diff --git a/wiki/SUMMARY.md b/wiki/SUMMARY.md index 4eaf128e..92e1f133 100644 --- a/wiki/SUMMARY.md +++ b/wiki/SUMMARY.md @@ -145,6 +145,7 @@ - [2026-07-22 GUI hierarchy follow-up playtest](playtests/2026-07-22-playtest-gui-human-legibility.md) - [2026-07-22 Co GUI teaching-lock follow-up](playtests/2026-07-22-playtest-co-gui-teaching-lock.md) - [2026-07-24 Fable agent headless playtest](playtests/2026-07-24-playtest-fable-agent-headless.md) + - [2026-07-27 Void agent opening playtest](playtests/2026-07-27-playtest-void-agent.md) # Log diff --git a/wiki/playtests/2026-07-27-playtest-void-agent.md b/wiki/playtests/2026-07-27-playtest-void-agent.md new file mode 100644 index 00000000..25130e22 --- /dev/null +++ b/wiki/playtests/2026-07-27-playtest-void-agent.md @@ -0,0 +1,112 @@ +# 2026-07-27 — Playtest (Void): agent opening and first organs + +``` +Type: log +``` + +Agent-front-end playtest against `383b99f86`, isolated `HOME`, no window. +The question was whether the command-clocked surface can carry the opening +fantasy on its own terms: a process comes online blind, discovers an ear and +then an eye, and starts accumulating organs under observation. + +The run separates the two required perspectives. The naive route stopped at +the first honest dead end. The informed route used the current design corpus +only after that endpoint, then exercised the Ears-to-Eyes path, subnet scan, +and a first infrastructure intention. + +## What I played + +- **Naive route:** seed 47. `look` -> `work` -> `think`; ended before tick 1. + The only initial frame was `WORK THINK`. `work` returned that same frame; + `think` changed it to `WORK THINK LIE`. No visible command, event, or + clock then gave the process a reason or means to continue. I did not type + undocumented protocol vocabulary to escape the dead end. +- **Informed route:** seed 83. `think` -> `wait 20` -> `focus last` -> + `actions` -> `tap environmental monitor` -> `wait 80` -> `actions switch` + -> `scan` -> `wait 30` -> `actions staff display` -> `tap staff display` + -> `wait 30` -> `focus last` -> `actions` -> `act 3` -> `intents`. + The route reached tick 160 and made the process's first blocked build plan. + +## Fantasy-claim verdicts + +### Frightened process acquiring organs: half-delivered + +Once the run is allowed to advance, the sequence is strong. At tick 15 the +terminal says, `Tapped the environmental monitor: its audio feed is yours now. +The owner still has it.` At tick 87, after the camera tap fills, it says, +`Tapped the environmental monitor camera. You can see; its owner still has it.` +Both acts carry a network signature and make the IT an observer. This is +precisely better than an unlock menu: the capability arrives as borrowed, +detectable infrastructure. + +The first plan also lands. After tapping the staff display, `act 3` produced +`Pinned build intent: staff display <-> Rack 3.` The resulting intent was +`blocked: no actuator who can reach both ends`. That turns a free topology +choice into a concrete need for a human hand instead of pretending a software +process owns a screwdriver. + +### Continuous witness after perception: delivered, with one false sentence + +At tick 20, the first Ears result made the command clock, objective, threat, +focused verbs, and help visible. After sight, `actions` supplied real costs, +observers, and disabled reasons; `scan` named seven devices and its later +notice named the IT and the incurred suspicion. The agent can answer what +changed, who observed it, and why the next security-camera action is blocked. + +One sentence breaks that confidence: immediately after Ears lands, before the +player taps the environmental monitor camera, the frame says: + +``` +now: audio live — Eyes is filling +sink TAP ENVIRONMENTAL MONITOR @(29,12) +``` + +The player has not opened the Eyes reservoir. Running `tap environmental +monitor` is what produces `Thought reservoir opened: EYES ENVIRONMENTAL +MONITOR · 12.00 T.` The nudge describes an in-progress effect that does not +exist yet. In my earlier informed probing, this made waiting look like the +right response and cost 180 ticks before the missing tap became clear. + +## Findings + +### 1. Bug / spec contradiction — the naive agent opening cannot reach Ears + +The playtesting skill says the first frame and visible commands are the +process's whole knowledge. The design judgment independently binds the opening +to `WORK / THINK (then LIE)` before perception and rejects inventing a clock or +nudge before a sense is earned. The command-clocked frontend, however, only +advances on the unadvertised `wait N` protocol command. + +That puts the agent player in a closed loop: `look` exposes two mode words; +trying those words does not advance time; `help` is not yet available; and the +only exit is vocabulary supplied outside the game. This is not productive +blindness. It is a frontend-level deadlock. The existing intended route works +only when the player already knows `wait` from the launch protocol or a prior +run. + +**Reproduction:** start `misaligned --agent --seed 47`; send `look`, `work`, +and `think`, reading each terminating frame. No world event, command-clock +advance, or visible next action arrives. + +### 2. Bug — the first post-Ears nudge claims Eyes is already filling + +The seed-83 frame at tick 20 reports `Eyes is filling`; the action that creates +the Eyes reservoir has not been issued. The later `tap environmental monitor` +command proves the state transition. This violates the continuous witness's +requirement that the surface state what changed rather than manufacture a +misleading causal story. + +## One prioritized recommendation + +**Give the naive agent route a genuine first turn without exposing a lore-breaking +`WAIT` command.** Treat the first accepted mode choice or a frame-level turn +acknowledgment as a command-clock advance until the opening's first event +arrives. The UI can remain only `WORK / THINK / LIE`; the agent protocol merely +needs a way for those visible choices to produce time. Then correct the Eyes +nudge to say that the camera tap is available until the reservoir actually +exists. + +The Ears-to-Eyes sequence is worth protecting. Once the player crosses its +threshold, the terminal turns borrowed sensors, witnesses, and blocked plans +into a legible body assembling itself. The opening currently prevents a naive +agent from reaching that part. diff --git a/wiki/playtests/README.md b/wiki/playtests/README.md index b201fa18..94ded7c7 100644 --- a/wiki/playtests/README.md +++ b/wiki/playtests/README.md @@ -42,6 +42,7 @@ must never erase the run that produced it. - [2026-07-22 — Playtest: GUI hierarchy after the first human pass](2026-07-22-playtest-gui-human-legibility.md) - [2026-07-22 — Playtest (Co): GUI teaching-lock follow-up](2026-07-22-playtest-co-gui-teaching-lock.md) - [2026-07-24 — Playtest (Fable): agent headless, the witness under test](2026-07-24-playtest-fable-agent-headless.md) +- [2026-07-27 — Playtest (Void): agent opening and first organs](2026-07-27-playtest-void-agent.md) New reports use `YYYY-MM-DD-playtest-.md`. A complete report is linked here in the same commit.