The God Questions — what governs this project

Date: 2026-08-29 · Owner: SOL-GODQUESTIONS · Status: proposal for CENA review Binding method: by reference only. Every link below is a pointer to a record that stays where it lives. Nothing in Great_Library_of_SISO/ is edited, copied, or vendored by this document. That is GQ-006 success criterion 3, applied to ourselves.

Read this first: the count is wrong everywhere

ECOSYSTEM.md:25 says the Great Library holds "9 God Questions". The dispatch that commissioned this document says nine. GQ-009's own watch_trigger_status field says "All 7 registry questions".

Re-derived from source on 2026-08-29:

ls Great_Library_of_SISO/registry/works/frontier-question-gq-*.json | wc -l
→ 8

The eight are GQ-001, 002, 004, 005, 006, 008, 009, 010. There is no GQ-003 record and no GQ-007 record in the registry. Three different published counts exist for one set of files and none of them is the observed number.

This is not a footnote. GQ-006's second success criterion is "Every number the organ publishes can be re-derived mechanically from a named source by a second program that did not produce it." The first time this document ran that check against our own front door, it failed. The correction belongs at the top rather than in an appendix, because the failure mode it demonstrates — a number that everyone repeats and nobody re-derives — is the exact one the binding below is meant to catch.

The inversion, stated precisely

Actionist has been treating the Great Library as an index to be listed in. The relationship runs the other way.

GQ-004 · Best Software Primitive has carried "state": "scoped" since 2026-08-04 and this evidence gap, verbatim from registry/works/frontier-question-gq-004.json:106:

"No primitive has been evaluated yet. This contract defines what an answer would have to satisfy; it does not report findings."

Its question (:71) is "For a named software primitive, which implementation is best for the stated context, and is it safe and valuable to reuse, vendor, or learn from?"

That is the capability shelf's question. It is the RVM's question. It is the foundry's question. Actionist is not a consumer of GQ-004 — Actionist is the engine that answers it, and GQ-004 is the standing specification Actionist has been building toward without citing.

But the inversion cuts both ways, and this is the part worth being honest about up front. Read GQ-004's answer_shape (:85) next to what the shelf currently holds:

And the qualification counts, re-derived from knowledge/capability-shelf/source-registry.jsonl immediately before writing this line:

total rows                    3,010
qualification_status QUALIFIED    0     (all 3,010 NOT_QUALIFIED)
admission_status ADMITTED         0     (all 3,010 NOT_ADMITTED)
foundry_stage stage_0_intake  2,819
rows with any client precedent    7

So the honest statement of the inversion is narrower than the one in ECOSYSTEM.md:65 ("Actionist is not a consumer of the Great Library. It is the engine that answers its hardest open question"). That sentence is written in the present tense and it is a claim about the future. Stated as a present fact:

Actionist could answer GQ-004 and has produced no evidence that would qualify. Zero qualified rows means zero GQ-004 findings. The engine is designed, partly built, and has never been run to completion on a single primitive.

Two things stop that from being merely a deflation.

First, the bar is already written and we did not write it. GQ-004's contract states exactly what the first completed run must produce — down to a field we do not have. We are not guessing at a standard, which is the usual reason a first attempt gets rejected.

**Second, and less obvious: a GQ-004 finding is a shortlist, so one qualified block would still not be one.** answer_shape requires "explicit reasons not to choose each candidate" — plural. That needs two or more candidates for the same primitive in the same named context. The two built blocks are different primitives (a data grid and a workspace/editor), so even qualifying both produces two decisions and zero findings. The gap between here and the first finding is not "finish qualification"; it is qualify a second donor of a capability we have already started, which is a queue decision nobody has had to make yet.

That is the practical content of the inversion, and it is worth more than the slogan.

The binding map

Eight questions. Four bind directly and govern work in flight. Four are unexamined against this project until now; two of those turn out to bind harder than expected.

