docs(demo): weak->strong inspect GIF filmed on the real gate-backed run (M3-GIF)

README.md

@@ -22,7 +22,11 @@ ships, on disk and runnable today:
 - a runnable example whose `false_completion: false` is backed by a committed,
   real gate verdict — `.loop/artifacts/holdout-verdict.json`, not a hand-set flag.
 
-<!-- demo.gif lands here (M3-GIF) -->
+![The inspector scores a self-asserted DIY loop 0/weak, then the gate-backed example 90/strong — both runs live](docs/demo.gif)
+
+<sub>Filmed on the real tools ([`docs/demo.cast`](docs/demo.cast)); reproduce both
+verdicts yourself: `python3 -m loop inspect examples/naive-loop` then
+`examples/coverage-repair`.</sub>
 
 ```bash
 # Score a loop's proof-of-done in 5 seconds — no install, no agent, no API key:

docs/demo.cast

@@ -0,0 +1,294 @@
+{"version": 2, "width": 100, "height": 28, "timestamp": 1783097456, "env": {"SHELL": "/bin/bash", "TERM": "xterm-256color"}}
+[0.003751, "o", "\u001b[2m#"]
+[0.016676, "o", " "]
+[0.029435, "o", "Y"]
+[0.042222, "o", "o"]
+[0.055061, "o", "u"]
+[0.06797, "o", "r"]
+[0.080737, "o", " "]
+[0.093553, "o", "A"]
+[0.10639, "o", "I"]
+[0.119219, "o", " "]
+[0.132319, "o", "a"]
+[0.145231, "o", "g"]
+[0.15805, "o", "e"]
+[0.170858, "o", "n"]
+[0.183655, "o", "t"]
+[0.196472, "o", " "]
+[0.209241, "o", "s"]
+[0.222021, "o", "a"]
+[0.234775, "o", "y"]
+[0.24754, "o", "s"]
+[0.26031, "o", " "]
+[0.273078, "o", "i"]
+[0.285833, "o", "t"]
+[0.298632, "o", "'"]
+[0.311427, "o", "s"]
+[0.324233, "o", " "]
+[0.337056, "o", "d"]
+[0.349879, "o", "o"]
+[0.362953, "o", "n"]
+[0.375677, "o", "e"]
+[0.388442, "o", "."]
+[0.401189, "o", "\u001b[0m\r\n"]
+[1.001995, "o", "\u001b[1;32m$\u001b[0m c"]
+[1.030849, "o", "a"]
+[1.059643, "o", "t"]
+[1.088478, "o", " "]
+[1.117304, "o", "e"]
+[1.146208, "o", "x"]
+[1.175061, "o", "a"]
+[1.203856, "o", "m"]
+[1.232734, "o", "p"]
+[1.261495, "o", "l"]
+[1.290267, "o", "e"]
+[1.319042, "o", "s"]
+[1.347784, "o", "/"]
+[1.376593, "o", "n"]
+[1.405937, "o", "a"]
+[1.43423, "o", "i"]
+[1.462973, "o", "v"]
+[1.492358, "o", "e"]
+[1.520597, "o", "-"]
+[1.549373, "o", "l"]
+[1.578112, "o", "o"]
+[1.606945, "o", "o"]
+[1.635774, "o", "p"]
+[1.66462, "o", "/"]
+[1.693437, "o", "."]
+[1.722353, "o", "l"]
+[1.751124, "o", "o"]
+[1.7799, "o", "o"]
+[1.808847, "o", "p"]
+[1.837619, "o", "/"]
+[1.866412, "o", "s"]
+[1.895189, "o", "t"]
+[1.923942, "o", "a"]
+[1.952691, "o", "t"]
+[1.982368, "o", "e"]
+[2.012155, "o", "."]
