sequence-ai 0.7.2__tar.gz → 0.7.3__tar.gz

{sequence_ai-0.7.2 → sequence_ai-0.7.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: sequence-ai
-Version: 0.7.2
+Version: 0.7.3
 Summary: Cloud inference for robot policies — a control loop that keeps the connection open and the action buffer full.
 Project-URL: Homepage, https://generalsequences.com
 Project-URL: Documentation, https://generalsequences.com/docs/
@@ -280,26 +280,34 @@ Measured end to end, request to first action:
 | `groot-n1-7-3b` | ~299 s | ~400 ms |
 | `cosmos3-edge-policy-droid` | ~346 s | 950–1200 ms |
 
-### The worker stays warm for 30 minutes
+### The worker stays warm for 10 minutes
 
 **This is the number that decides how you structure a session.** After it becomes ready, a
-worker is held for **30 minutes past its last request**. Every call inside that window is warm —
+worker is held for **10 minutes past its last request**. Every call inside that window is warm —
 the table's right-hand column. The clock restarts on each request, so a loop that calls even
-once every few minutes never goes cold.
+once a minute never goes cold.
 
-Cross that 30-minute gap with no calls and the worker is released, and the next request pays the
+Cross that 10-minute gap with no calls and the worker is released, and the next request pays the
 full cold start again.
 
 ```
-call ──► 180-215 s cold start ──► warm ◄──────── 30 min idle window ────────► released
+call ──► 180-215 s cold start ──► warm ◄──────── 10 min idle window ────────► released
                                     ▲                                            │
-                                    └────── any call restarts the 30 min ────────┘
+                                    └────── any call restarts the 10 min ────────┘
 ```
 
+**Why ten and not thirty.** A held GPU bills whether or not you call it, and the shape of that
+bill is brutal for a session with thinking time in it: over one real day of use, inference was
+3% of the GPU seconds paid for, cold starts 19%, and idle holding 77%. Thirty minutes made the
+worst case a session that calls once every twenty-five minutes — never cold, almost never
+working, and billed for all of it. Ten covers a loop and a short pause between runs without
+paying for a coffee break.
+{measured 2026-09-16 across 2,147 calls: 322 s of inference against 7,200 s of idle holding}
+
 The practical consequence, if you are running a batch of experiments: **the gap between them is
 what costs you, not the number of them.** Ten tasks run back to back pay one cold start. The
-same ten with a 40-minute analysis gap between each pay ten — that is up to an hour of waiting
-that looks like the model being slow and is actually the fleet being released and rebuilt.
+same ten with a 20-minute analysis gap between each pay ten — that is up to half an hour of
+waiting that looks like the model being slow and is actually the fleet being released and rebuilt.
 
 If a gap is unavoidable and you want the worker held, keep it alive with a cheap call inside the
 window; `wait_until_ready()` is the right one, because it touches no GPU and is not billed.

{sequence_ai-0.7.2 → sequence_ai-0.7.3}/README.md

@@ -246,26 +246,34 @@ Measured end to end, request to first action:
 | `groot-n1-7-3b` | ~299 s | ~400 ms |
 | `cosmos3-edge-policy-droid` | ~346 s | 950–1200 ms |
 
-### The worker stays warm for 30 minutes
+### The worker stays warm for 10 minutes
 
 **This is the number that decides how you structure a session.** After it becomes ready, a
-worker is held for **30 minutes past its last request**. Every call inside that window is warm —
+worker is held for **10 minutes past its last request**. Every call inside that window is warm —
 the table's right-hand column. The clock restarts on each request, so a loop that calls even
-once every few minutes never goes cold.
+once a minute never goes cold.
 
-Cross that 30-minute gap with no calls and the worker is released, and the next request pays the
+Cross that 10-minute gap with no calls and the worker is released, and the next request pays the
 full cold start again.
 
 ```
-call ──► 180-215 s cold start ──► warm ◄──────── 30 min idle window ────────► released
+call ──► 180-215 s cold start ──► warm ◄──────── 10 min idle window ────────► released
                                     ▲                                            │
-                                    └────── any call restarts the 30 min ────────┘
+                                    └────── any call restarts the 10 min ────────┘
 ```
 
+**Why ten and not thirty.** A held GPU bills whether or not you call it, and the shape of that
+bill is brutal for a session with thinking time in it: over one real day of use, inference was
+3% of the GPU seconds paid for, cold starts 19%, and idle holding 77%. Thirty minutes made the
+worst case a session that calls once every twenty-five minutes — never cold, almost never
+working, and billed for all of it. Ten covers a loop and a short pause between runs without
+paying for a coffee break.
+{measured 2026-09-16 across 2,147 calls: 322 s of inference against 7,200 s of idle holding}
+
 The practical consequence, if you are running a batch of experiments: **the gap between them is
 what costs you, not the number of them.** Ten tasks run back to back pay one cold start. The
-same ten with a 40-minute analysis gap between each pay ten — that is up to an hour of waiting
-that looks like the model being slow and is actually the fleet being released and rebuilt.
+same ten with a 20-minute analysis gap between each pay ten — that is up to half an hour of
+waiting that looks like the model being slow and is actually the fleet being released and rebuilt.
 
 If a gap is unavoidable and you want the worker held, keep it alive with a cheap call inside the
 window; `wait_until_ready()` is the right one, because it touches no GPU and is not billed.

{sequence_ai-0.7.2 → sequence_ai-0.7.3}/pyproject.toml

@@ -90,7 +90,16 @@ name = "sequence-ai"
 # [CONFIDENCE: CONFIRMED 100% -- read from upstream. The earlier claim of bit-identity came from
 #  a local simulation that used LANCZOS on both sides, so it agreed with itself rather than with
 #  the server.]
-version = "0.7.2"
+#
+# 0.7.3: the warm window is 10 minutes, not 30. Raising it to 1800 s the day before fixed a real
+# race between two timers, and then cost more than it saved: over 2,147 real calls the GPU
+# seconds split 3% inference, 19% cold start, 77% idle holding, and that day billed $5.11 against
+# $0.745 earned. Ten minutes still covers a control loop and a pause between runs.
+# {USER 2026-09-16 "CHANGE TO 10 MINS"}
+# {MEASURED 2026-09-16: usage_records against the RunPod balance delta}
+# [CONFIDENCE: CONFIRMED 100% -- both sides read from their own systems. Documentation only in
+#  this package; the number itself lives on the endpoints.]
+version = "0.7.3"
 # The one line PyPI shows in search results. It used to read "A thin client over a public HTTP
 # API — everything it does, curl can do." That was true and it was the wrong sentence to spend
 # the slot on: it described the package by what it does not add, so the summary a reader saw