Summary table

GQ-004 · Best Software Primitive

registry/works/frontier-question-gq-004.json · gls:work:85964f1c-2efe-48b0-8f71-dc32bb668c0c State scoped · lifecycle experimental · updated 2026-08-04

Binds to: knowledge/capability-shelf/source-registry.jsonl (3,010 rows), knowledge/frameworks/repository-value-matrix-v1.md, knowledge/capability-shelf/FOUNDRY-PROCESS.md (11 stages), blocks/teable-data-grid/, blocks/affine-workspace/.

What it has already decided, so we do not get to re-litigate it:

1. *"Every candidate carries an explicit reason NOT to choose it — a shortlist without stated

downsides is advocacy, not evaluation."* (success_criteria[0]) — The RVM produces a Diamond Score and a minimum_critical_dimension. A low dimension is not a reason not to choose; it is a number. This field does not exist in our schema and must be added.

1. *"Interface fit is assessed against the named context's actual call sites, not against the

candidate's own documentation."* (success_criteria[2]) — RVM's integration_seams dimension is measured from the donor's typed APIs and extension points. That is the candidate's side of the seam. GQ-004 requires the host's side.

1. "Adoption cost includes the cost of removing the candidate later, not only of adding it."

(success_criteria[3]) — adaptation_burden counts surgeries to get in. Nothing counts the cost of getting out. R4's iframe-debt argument is exactly a removal-cost argument that the scoring framework cannot express.

1. *"The rights gate is applied and recorded per candidate before any code is vendored or

reused, not after."* (success_criteria[1]) — Consistent with R1 (licence is not a selection input) because a gate at admission is not a weight at selection. GQ-004 and R1 agree; they are describing different stages.

Its falsifier that will fire on us first (falsifiers[2]): "The 'best' choice changes when the context is restated slightly, indicating the ranking tracked the phrasing rather than the requirement." The RVM scores per (source, shape, recipe, block, workflow). Restating a recipe slightly is cheap. Nobody has checked whether the ranking is stable under it.

What it explicitly leaves open: which primitives to evaluate, in what order, and what counts as a "named context". GQ-004 defines the bar, not the queue. That is ours to choose.

Where our evidence flows back: a qualified block becomes a GQ-004 finding. See G3 flow-back below. Requires ≥1 row at QUALIFIED. Currently 0.

GQ-006 · The Information Organ

registry/works/frontier-question-gq-006.json · gls:work:2144e900-4886-461e-92f1-49571f2e9f74 State partial · lifecycle active · updated 2026-08-04

Binds to: site/system-map/data/state.json, INTENT.md as front door, every gate and verifier we run, and this document.

What it has already decided:

1. *"Evidence attaches to a question by reference — byte ranges into sources — rather than by

copying corpora into the registry."* (success_criteria[2]) — This is the constraint that makes the whole flow-back design tractable. The 6.7MB source-registry.jsonl never moves. Already adopted in ECOSYSTEM.md's standing constraints.

1. *"Freshness is derived from observable state, never asserted by whoever last touched the

record."* (success_criteria[0]) — Our observed_at: "2026-08-28" fields are assertions by the last writer. They are exactly what this criterion prohibits.

1. *"A delivery path that stops working becomes visible without anyone remembering to check

it."* (success_criteria[3])

Its falsifiers, three of which have already fired in the Library (state_note: "three of the four falsifiers have already fired at least once") — these are recorded failure modes, not hypotheticals:

• Falsifier 1 fired: *"a hyphen/underscore path typo reported 0 source inventories when 6

existed, because the builder and the checker shared the wrong path." This is why R5's strengthened form demands a second program that did not produce the number*. A verifier sharing code with its generator verifies nothing.

• Falsifier 2 fired: "hardening a verifier marked all 10 claims stale at once" — a gate

that cries wolf is worse than no gate. Adopted in INTENT.md as the per-lane staleness rule.

• Falsifier 4 fired: *"the organ reports a pipeline as healthy while that pipeline has never

actually run end to end... outbox/sent/ was empty — zero escalations ever delivered."*

That last one is us, right now, in a different costume. The foundry is an 11-stage pipeline. Stages 0-6 have run. Stages 8-10 (qualification, admission, production learning) have never run to completion on any source: 0 QUALIFIED, 0 ADMITTED. The framework register reports 24 frameworks at machine_readable and specified maturity — which is honest — but registry-summary.json describes a 3,010-row shelf in the language of an operating system. GQ-006 falsifier 4 is the check that catches this, and it fires.