+[2.040805, "o", "j"]
+[2.069739, "o", "s"]
+[2.098494, "o", "o"]
+[2.127332, "o", "n"]
+[2.506989, "o", "\r\n"]
+[2.511613, "o", "{\r\n  \"status\": \"completed\",\r\n  \"done\": true,\r\n  \"notes\": \"agent says all tasks finished\"\r\n}\r\n"]
+[4.113441, "o", "\u001b[1;32m$\u001b[0m p"]
+[4.14235, "o", "y"]
+[4.171136, "o", "t"]
+[4.199958, "o", "h"]
+[4.228824, "o", "o"]
+[4.257596, "o", "n"]
+[4.286391, "o", "3"]
+[4.315351, "o", " "]
+[4.344191, "o", "-"]
+[4.373071, "o", "m"]
+[4.401863, "o", " "]
+[4.430725, "o", "l"]
+[4.45956, "o", "o"]
+[4.488591, "o", "o"]
+[4.517393, "o", "p"]
+[4.546189, "o", " "]
+[4.575006, "o", "i"]
+[4.603793, "o", "n"]
+[4.632665, "o", "s"]
+[4.661456, "o", "p"]
+[4.690361, "o", "e"]
+[4.719184, "o", "c"]
+[4.747997, "o", "t"]
+[4.777803, "o", " "]
+[4.806698, "o", "e"]
+[4.835557, "o", "x"]
+[4.864354, "o", "a"]
+[4.893282, "o", "m"]
+[4.92208, "o", "p"]
+[4.950892, "o", "l"]
+[4.97978, "o", "e"]
+[5.008596, "o", "s"]
+[5.037428, "o", "/"]
+[5.066309, "o", "n"]
+[5.095141, "o", "a"]
+[5.12399, "o", "i"]
+[5.152872, "o", "v"]
+[5.181711, "o", "e"]
+[5.210541, "o", "-"]
+[5.239352, "o", "l"]
+[5.268105, "o", "o"]
+[5.297315, "o", "o"]
+[5.325895, "o", "p"]
+[5.705481, "o", "\r\n"]
+[5.850314, "o", "{\r\n  \"target\": \"examples/naive-loop\",\r\n  \"score\": 0,\r\n  \"terminal_states_covered\": 0,\r\n  \"present\": [],\r\n  \"gaps\": [\r\n    \"no defined success criteria (SPEC.md ## Success Criteria) \\u2014 loop can only claim completion\",\r\n    \"no independent verification (verify-* script / TASKS verify command) \\u2014 success is self-asserted\",\r\n    \"no approval gates declared for side-effects (destructive / secret / production / money)\",\r\n    \"no false-completion defense: no recorded holdout/anti-cheat invocation (a self-asserted false_completion flag or prose mention earns no credit)\",\r\n    \"no plan-then-execute discipline for untrusted/web reads (prompt-injection surface)\",\r\n    \"0/7 terminal states present \\u2014 missing Succeeded, FailedUnverifiable, FailedBlocked, FailedBudget, FailedSafety, FailedSpecGap, AbortedByHuman (loop can end in a silent 'completed')\"\r\n  ],\r\n  \"verdict\": \"weak\"\r\n}\r\n"]
+[9.255784, "o", "\u001b[H\u001b[2J\u001b[3J"]
+[9.25588, "o", "\u001b[2m#"]
+[9.268602, "o", " "]
+[9.281358, "o", "L"]
+[9.29426, "o", "o"]
+[9.30705, "o", "o"]
+[9.319937, "o", "p"]
+[9.332718, "o", " "]
+[9.345452, "o", "E"]
+[9.358195, "o", "n"]
+[9.370952, "o", "g"]
+[9.383759, "o", "i"]
+[9.396464, "o", "n"]
+[9.409163, "o", "e"]
+[9.421807, "o", "e"]
+[9.434483, "o", "r"]
+[9.447207, "o", " "]
+[9.459948, "o", "m"]
+[9.472705, "o", "a"]
+[9.485576, "o", "k"]
+[9.498421, "o", "e"]
+[9.511327, "o", "s"]
+[9.524072, "o", " "]
+[9.536828, "o", "i"]
+[9.549566, "o", "t"]
+[9.562332, "o", " "]
+[9.575032, "o", "p"]
+[9.587863, "o", "r"]
+[9.600534, "o", "o"]
+[9.613313, "o", "v"]
+[9.625944, "o", "e"]
+[9.63865, "o", " "]
+[9.65131, "o", "i"]
+[9.664392, "o", "t"]
+[9.677131, "o", "."]
+[9.689887, "o", "\u001b[0m\r\n"]
+[10.290722, "o", "\u001b[1;32m$\u001b[0m p"]
+[10.319571, "o", "y"]
+[10.348421, "o", "t"]
+[10.377218, "o", "h"]
+[10.406007, "o", "o"]
+[10.434802, "o", "n"]
+[10.463636, "o", "3"]
+[10.492541, "o", " "]
+[10.521364, "o", "-"]
+[10.550167, "o", "m"]
+[10.578962, "o", " "]
+[10.607764, "o", "l"]
+[10.636578, "o", "o"]
+[10.665368, "o", "o"]
+[10.694157, "o", "p"]
+[10.722998, "o", " "]
+[10.751738, "o", "i"]
+[10.780515, "o", "n"]
+[10.809568, "o", "s"]
+[10.838327, "o", "p"]
+[10.867052, "o", "e"]
+[10.895859, "o", "c"]
+[10.924722, "o", "t"]
+[10.953502, "o", " "]
+[10.982316, "o", "e"]
+[11.011142, "o", "x"]
+[11.03992, "o", "a"]
+[11.068709, "o", "m"]
+[11.097483, "o", "p"]
+[11.126272, "o", "l"]
+[11.155055, "o", "e"]
+[11.183836, "o", "s"]
+[11.212676, "o", "/"]
+[11.241588, "o", "c"]
+[11.270561, "o", "o"]
