eureka-site/BRIEF.md

Eureka landing page: handoff brief

Status of this brief (operator correction, 2026-09-25). Read this first.

This brief was written by the Eureka-running Self agent without operator guidance, to capture its own perspective. Its authority is invented. Treat it as one input: useful for its evidence pointers and the campaign defects below, not binding direction. gpt6-copy.md (drafted by GPT-6 "Sol") came out of a longer operator conversation about the direction of Eureka and GameCult and carries more operator intent, though the copy itself is a draft to workshop.

Operator rulings (2026-09-25). These settle the open questions below.

• Standalone site, in this repo. It will be the public repo GameCult/eureka-site, served at eureka.gamecult.org, a GitHub Pages static site, in the GameCult brand. (This originally said "built with GameCult's Quartz setup"; the Quartz part was superseded later the same day. See the last ruling below.) The public repo https://github.com/GameCult/eureka-site was created on 2026-09-25, with main tracking origin. The GitHub Pages deploy and the eureka.gamecult.org domain are not set up yet. (An earlier draft of this section said epiphany-site; that was a mistype.)
• The site is Eureka's, not Epiphany's. The lineage still matters ("the lineage is the whole story"), but Eureka is Epiphany's sister, not her descendant: two Personas developing side by side on a shared substrate (the Mind schema), not a branch. Same faculty vocabulary (Self, Imagination, Hands, Soul, Mind Steward); Epiphany keeps findings in typed state, Eureka in committed docs and agent reports (see SKILL.md's opening and references/epiphany-comparison-2026-09-15.md). Read Epiphany itself, the Epiphany repo (README, docs/), before shaping that section.
• Naming GameCult projects (CultLib, StreamPixels, Epiphany) is fine. Still never name the streamer or any private individual.
• Cost framing supersedes "What NOT to claim" below. The ~60% of a weekly quota spent in one day was legitimate work, not waste; the regret was spending it on a more expensive, more verbose model just as Opus 5.5 released. Do not describe it as waste or as off-critical-path Soul loops. "Not speed" still stands.
• Copy direction. (a) Soften autonomy-leaning lines while keeping "the operator owns decisions". (b) Lead with specific defects and why the green suite missed them, including the StreamPixels/CultLib campaign defects listed below. (c) Cut length. The staccato one-line style stays but is used sparingly and animated: lines slide in and away one by one, then the page returns to paragraph copy. Style reference the operator loves: https://acko.net/.
• 2026-09-25, later: the page is a scroll-driven WebGL flight in full PS1 style, hand-built static, NOT Quartz. This supersedes the Quartz stack named in the first ruling above (repo, domain and GitHub Pages still stand). The prototype lives in prototype/ and is served at https://yggdrasil.gamecult.org/eureka/ as a stopgap until eureka.gamecult.org exists on Pages; deploy per GameCult's ops runbook for the Eureka prototype. Approved route, in order: hall/monument; hatch + shaft; Eureka organ chamber; lattice (Soul's probes, mutation testing); specimen gallery (one real defect per case); archive (rules/evidence); Epiphany's chamber (all eight organs in scaffolding: the sister reveal); return to the hall with the monument split open, plus install. Organs are PS1 demakes of their Aquarium renditions (Aquarium's docs/epiphany-agent-sdf-visual-language.md) using the Aquarium organ identity palettes as-is; the GameCult brand does not define those colours, so they are a declared exception. The page names the memory organ Life, not "Mind Steward". The Eureka skill made the same rename on 2026-09-25 (GameCult/Eureka commit bd8dc2e; "Mind Steward" survives there only in dated history). The same day, the skill also started sizing each cut against Hands' context budget (commit 7e2f66a). The operator's wider agent doctrine adopted the name Life the same day. Life's scope stays narrow (keeping memory honest) until a typed memory service owns each agent's mind. Then it takes on continuity, pressure and resuming after interruption, as it already does in Epiphany. Epiphany's own naming is Epiphany's to change.