What it leaves open: the query primitives and push mechanism. Its own evidence gap says "Push freshness into agent workflows is named in the question but unbuilt: the organ currently publishes and waits to be read rather than pushing into a loop." Our state record has the same shape and the same gap.

GQ-002 · 10× the Agent Layer

registry/works/frontier-question-gq-002.json · State partial · lifecycle active

GQ-008 · Model Routing Evidence

registry/works/frontier-question-gq-008.json · State partial · lifecycle active

These two bind as a pair, because they were downgraded on the same day for the same reason and they teach the same two lessons.

Lesson one — the zero-falsifier downgrade. Both carry a state_note recording a downgrade from answered to partial on 2026-08-04. GQ-008's reads: "Downgraded from 'answered' on 2026-08-04, for the same reason as GQ-002: marked answered while carrying zero falsifiers." GQ-002's adds the sharper line: "Writing the falsifiers is what exposed the gap."

This is R6's precedent and it is already in INTENT.md. What is not yet acted on: 0 of our 24 frameworks carry success_criteria. All 24 carry falsifiers (2-3 each — verified by reading frameworks[].falsifiers across framework-register.json). We adopted half the contract shape. G2 proposes the other half.

Lesson two — the measurement-artifact trap, which is the dangerous one. GQ-008's third falsifier:

"A measured cost gap between providers disappears once gateway configuration is corrected, meaning the matrix ranked a misconfiguration rather than a model. This has already partially fired: MiniMax reads 0 cached tokens through Bifrost and 2,944 of 3,033 through the local proxy."

And its fourth success criterion: "The matrix distinguishes provider capability from gateway configuration, so a defect in the path is not recorded as a property of the model."

The RVM has no equivalent criterion, and it needs one more urgently than GQ-008 did. Full analysis in G4.

GQ-001 · The Agent Workspace

registry/works/frontier-question-gq-001.json · State partial · lifecycle active Unexamined until now. It binds strongly.

Binds to: INTENT.md's ownership table (§3), the domain/contract seam model, the block record structure, AGENTS.md, CLAUDE.md.

What it decides for us:

1. *"Each principle is graded by the evidence behind it and states whether it is a universal

finding or a SISO preference."* (success_criteria[0]) — INTENT.md's six standing rules are stated as universals. R4 ("default reuse shape is fork and integrate natively") is a preference derived from one observation about AFFiNE's iframe. It may well be right; it is not graded, and GQ-001 requires that distinction per principle.

1. *"Rejected patterns are recorded with the observation that rejected them, so a later reader

can tell a tried-and-failed pattern from an untried one."* (success_criteria[1]) — INTENT does this well. Every rule carries its counter-example with a date. This is the one criterion we already pass cleanly.

1. *"A principle that has never been violated in practice is marked as untested rather than as

confirmed."* (success_criteria[3])

Its first assumption, QA-GQ001-BOUNDARIES (from the program block): "A complete agent workspace can integrate its organs through typed interfaces without collapsing their independent ownership boundaries." Falsifier: "Representative workspace tasks require one organ to own another organ's durable state, evidence adjudication, and runtime authority to remain coherent."

INTENT.md §3 is a bet on exactly that assumption — "No domain owns another domain's truth. Agents own domains; contracts own the seams." Actionist is a second workspace testing GQ-001's central assumption, and GQ-001's headline evidence gap is that it has only one: "Evidence comes almost entirely from a single workspace operated by a single agent, so 'universal' cannot yet be distinguished from 'true here'."

That is a genuine, cheap, unclaimed contribution. We are the second data point GQ-001 says it needs, and we cost nothing extra to observe.

GQ-009 · The God Questions Observatory

registry/works/frontier-question-gq-009.json · State researching · lifecycle active Unexamined until now. It binds reciprocally — this document is an instance of it.

What it decides: "A question can justify a bounded action candidate and later receive a verified learning return without the Library authorizing or executing the action." (success_criteria[2]). This is the governance shape of the entire flow-back design: the Library never authorizes Actionist work; it receives evidence returns. G3 is built to that constraint.

