Eureka/references/epiphany-comparison-2026-09-15.md

Epiphany pipeline model (source-grounded) vs. the Eureka skill

Written 2026-09-15 by a read-only Modeling/Eyes pass. Epiphany HEAD is 71324718 (2026-09-14), branch as checked out, worktree clean. The voidbot MCP was not used; everything below comes from a direct Grep/Read of F:\Projects\Epiphany. .epiphany-run/ was ignored as stale.

Faculty names are as of this date. The skill's Mind Steward was renamed Life on 2026-09-25 (see the changelog); rows below that describe the skill's steward mean that faculty. Epiphany's own mind_steward doctrine and its Continuity (Reorientation) worker are described as they stood.

Resolved (Mind Steward, 2026-09-15). While this pass ran, the skill was renamed from epiphany-pipeline to eureka, and its frontmatter and changelog were then updated, so the drift this pass saw no longer exists.

Not covered by this pass: MCP. Epiphany has an rmcp client only (epiphany-tool-mcp-runtime/Cargo.toml:18, client feature). It has no MCP server.

---

0. What Epiphany actually is (one paragraph)

Epiphany is a Rust workspace. It is a library (epiphany-core, 55,650 lines of tracked .rs) plus a release bundle that owns 9 [[bin]] targets (Cargo.toml). It schedules LLM worker passes (Modeling, Research/Eyes, Imagination, Mind plan-review, Verification/Soul, Reorientation, Persona) through a pure priority projection over typed CultCache state. Each pass returns a closed typed decision, and a per-family admission owner commits it to the Mind store with batch compare-and-swap. The "pipeline" is state-driven, not conversational: nobody writes a brief. The coordinator derives the next action from which typed obligation exists.

---

1. Faculties and organs

Faculty	Embodiment in source	Status
Self	resident_self.rs (3,105 lines): typed pressure, grant, lease, cooldown (cooldown_seconds :284), retry, settlement. surfaces/coordinator_decision.rs:7-103 recommend_coordinator_action, a pure priority function. Binaries epiphany-mvp-coordinator, epiphany-swarm.	Implemented as deterministic code, not an agent. map.yaml calls it "partial".
Modeling	Worker role modeling; instructions at current_work.rs:2885 (Body), :3063 (proposal), :3189 (frontier-verdict). Output schemas in epiphany-openai-runtime/src/lib.rs:937-968. Writes keyed RepoModel docs (repo_model_documents.rs).	Implemented.
Eyes	Worker role research, current_work.rs:1610. Immutable GitHub source identity. Evidence enters Mind as keyed evidence/observation documents.	Implemented. Launched only for an explicit external-evidence obligation (algorithmic map :856).
Imagination	Worker role imagination, current_work.rs:1170 (frontier planning candidate), :2738 (Persona-feedback consideration); imagination_consideration.rs. It proposes, and the next organ can only be Eyes or Imagination (6e600e8d, algorithmic map :876-883).	Implemented.
Mind (admission)	reasoning_context.rs commit owner: atomic batch CAS plus EpiphanyMindCommitReceipt. A Mind plan-review worker (current_work.rs:1327) adopts, refuses, or holds an Imagination candidate. anatomy.md:49-52 says this is a stateless admission procedure, not a sub-agent.	Implemented.
Soul	Worker role verification, current_work.rs:1466 ("Audit only the exact typed Hands consequence..."). soul_gateway.rs:3-58: SoulVerdictReceipt (verdict is a String) and RepoFrontierVerificationRequest, which binds the Hands intent, patch, command and commit receipt IDs.	Implemented as a model pass over sealed receipts.
Hands	hands_gateway.rs: HandsActionIntent, and Patch/Command/Commit receipts. Writers are runtime_spine.rs:7386,7422,7456. Every caller is in current_work.rs tests (lines 4383-4471, inside #[cfg(test)] from :3309). ContinueImplementation is emitted by coordinator_decision.rs:92-96, but Resident Self excludes it from the launchable set (resident_self.rs:974-993, which returns None). No code path in core, OpenAI runtime or tool-mcp runtime produces Hands receipts. The writers are still re-exported as public crate API (lib.rs:230-232), even though their only callers are tests.	Types and verification input only; no live actuator in the current tree (verified by grep). map.yaml:275 still says "hands: proven for bounded consequences" and cites historical c011-era runs.
Persona	persona_*.rs (6 files) plus epiphany-persona-service (Projector, Persona, Interpreter stages) and epiphany-persona-discord-permit.	Implemented, "partial" per map.yaml.
Continuity	reorientation_work.rs worker (:475, "Decide resume or regather"). EpiphanyCoordinatorDeathRecovery, archived session tombstones.	Implemented as a worker plus runtime tombstones.
Mind Steward	No code. It exists only as agent doctrine: AGENTS.md:81-103 requires a Codex/Claude mind_steward sub-agent over state/map.yaml, the handoff, and the algorithmic map.	Docs/agent-procedure only.
Nervous system	Heartbeat store state/agent-heartbeats.msgpack, Resident Self grants/cooldown, runtime-spine jobs. AGENTS.md:197-201 heartbeat physiology rules.	"partial" (map.yaml organ_readiness).
Substrate Gate	substrate_gate.rs: grants/refusals, and HandsActionIntent.substrate_gate_grant_receipt_id.	Implemented (types).
Operator	Not an organ. See §5.	—