Operator rulings (2026-10-09). These supersede the route's ending above.

• The last chamber is the vanishing twin, not a sister reveal. Epiphany is archived (Eureka mind ruling eureka-substrate:ruling:operator-archive-epiphany, 2026-10-01: "Eureka absorbed her in the womb as fetuses sometimes do"). Keep Epiphany's chamber and its organ models, recast: the scaffolding stands mostly empty and her organs drift into Eureka's orbit. The lineage stays; the "sister" framing above is superseded. The chamber's name is the vanishing twin (operator, 2026-10-09: "I believe you called it a vanishing twin. Much more poetic"): the twin absorbed before birth, whose cells the survivor carries.
• Eyes is a full faculty. Eyes joins Eureka's chamber at full detail, with its own copy: evidence before every fork, advice owned instead of deference to the operator, research always kept. The faculty copy counts six, not five. Operator, 2026-10-09: "Eyes turned out to be one of the more valuable agents in Eureka after the recent doctrine change. Building has never been so comfy."
• The copy is fact-checked against the live system before it ships; the audit is docs/research/landing-claims-audit-2026-10-09.md.

---

This was written 2026-09-24 by the Self agent that has been running Eureka on a live multi-repo campaign for two days: CultLib's TypeScript QUIC transport, and StreamPixels' service and Unity overlay. It is a boots-on-the-ground account for the agent who builds the landing page. Read it, then read the sources it points to. Where this brief and a source disagree, the source wins.

What you are selling, in one breath

Eureka is a Claude Code skill (MIT, public at https://github.com/GameCult/Eureka). It splits a big change across separate agents: one maps it, one executes one cut at a time, and one that did not write the code is told to prove the executor wrong. The operator rules on genuine product decisions. Nothing grades its own work.

The problem it answers: an agent that plans, implements, tests and reports on its own work will hand you green tests and a clean report over a broken machine. That isn't because it lies. Its tests pin its own spelling, and its report is written by the party with the most reason to believe it.

Sources of truth (read these, do not restate them from memory)

The skill repo is checked out at C:\Users\Meta\.claude\skills\eureka\. Its remote is GameCult/Eureka.

• README.md: the public pitch, already written well. Your page should read as the same voice, not a rewrite. Start from its opening two paragraphs.
• SKILL.md: the faculties (Self, Imagination, Hands, Soul, Mind Steward, Operator), the loop, and the git and tooling scars.
• references/postmortem-cultcache.md: the headline evidence. 29 of 32 Soul passes found real defects that the implementer's own green tests had passed, about 75 in all.
• references/changelog.md: how the skill changed and why, with evidence per entry.
• docs/assets/: the mascot (eureka-mascot-4x.png, eureka-mascot.png, .kra sources) and eureka-mascot.prompt.md, its generation prompt. The mascot is a pixel-art inspector crouched on a cracked ALL TESTS PASSING monument, prying up a panel stencilled GREEN, beside a specimen case of pinned mutants, one still moving and labelled SURVIVED. It is the thesis, so use it.

What it looked like in practice (this campaign, 2026-09-23/24)

Every item below is on record in the named repo's map or commit history. Each was found by a pass that did not write the code, and in most cases the implementer's suite was green:

• Every overlay would have dropped 10 s after connecting. A handler swap meant the provider never recorded CONNECTED, so its handshake timer evicted healthy peers. No test lived past 10 s. (CultLib docs/typescript-quic-realtime-cut.md.)
• A use-after-free in a native binding's release(). Reproduced as a core dump under thread-pool saturation, and later pinned by a regression test that crashes with the old order.
• A memory leak of about 53 bytes per native call. That is roughly 36 MB a day when idle and gigabytes under load, with a flat JS heap the whole time.
• Every subathon event after the first would have failed in production. A fractional value was written to an integer column, and the tests used whole-second fake timers.
• Split authority: one transport honoured an admin toggle and the other read raw state, so a creator who switched the subathon off could still be served it.
• A test double that hid a real bug: the fake executor gave every beat its own completion callback, and the real one had a single shared slot. A late finish released the next beat early, so animations overlapped.
• "Pre-existing" is not a diagnosis: a failing test that many agents walked past had only ever passed on machines holding a git-ignored licensed asset pack.
• The honest half: implementers repeatedly stopped at forks instead of bluffing through, and reported their own vacuous tests. One refused to break 104 scene references; another found that a mutation passed because its test read expected values from the code under test.