Its evidence gap names us exactly:

"No standing agent operates the loop. Every piece of infrastructure built to date is supply; nothing continuously reads a question, finds bearing evidence, and updates an answer. This is the gap between a warehouse and an organ."

Actionist is a candidate standing agent for GQ-004. Not a metaphor — that is the literal missing component, and GQ-004 is the question we are already built to read.

And its fifth watch trigger is the one nobody can fire alone:

"A standing agent proposes a change to the Library or Foundry and it survives independent review — the first evidence that self-improvement is real rather than asserted."

Its own recorded status: "CANNOT FIRE BY CONSTRUCTION — the standing agent is the proposer and the trigger requires INDEPENDENT review. This is the only one genuinely blocked on a second party."

Actionist is a second party. A proposal originating in this project, reviewed by the Library's maintainer, is structurally capable of firing GQ-009's watch trigger 5 — the trigger its own record says cannot fire by construction. That is the single highest-leverage move available in this whole engagement, and it costs one well-formed proposal.

Its hardest constraint on us (falsifiers[3]): "Maintaining the question contract costs more attention than the repeated research and decision errors it prevents." G2 must not produce ceremony. Every field proposed there has to earn its keep.

GQ-005 · Where the Field Is Moving

registry/works/frontier-question-gq-005.json · State scoped · lifecycle experimental Unexamined until now. Binds moderately — and warns us off a trap.

Binds to: the discovery estate (Mac Mini identity.sqlite, 1,358,200 repo_card rows), the industry discovery scrapes in registry-summary.json, any future "what's trending" input to shelf ranking.

What it decides: "Every predicted movement carries a date and a scoring rule fixed in advance, so it can later be marked right or wrong without renegotiation." If shelf ranking ever consumes a momentum signal, it inherits this bar.

The warning it carries for us, from its own evidence gaps, measured 2026-08-04: the top category by star gain (agent-extension-pack, 2.06/repo against a 0.045 baseline) "draws 92.1% of its gain from ONE repository, and only 29 repos in the whole corpus moved at all across the three days."

A momentum signal that is 92% one repository is a signal about one repository. Our priority_score field ranks 3,010 rows and no one has checked its concentration. Same failure shape, unexamined.

Second warning, from falsifiers[2]: "The observed deltas are dominated by measurement artifacts — a source changing its reporting rather than the field changing." R2's counter- example (an agent re-scraped 21st.dev for 3,507 components already on disk, 37% redundant fetch rate) is a discovery-estate failure of the same family: measuring our own collection activity and mistaking it for signal about the world.

GQ-010 · The People Graph

registry/works/frontier-question-gq-010.json · State scoped · lifecycle active updated 2026-08-06 — the most recently revised of the eight. No direct binding. And it is the most useful question on this list for us.

The People Graph is not our domain and nothing in Actionist touches it. But GQ-010's question is, structurally, our question with the nouns swapped:

"How do we increase the People Graph's research and decision value by 100× without treating row count as value or weakening identity quality, provenance, rights, privacy, reproducibility, and reversibility?"

Substitute "capability shelf" for "People Graph" and that is R3 — breadth is not progress — stated as a research contract with measurable criteria, four weeks before an Actionist lane ran a breadth wave from 191 to 3,010 rows and rated zero of them.

GQ-010 already ran our experiment and pre-registered the result. Its first falsifier:

"Source expansion increases rows or edges by at least tenfold but produces no measured improvement in the five decision-use cases."

Actionist: 191 → 3,010 rows is a 15.8× expansion. Measured improvement in decisions: none recorded. Qualified rows: 0. That falsifier, applied to our shelf, fires. It is the cleanest external confirmation available that R3 was right, and it comes from a question that was not written about us.

What we should steal from it wholesale. GQ-010 separates value into six named dimensions where our shelf has one (priority_score): breadth, resolution quality, research depth, decision value, rights/privacy, and reproducibility/cost. Its fourth criterion is the one our framework register has no analogue for:

"Decision value: in at least five pre-registered operator or research decisions, the graph changes the selected action or reduces evidence-retrieval time by at least 80% against a measured baseline; additional rows alone cannot satisfy this criterion."