The docs-only faculty surfaces are the Mind Steward and the dependency matrix. notes/organ-dependency-contracts.md still says "every embodied sub-agent depends on all the others", but the code matrix was deleted in 2aeab7a9.

2. Routing and scheduling

• The priority projection is pure (coordinator_decision.rs:7-103; algorithmic map :1128-1143), in this order:

  1. Mind missing, then Reorientation, then operator regather;
  2. frontier Planning stages (Imagination, then Mind review, then commit);
  3. proposal, Body, and frontier-verdict Modeling;
  4. Verification;
  5. Research;
  6. Imagination considerations;
  7. Hands-ready;
  8. otherwise await a proposal.

  "No default Modeling job, latest lane, accepted-at comparison, runtime event, or generic interrupt can manufacture work" (:1142).

• Input: current_work.rs projects one pulled CultCache snapshot into family obligations. Each family carries an EpiphanyAgentPassAttemptProjection (Launch/Wait/Review) (:1183-1197). A failed attempt changes the projection digest, so Resident Self grants exactly one fresh pressure. Replaying an unchanged failure is idempotent.

• Launch: Resident Self converts the action into a grant (resident_self_current_work_action, :965-1001). The coordinator launches a worker process. The sealed typed worker launch document owns family identity (edb5c3a3).

• Briefing: the instruction is a one-sentence hardcoded string per family (current_work.rs:1170,1327,1466,1610,2738,2885,3063,3189; reorientation_work.rs:475). The OpenAI runtime appends the sealed typed projection (dynamic_prompt_context), a family output contract text, and tool/evidence mandates (epiphany-openai-runtime/src/lib.rs:326-339, 922-935). There are no free-form briefs. Context is the sealed EpiphanyReasoningBasis (algorithmic map :1029-1035). The legacy prompt-TOML specialist prompts exist only in .epiphany-run (stale).

• Tracked residue: epiphany-state-model/src/prompts/*.md (2 files) is still tracked although the package was deleted in 07529fb5.