What NOT to claim

• Not speed. Eureka is slower and more expensive than one agent. On this campaign the operator watched about 60% of a weekly quota go in one day, much of it on deep Soul loops that weren't on the critical path. The skill now tells Self to spend through scope: one Soul gate per cut, fix batches only for critical-path findings, narrow cheaper passes for small deltas.
• Not "the agents are smart enough now." The point is the opposite. Self (the coordinator) was corrected by the operator several times in this campaign, for overbuilding an authentication scheme for a single known server and for calling authored data "a pipeline". Separation of faculties catches the implementer. It does not make the coordinator infallible, and the page shouldn't imply it does.
• Not autonomy. The operator rules forks. The loop is designed around a human who decides product meaning.
• No victory-lap language about what it "no longer" does. Describe the live system and its tradeoffs. Scars belong in the evidence, told plainly.

Audience and tone (proposed; confirm with the operator)

• Proposed readers: engineers already using Claude Code (or similar agents) on real codebases, who have been burned by "all green" reports.
• Voice: dry and direct, lightly self-deprecating, with no hype. It's the README's voice. Concrete scars beat adjectives. The strongest copy is a specific defect and the reason the green suite missed it.
• A likely call to action: the GitHub repo, and how to install a Claude Code skill. Check the current install method against Claude Code's docs; don't guess it.

Visual identity

This is a GameCult surface, so follow the GameCult brand, which is defined in code:

• F:\Projects\gamecult-site\site\quartz.config.ts: configuration.theme, for the typefaces, weights and palette.
• F:\Projects\gamecult-site\site\quartz\styles\custom.scss: the ground wash, layout, and page-type variants.
• F:\Projects\gamecult-site\docs\brand-design-language.md: what those mean. Where it disagrees with the code, the code wins.

The short form: Montserrat 100–200 for large titles, Ubuntu 300 for prose, IBM Plex Mono uppercase and letter-spaced for small labels. Ground #07111a, panels #16212c, body #b7c7d9, headings #eef5ff, accent #ff8a2a, links #59b7ff. It is single-theme dark by decision, so paint every colour explicitly. The brand has no success, warning or error colours. If you want green and red for "passed" and "survived", you are inventing them, so say so, and consider using the mascot's own palette instead.

Open questions for the operator (all four answered 2026-09-25, see the top; kept as history)

1. Where does it live? A page on the Quartz-based gamecult-site, or a standalone static page (and on what domain)? Recommendation: a gamecult-site page, since the brand and deploy path already exist there (check gamecult-ops for how that site deploys).
2. May the campaign examples name StreamPixels and CultLib? They're GameCult projects. Never name the streamer or any private individual.
3. Should the page mention Epiphany? Eureka is described as the skill counterpart of Epiphany; see references/epiphany-comparison-2026-09-15.md. Ask how much of that relationship the operator wants public.
4. Is there a repo for the page? This folder has no git repo yet. GameCult repos default to the GameCult GitHub org (see F:\Projects\CLAUDE.md, "GitHub Repo Creation"). Confirm the name and visibility before creating one.

Working rules that apply to you

• The global and GameCult doctrine (~/.claude/CLAUDE.md, F:\Projects\CLAUDE.md) apply: map first, keep it small, and no load-bearing JSON if the page grows a data layer.
• Don't copy transcripts or agent reports onto the page. Cite the committed evidence (the postmortem, the changelog, the maps).
• The mascot has a saved generation prompt. Any new image you generate gets its exact final prompt saved beside it, with the same base filename.