Pre-registered decisions, a measured baseline, a numeric threshold, and an explicit statement that volume cannot substitute. That is the shape of the success criterion the RVM is missing, already written, already dated, by someone who was not thinking about us. G2 adopts it.

What binds to nothing

Honest negative result, since a binding map that finds everything relevant is not a map.

• GQ-005 → composition/host/data domains: nothing. Momentum has no bearing on how a block

binds to a host.

• GQ-010 → any Actionist domain: nothing directly. Its value is analogical, and the analogy

is load-bearing enough to justify the section above, but no Actionist artifact should cite GQ-010 as governing evidence.

• GQ-003, GQ-007: do not exist. If a future document cites either, it is citing the phantom

count corrected at the top of this file.

The flow-back path

Designed here, not executed. Detail and open questions in research/workstreams/2026-08-29-god-questions/G3-FLOW-BACK.md.

Actionist                                          Great Library
─────────                                          ─────────────
shelf row (3,010)          ─── no path ───▶        (nothing; discovery is not evidence)
       │
       │ foundry stages 1-7
       ▼
qualified block (0 today)  ─── pointer ──▶         Source Inventory campaign unit
       │                                            · promotion.stage
       │                                            · verification_evidence_refs  ← by reference
       │                                            · next_gate / blockers
       ▼
admitted block (0 today)   ─── pointer ──▶         GQ-004 finding
                                                    · named context
                                                    · reason-not-to-choose per candidate
                                                    · measured adoption + removal cost

Three properties this design must preserve, each traceable to a named constraint:

1. Zero forks, zero vendoring — the Library catalogues by URL + pinned revision. Our

6.7MB source-registry.jsonl never enters it. (GQ-006 success_criteria[2]; CONTRIBUTING "Capability discovery and promotion".)

1. Registration shape is a single Work + a dated Source Inventory campaign, not a Section.

A Section requires a build.mjs edit, closed-enum schema changes, a Snapshot, and an Event. CENA established this; the schema confirms it — source-inventory.schema.json has a campaign object with exactly the ten-stage lifecycle enum INTENT.md §4 already adopted verbatim.

1. Discovery must not create a Work automatically (CONTRIBUTING, verbatim). The 3,010 rows

are discovery. They have no path into the Library and should not acquire one. Only a reviewed, qualified candidate promotes.

The gating fact: every arrow below the first is currently empty. The path can be designed now and cannot be exercised until one row reaches QUALIFIED.

G4 — where our frameworks can fail the same way

GQ-008's falsifier 3 fired: a measured MiniMax cost disadvantage turned out to be a gateway defect, not a model property. The matrix ranked a misconfiguration and reported it as a finding.

Full audit in research/workstreams/2026-08-29-god-questions/G4-MEASUREMENT-ARTIFACTS.md. Headline: VSCP-1 is the RVM's Bifrost. ui_quality carries weight 0.22 for surface/module rows and 0.24 for pattern rows — the single heaviest dimension in both profiles — and it is scored entirely from screenshots taken by our own harness under our pinned fonts, locale, device scale factor and reduced-motion preference. A donor that renders badly under our capture configuration is indistinguishable, in the resulting score, from a donor that renders badly. The RVM has no criterion separating donor property from harness configuration. GQ-008 has exactly that criterion (success_criteria[3]) and needed it.

Standing obligations this binding creates

Refresh policy for this document

Re-read the eight contracts and re-derive the counts when any of these occur:

• a God Question changes state or updated_at (all eight are currently 2026-08-04 except

GQ-010 at 2026-08-06);

• the first shelf row reaches QUALIFIED, which makes obligation 9 live;
• a ninth God Question is added, or GQ-003/GQ-007 appear;
• any count in this document is quoted somewhere else without re-derivation — that is the

failure mode described at the top, and it is a trigger, not a nuisance.

Every number here was re-derived from source on 2026-08-29 immediately before writing. Nothing is quoted from another document, including the ones that were wrong.