3. Verification

• Chain: Hands receipts, then a RepoFrontierVerificationRequest (binds the exact receipt IDs, frontier item hash, and model projection digest), then a Verification worker pass, then SoulVerdictReceipt, then a derived frontier Modeling request (runtime_spine.rs:7154-7215).
• Gate vocabulary: "pass" resolves to Resolved; "needs-review"|"needs-evidence"|"fail" map to Blocked; anything else is an error (runtime_spine.rs:7197-7200). This is a string match, not an enum. The verdict must byte-match the worker result's verdict, summary, risks and evidence (:7207-7214).
• Anti-self-report:
  • The Verification projection is sealed. Since 24023265 it is the sole Hands receipt input, so Soul cannot reread mutable runtime state (map.yaml:46-65).
  • Provider success cannot launder malformed output: native decoding and family admission are final (553f79d9).
  • A structurally valid but semantically refused result writes a typed admission-refusal and schedules a retry (algorithmic map :1390-1394).
  • a01c6842 deleted the coordinator's "same-function always-approved Hands review".
• What is absent:
  • No mutation testing anywhere (grep for mutants found nothing).
  • No requirement that Verification runs a different model or worker identity than Hands (grep found no distinctness check). Independence comes from separate job/launch/basis, not from an adversarial brief.
  • No CONFIRMED/PLAUSIBLE grading.
  • The Soul brief is one sentence, "Audit only the exact typed Hands consequence", so it is an audit, not a falsification mandate.
  • No INTEGRATE / INTEGRATE_WITH_FOLLOWUPS verdicts.
• Repo-development verification (humans/Codex building Epiphany itself):
  • Focused cargo tests per cut, recorded as counts in map.yaml (for example "Core 152/152").
  • Idunn CI gate on Yggdrasil with SHA-256 test receipts (algorithmic map :1414-1418).
  • Live "capstones" (Ox10..Ox17, Capstone 17) as end-to-end falsification. Each failure is sealed as evidence and never resumed (:1440-1502).
• Lessons doc: notes/faculty-workflow-lessons-2026-09-04.md records that a Soul writing tests produced the best findings (:49-55), and proposes a typed Verdict {HOLDS|FALSIFIED|UNPROVEN} (:151-152). That is not implemented.

4. Plans and maps

• Durable executive map: state/map.yaml, 292 lines, AGENTS.md:60-80. It holds objective, invariants, current_status (one cut entry per landed SHA with net lines and verification), active_subgoals, organ_readiness, forbidden_actions, and memory_policy. It plays the role of the skill's status header, but it is YAML, not CultCache.
• Mechanism map: notes/epiphany-current-algorithmic-map.md (1,526 lines) carries owner/inputs/outputs/invariant tables (:841-862) and per-SHA cut narratives. Campaign plan: notes/epiphany-fork-implementation-plan.md (945). Handoff: notes/fresh-workspace-handoff.md (2,343 lines).
• Evidence ledger: state/ledgers.msgpack via state_ledger.rs. It is a single-key CultCache entry holding Vec<branches> and Vec<evidence{ts,type,status,note}>, and each append rewrites the entry (:52-64). The CLI is epiphany-state add-evidence.
• In-runtime plans: Imagination frontier candidates, then a Mind plan decision (adopt/refuse/hold) as keyed RepoFrontierRoute, plus repository_scope and authorized_paths ceilings (5b799b12). These are typed and CAS'd.
• No per-migration target doc or cut map in the skill's sense. Cuts are recorded after the fact in map.yaml and the algorithmic map, not specified ahead of time to file:line.
• Operator rulings: there is no typed ruling artifact. In the runtime, the operator objective enters via coordinator_objective_intake.rs:45-78 (UserObjectiveIntake, which refuses to replace an existing objective). In repo development, rulings live as prose invariants in map.yaml and AGENTS.md. The lessons doc (:153-158) proposes "Rulings as a first-class artifact". That is not implemented.

5. Operator interaction

• RegatherManually is the only operator-question action. It fires when an accepted continuity decision sets operator_regather_required (coordinator_decision.rs:27-32, current_work.rs:533-539).
• PrepareCheckpoint fires when Mind is missing and requires an operator objective (:10-15).
• The consent/approval surface is the canonical swarm brake plus deployment brake, and Idunn operator grants that open and expire (map.yaml:40-41). forbidden_actions (map.yaml) forbids speculative deployment grants.
• "Humans talk to Face" (AGENTS.md:189-192). The Aquarium UI moved to a sibling repo (apps/README.md) and is not in this body. The former Discord operator bridge is deleted (algorithmic map :1507).
• There is no bundled fork/question document with a recommendation, and no ruling write-back path besides objective intake and brake/grant transactions.

