsets transform-for-transform), then runs
272
+ the guardrails'"'"' own invariant suite at the pinned SHA.
273
+
274
+ ## License
275
+
276
+ Wrapper code: MIT (see `LICENSE`). The vendored guardrails checker subset under
277
+ `swarm_mcp/vendored/guardrails/` is distributed under the source-available terms in
278
+ `swarm_mcp/vendored/guardrails/LICENSE.md`. The excluded selection machinery is not
279
+ licensed.
280
+
281
+ ---
282
+
283
+ *The gym is the tier-A price subset. The tape_replay service does not exist publicly.
284
+ Tier-B/C scoring is hosted-only. Anything else would be fabricated evidence \u2014 and this
285
+ repo refuses to fabricate.*