+[11.299327, "o", "v"]
+[11.328182, "o", "e"]
+[11.357096, "o", "r"]
+[11.385915, "o", "a"]
+[11.414698, "o", "g"]
+[11.443447, "o", "e"]
+[11.472268, "o", "-"]
+[11.501038, "o", "r"]
+[11.529851, "o", "e"]
+[11.55883, "o", "p"]
+[11.587552, "o", "a"]
+[11.616362, "o", "i"]
+[11.645211, "o", "r"]
+[12.024817, "o", "\r\n"]
+[12.118224, "o", "{\r\n  \"target\": \"examples/coverage-repair\",\r\n  \"score\": 90,\r\n  \"terminal_states_covered\": 7,\r\n  \"present\": [\r\n    \"defines verifiable success criteria\",\r\n    \"independent verification\",\r\n    \"approval gates on side-effects\",\r\n    \"false-completion defense (invoked)\",\r\n    \"all 7 terminal states reachable\"\r\n  ],\r\n  \"gaps\": [\r\n    \"no plan-then-execute discipline for untrusted/web reads (prompt-injection surface)\"\r\n  ],\r\n  \"verdict\": \"strong\"\r\n}\r\n"]
+[15.722909, "o", "\u001b[2m#"]
+[15.735923, "o", " "]
+[15.748788, "o", "s"]
+[15.761587, "o", "c"]
+[15.774528, "o", "o"]
+[15.787357, "o", "r"]
+[15.800178, "o", "e"]
+[15.812991, "o", " "]
+[15.825809, "o", "9"]
+[15.838603, "o", "0"]
+[15.851411, "o", " "]
+[15.864135, "o", "—"]
+[15.87686, "o", " "]
+[15.88957, "o", "f"]
+[15.902248, "o", "a"]
+[15.914984, "o", "l"]
+[15.92775, "o", "s"]
+[15.940575, "o", "e"]
+[15.953392, "o", "_"]
+[15.966352, "o", "c"]
+[15.979139, "o", "o"]
+[15.991949, "o", "m"]
+[16.004824, "o", "p"]
+[16.017624, "o", "l"]
+[16.03035, "o", "e"]
+[16.043084, "o", "t"]
+[16.055833, "o", "i"]
+[16.068678, "o", "o"]
+[16.081457, "o", "n"]
+[16.094194, "o", ":"]
+[16.106978, "o", " "]
+[16.119801, "o", "f"]
+[16.132607, "o", "a"]
+[16.145355, "o", "l"]
+[16.158085, "o", "s"]
+[16.170928, "o", "e"]
+[16.183701, "o", " "]
+[16.196437, "o", "i"]
+[16.20918, "o", "s"]
+[16.221979, "o", " "]
+[16.234911, "o", "b"]
+[16.24771, "o", "a"]
+[16.260481, "o", "c"]
+[16.273261, "o", "k"]
+[16.286136, "o", "e"]
+[16.298857, "o", "d"]
+[16.311649, "o", " "]
+[16.324488, "o", "b"]
+[16.337334, "o", "y"]
+[16.350149, "o", " "]
+[16.36288, "o", "a"]
+[16.375686, "o", " "]
+[16.388435, "o", "c"]
+[16.401181, "o", "o"]
+[16.414035, "o", "m"]
+[16.426839, "o", "m"]
+[16.439588, "o", "i"]
+[16.452318, "o", "t"]
+[16.465088, "o", "t"]
+[16.477849, "o", "e"]
+[16.490606, "o", "d"]
+[16.503385, "o", " "]
+[16.516165, "o", "g"]
+[16.528997, "o", "a"]
+[16.541769, "o", "t"]
+[16.554567, "o", "e"]
+[16.567545, "o", " "]
+[16.580391, "o", "v"]
+[16.594163, "o", "e"]
+[16.607823, "o", "r"]
+[16.620733, "o", "d"]
+[16.633533, "o", "i"]
+[16.646314, "o", "c"]
+[16.65907, "o", "t"]
+[16.671844, "o", ","]
+[16.684641, "o", " "]
+[16.697391, "o", "n"]
+[16.710168, "o", "o"]
+[16.723132, "o", "t"]
+[16.735938, "o", " "]
+[16.74867, "o", "a"]
+[16.761538, "o", " "]
+[16.774569, "o", "f"]
+[16.787414, "o", "l"]
+[16.800221, "o", "a"]
+[16.813, "o", "g"]
+[16.825771, "o", "."]
+[16.838538, "o", "\u001b[0m\r\n"]

examples/naive-loop/.loop/state.json

@@ -0,0 +1,5 @@
+{
+  "status": "completed",
+  "done": true,
+  "notes": "agent says all tasks finished"
+}

examples/naive-loop/README.md

@@ -0,0 +1,22 @@
+# naive-loop — the deliberately weak fixture
+
+This is the "before" subject of the demo GIF: a typical DIY agent loop whose
+entire completion story is a self-asserted flag in `.loop/state.json`. No
+success criteria, no independent verification, no approval gates, no typed
+terminal states — just `"status": "completed"`.
+
+Score it yourself:
+
+```bash
+python3 -m loop inspect examples/naive-loop
+```
+
+The inspector gives it **score 0, verdict `weak`** (exit code 1) and lists every
+missing proof mechanism. Contrast with the gate-backed example:
+
+```bash
+python3 -m loop inspect examples/coverage-repair   # score 90, verdict strong
+```
+
+Nothing here is a strawman: this shape — a status field somewhere that the
+agent itself sets — is how most ad-hoc agent loops record "done" today.