6. Memory

• Runtime Mind: keyed CultCache documents in the runtime .cc store, with singleton objective/focus/mode/RepoModel binding and keyed evidence, invariants, checkpoints, Persona memories and more (algorithmic map :981-1001). Merge law :1105-1114. There is no aggregate head and no semantic cache (856648de deleted Qdrant/Ollama/Postgres projection).
• Persona memory: admit_persona_state_notes writes keyed Mind documents (persona_conversation.rs:1018). The repo publishes schemas/cultnet/gamecult.persona_state.v0.schema.json but does not consume it (consistent with F:\Projects\CLAUDE.md).
• Repo-development memory: map.yaml, handoff, and ledger, maintained by the agent-side mind_steward procedure (AGENTS.md:81-103, 144-151, 216-221). No steward organ exists in code, and nothing triggers the steward structurally. The lessons doc :165-168 proposes boundary triggers; they are not implemented.
• epiphany-agent-memory-store was removed in 387afe49 (verified: that commit deletes it along with about 28 other bins).

7. Observability

• Eve surface: only Model Atlas has one (atlas/eve_surface.rs, 1,597 lines). Coordinator status uses string operator_status() values (resident_self.rs:582-593) and the epiphany-state status CLI. perfect-machine-audit-roadmap.md:53 lists "Eve/CultUI operator interface" as open.
• Audit: epiphany-model-runtime audit-decision / list-decisions reconstruct any terminal pass from typed records without transcripts (algorithmic map :1403-1409). CultMesh projects reasoning basis, decision context, and commit receipt read-only (map.yaml:118).
• Budgets:
  • Time: one outer worker budget; Persona --turn-timeout-seconds, default 600s (algorithmic map :1041-1047).
  • Tokens: recorded on receipts only (input_tokens, output_tokens, reasoning_output_tokens, cached_input_tokens, from epiphany-model-adapter/src/native.rs:153-167). No token-budget enforcement was found.
  • reasoning_effort is an optional per-request field.
  • No model-selection-by-criticality policy.
• Build/verify footprint: each cut in map.yaml records verification output size (for example "Verification generated 6.349 GiB across 7,376 files"; map.yaml:79-80).

8. Failure handling

• Typed failure terminality: EpiphanyModelPassFailure, bound to the sealed context. One shared failure owner covers role, reorient, and Persona (bb823c54, algorithmic map :1082-1097).
• Retry: a changed attempt digest yields one fresh grant, and proposal attempt ordinals are contiguous (:1191-1197). Admission refusal writes a typed refusal and a retry with prior refusals in context (:1390-1396).
• Crash: EpiphanyCoordinatorDeathRecovery competes with success/failure receipts through one EpiphanyCoordinatorRunTerminality identity (:1220-1231). Archived session tombstones prevent identity resurrection (:941-951).
• Compaction (agent-side): AGENTS.md:153-180 re-entry rite.
• Schema epochs: hard cuts refuse old writable stores, with no migrator or dual reader (for example Runtime v45 / Mind v11, map.yaml:76).
• Wrong organ output: refusal documents; no Mind write; the run is sealed and never resumed (Ox10/12/13/15/16 narratives, algorithmic map :1440-1502).
• Scars recorded in-repo:
  • capstone narratives;
  • faculty-workflow-lessons (integration is a gated sequence, :116-125; push gated on the result line, not the launch, :127-135);
  • AGENTS.md build-economy rules (:250-277).

9. Scale and structure

• Workspace members (Cargo.toml): 5 crates plus the root release bundle, with 9 [[bin]]: release, state, repository-body, swarm, persona-service, persona-discord-permit, mvp-coordinator, model-runtime, tool-mcp-runtime. That is down from about 75 core bins; 387afe49 alone removed about 28.
• Tracked source: 77 .rs files, 65,113 lines. The giants are runtime_spine.rs 8,899; current_work.rs 5,714; openai-runtime/lib.rs 3,350; resident_self.rs 3,105; reasoning_context.rs 2,715.
• Liability observations:
  • The repo-local target/ is 39 GiB across 113,504 files. AGENTS.md:117-123 says builds should use the shared C:\Users\Meta\.cargo-target-codex target, so this root is exactly the kind of unowned output root the build-economy rules forbid (AGENTS.md:273-277).
  • The git-ignored .epiphany-run/ sandbox is 105 GiB. It is stale, not a tool (per F:\Projects\CLAUDE.md), but it is the single largest disk liability in the repository. Together with target/, about 144 GiB of output sits beside 65k lines of source.
  • The deployed package still lists "26 binaries plus witness" (algorithmic map :1418, historical d2ca6630).
  • Untracked empty crate directories remain: epiphany-openai-auth-spine, epiphany-openai-codex-spine, epiphany-release-construction, epiphany-self-policy (0 files each).
  • Tracked orphan prompts remain in epiphany-state-model/src/prompts/.
  • The .gitignore still names deleted crates (epiphany-openai-adapter, epiphany-codex-bridge, apps/epiphany-gui).
  • Note bulk: notes/ totals about 10.6k lines. The handoff alone is 2,343 lines and repeats cut narratives also in map.yaml and the algorithmic map. That is triple-recorded cut history.
  • The Model Atlas code (about 10k lines under atlas/) is "code-complete", but Gate 1 is paused and it is operationally unaccepted.
  • The ledger uses .msgpack single-file stores, not .cc.

