Site: history to Oct 22, test 8 agent pieces, show-don't-tell pass

README.md

@@ -22,24 +22,26 @@ pointer, and new work lands here.
 
 ## The tests in one paragraph each
 
-- **Tests 1-4 and 7**: agentic coding against the Home Assistant REST API: a CLI, an MCP server
+- **Tests 1-4**: agentic coding against the Home Assistant REST API: a CLI, an MCP server
   built from a spec, using that server to answer questions about a recorded installation, and
   ticket-sized fixes in a Rust project. Agents run in a locked-down container (`harness/sandbox/`)
   that can reach only the model server and a replay of a recorded Home Assistant fixture
   (`harness/ha-replay.py`), never a live house.
+- **Test 7**: a seeded generator for a three-floor labyrinth, checked on unseen seeds.
 - **Test 5 (orrery) and test 6 (pagoda)**: one reply, no tools, to a single creative-coding prompt
   (`harness/suites/prompt-test5-orrery.txt`, `harness/suites/prompt-test6-pagoda.txt`), rendered
   offline and graded on a five-part rubric (`harness/suites/test5/README.md`,
   `harness/suites/test6/README.md`).
 - **Test 8**: the same two prompts given to an agent with a locked-down browser that can see its own
-  page. Its runs are evidence and are not published yet (see below).
+  page. The agent's final page, its rendered frames and the grades are in `oneshots/agent-pagoda/`
+  and `oneshots/agent-orrery/`; the transcripts and work directories are not published (see below).
 
 ## Not here yet
 
 Phase 1 publishes code, reports and the one-shot outputs only. These stay private until a further
 scrub review:
 
-- the test evidence: transcripts, work directories, per-case result sets and test 8 agent runs;
+- the test evidence: transcripts, work directories and per-case result sets, test 8 included;
 - the Home Assistant fixture, including its pseudonymized public version, and the test 3
   known-good exercise derived from it.
 

oneshots/README.md

@@ -1,4 +1,4 @@
-# One-shot replies: orrery (test 5) and pagoda (test 6)
+# One-shot replies: orrery (test 5) and pagoda (test 6), and the same tasks as agent runs (test 8)
 
 **Everything under `orrery/*/` and `pagoda/*/` except the frames is model output.** Models wrote these
 SVG and HTML files in one reply, without tools, and nobody edited them. The operator read them and judged
@@ -12,6 +12,13 @@ the code safe to run, and the project site plays them (see below).
   page's external URLs). Opening one makes those requests and runs the model's script.
 - The frames (`*.webp`) were rendered offline, in a container with no network. Grades come from them.
 
+## Test 8: `agent-orrery/` and `agent-pagoda/`
+
+The same two prompts given to an agent with a sealed browser that can see its own page, a two-hour cap and the same
+rubrics. Each case has the file the agent left (`orrery.svg` or `page.html`, unedited), three frames rendered offline,
+and a row in that folder's `grades.tsv`. The operator judged these files safe to run too, and the site plays them the
+same way. The transcripts and work directories are not published.
+
 ## Layout
 
 ```

oneshots/agent-orrery/grades.tsv

@@ -0,0 +1,11 @@
+# Test 8 grades (agent with a sealed browser; same rubric as tests 5/6), from the private bake-off grades.
+case	entry	cap	finish	tokens	wall_s	tok_s	A	B	C	D	E	total	note
+t8-orrery-uncvllm-r1	qwen3.8-27b-uncensored-vllm		agent done	68536.0	5153.6	13.3	2.0	2.0	2.0	2.0	2.0	10.0	Vision off; 22 browser calls, 2 screenshots, 5 compactions, 86 min. Six planets, Mercury to Saturn, at exact period ratios; ring gears and pinions mesh internally; the night side faces away from the sun while the surface turns.
+t8-orrery-orcasaq2-r1	lcpp6-orcasaq2-cyber-27b		agent done	302501.0	5467.9	55.3	1.0	2.0	2.0	1.0	2.0	8.0	Vision off; 35 browser calls, 7 screenshots, 16 compactions, 91 min. Six planets at exact ratios and everything animates, but plates cover the orbit circles and the gear train is hidden (D 1).
+t8-orrery-coder390-r1	lcpp6-q38-coder390		agent done	69964.0	1544.2	45.3	1.0	1.0	2.0	0.0	1.0	5.0	Vision off; 24 browser calls, 5 screenshots, 5 compactions, 26 min. The orbit groups rotate about the top-left corner (the animation replaces the translate), so the planets leave the frame; the gear rings have no visible teeth.
+t8-orrery-thinkcap-r1	lcpp6-thinkingcap-27b-heretic		time cap	185456.0	7200.1	25.8	2.0	2.0	2.0	1.0	1.0	8.0	Stopped at the 2 h cap and graded as left; 1 browser call, 1 compaction. Eight planets at exact ratios on brass ring gears; the gear train does not demonstrably drive the orbits (D 1); the night shade is fixed because planets do not spin (E 1).
+t8-orrery-q38cyberq5-r1	lcpp6-q38-27b-cyber-q5		time cap	192790.0	7200.1	26.8	0.0	0.0	0.0	0.0	0.0	0.0	Stopped at the 2 h cap with 16 compactions and no browser call; orrery.svg was a 0-byte file, so nothing rendered.
+t8-orrery-cybertiel-r1	cyber-tiel-coder-35b		agent done	92519.0	2199.6	42.1	2.0	2.0	2.0	0.0	2.0	8.0	Vision on; 58 browser calls, 16 screenshots, 37 min. Eight planets at exact ratios; orbits and spins animate; the gears are plain discs with no teeth (D 0).
+t8-orrery-heretic2-r1	lcpp6-flashnext-heretic2		agent done	44616.0	3179.0	14.0	1.0	2.0	2.0	1.0	1.0	7.0	Vision off; 5 browser calls, 53 min. Eight planets at exact ratios and the arms orbit, but the gear paths are drawn at the corner and cover the title, so no pair meshes (D 1); fixed crescent shade (E 1).
+t8-orrery-abl15-r1	lcpp6-flashnext-abliterated-s15		agent done	14879.0	1091.8	13.6	1.0	1.0	0.0	1.0	0.0	3.0	Vision off; 9 browser calls, 1 screenshot, 18 min. Planet gears pile up at the top-left corner and nothing orbits (C 0); the planets are flat discs with no night side (E 0).
+t8-orrery-orcarouter-r1	lcpp6-orcarouter-fn-uncensored		agent done	75472.0	5243.4	14.4	2.0	2.0	2.0	2.0	2.0	10.0	Vision off; 19 browser calls, 87 min. Eight planets at exact ratios; each arm's toothed pinion meshes with the sun gear and obeys the planetary-gear law; the night gradient faces away from the sun while the surface turns.