10. Known gaps and stale docs

• Hands claimed "proven" (map.yaml:275) while current source has no live receipt producer and Resident Self won't launch ContinueImplementation (§1). Verified by grep.
• No accepted end-to-end run on the current package. The objective (map.yaml:7) is still diagnosing the Aug 25 Luna connector failure. Capstone 17 and Atlas Gate 1 are open, and production is inactive and braked.
• notes/epiphany-anatomy.md:61-62 still expects "typed continuity packets, recovery receipts, compaction checkpoints, stale-turn repairs". Those contracts were deleted in ca2d2cf2 and fed4b857 (map.yaml:183).
• notes/organ-dependency-contracts.md describes the dependency matrix deleted in 2aeab7a9.
• notes/perfect-machine-audit-roadmap.md (dated 2026-08-12) cites the deleted "coordinator state transaction" and old test counts (684/684). It is still listed as "Active Design Reference" in notes/README.md:28.
• The README.md "What She Is" list names "heartbeat, sleep, memory" machinery, and the Continuity/Substrate Gate receipt families are partly deleted (inferred mismatch; not line-audited).
• docs/epiphany_body_whitepaper is explicitly marked historical (README:220).
• faculty-workflow-lessons proposes 7 typed artifacts: faculty artifact schemas, rulings, landed-names digest, owner-scoped edit queues, steward triggers, Self budget, swarm visibility. None exist in code (grep for verdict/ruling/digest types found only SoulVerdictReceipt with a string verdict).
• The skill frontmatter name mismatches its directory (see header).

---

Comparison table: Eureka skill vs. Epiphany

Dimension	Skill (Eureka)	Epiphany	Evidence
Self / router	Root LLM agent routes by judgment, keeps and commits maps, talks to operator.	Deterministic pure priority function over typed obligations; Resident Self grants/leases/cooldowns. No LLM router.	Skill SKILL.md:24-27, 210-232. Epiphany coordinator_decision.rs:7-103, resident_self.rs:965-1001.
Unit of work	A "cut" from a hand-authored cut map (repo-scale refactor).	A typed family obligation (Body Modeling, frontier item, Verification request...) in a RepoModel graph.	cut-map.md:31-56. Algorithmic map :1128-1197.
Planning artifact	Target doc (ends) and cut map (means), specified ahead to file:line with authority maps and a subtraction ledger.	Imagination candidate, then Mind plan-review adopt/refuse/hold, then keyed RepoFrontierRoute with repository_scope/authorized_paths. Repo dev records cuts after landing in map.yaml.	SKILL.md:60-112. current_work.rs:1170,1327; 5b799b12; map.yaml:45-188.
Briefs	Rich templates: invariants, promises to falsify, standing rulings, git rules.	One-sentence hardcoded instruction per family, plus sealed typed projection, output schema text, and tool mandates.	briefs.md:16-173. current_work.rs:1466 etc.; openai-runtime/lib.rs:326-339,922-935.
Hands	Subagent commits and pushes small cuts, runs its own mutations, reports structural delta.	Typed Intent/Patch/Command/Commit receipts with path ceilings; no live producer in current source; ContinueImplementation not launchable.	SKILL.md:126-147. hands_gateway.rs; writers only in current_work.rs tests (:3309+); resident_self.rs:974-993.
Soul independence	Different agent; adversarial brief listing Hands' promises; reruns builds; designs non-revert mutations; probes at the failing layer.	Separate worker/job over a sealed receipt projection; audit-framed one-line brief; no mutation testing; no distinct-model/worker check.	SKILL.md:149-167. current_work.rs:1466; 24023265; grep (no mutants, no distinctness).
Verdict shape	Findings CONFIRMED/PLAUSIBLE, with file:line, scenario and severity, plus promises that held.	SoulVerdictReceipt.verdict: String; "pass" becomes Resolved, "needs-review/needs-evidence/fail" become Blocked; byte-equality with the worker result.	SKILL.md:163-164. soul_gateway.rs:13; runtime_spine.rs:7197-7214.
Self-report protection	Structural: a different agent reruns everything. Rule "never claim a result before its notification".	Structural: sealed basis/context; native decode plus family admission final; provider success can't launder; typed refusal; always-approved review deleted.	SKILL.md:232. Algorithmic map :1049-1072, :1390-1396; a01c6842.
Soul loop cap	Explicitly uncapped on foundation cuts; track hit rate.	Retry is per failed attempt (one fresh grant per changed digest); no concept of repeated Soul passes over the same consequence.	SKILL.md:178-185. Algorithmic map :1191-1197.
Operator rulings	Bundled questions with options and a recommendation; dated rulings in the map; superseded text marked history.	No typed ruling artifact. Objective intake (refuses replacement); RegatherManually; brakes and expiring Idunn grants.	SKILL.md:114-124; cut-map.md:22-25. coordinator_objective_intake.rs:45-78; coordinator_decision.rs:27-32; map.yaml:40-41. Lessons doc :153-158 (proposed).
Consent / approval gates	Operator rules forks; git push per cut without further gating.	Swarm brake, deployment brake, Substrate Gate grants, Bifrost signed receipts for public consequence.	map.yaml:33-42; substrate_gate.rs; HandsActionIntent.substrate_gate_grant_receipt_id.
Memory / steward	Mind Steward subagent at phase boundaries; falsifies one claim; names mutations.	Runtime: keyed CultCache Mind with CAS merge law. Repo dev: same steward procedure as agent doctrine only; no steward organ or structural trigger.	briefs.md:142-158. Algorithmic map :981-1124; AGENTS.md:81-103.
Durable state format	Markdown docs committed in the changed repo.	Runtime .cc typed documents (CultCache); repo dev uses YAML map, markdown notes, and a msgpack ledger.	cut-map.md:3-10. reasoning_context.rs; state/map.yaml; state_ledger.rs:14-64.
Audit / provenance	Commit SHAs, Soul reports, postmortem from evidence.	Transcript-free audit-decision / list-decisions from sealed basis, context and commit receipt; SHA-256 CI test receipts.	postmortem-template.md. Algorithmic map :1403-1418.
Observability	Status header in the map; Self relays summaries.	map.yaml current_status; epiphany-state status; string operator statuses; Eve only for Atlas; operator Eve UI open.	cut-map.md:12-29. resident_self.rs:582-593; atlas/eve_surface.rs; audit-roadmap.md:53.
Budgets	Model choice by criticality; subtraction budget as pressure.	One outer time budget per pass; tokens recorded, not enforced; per-cut net-lines and verification-GiB recorded.	SKILL.md:220-227. Algorithmic map :1041-1047; model-adapter/native.rs:153-167; map.yaml:79-80.
Failure / recovery	Stop at a building commit boundary; triage into fix, fork or follow-up; git scars.	Typed EpiphanyModelPassFailure; death recovery; tombstones; epoch hard cuts; failed capstones sealed and never resumed.	briefs.md:92-93; SKILL.md:169-176, 234-254. Algorithmic map :1082-1097, :1220-1231, :1440-1502.
Compaction / re-entry	Maps committed; steward at boundaries.	AGENTS.md re-entry rite plus steward; runtime Reorientation worker decides resume or regather from a sealed continuity projection.	AGENTS.md:130-180; reorientation_work.rs:475.
Build economy	Detached builds with log and PID; tag-push limit; reproducible byte checks.	One package per entrypoint; focused single-package cargo only; Idunn owns broad builds; interrupted builds settled before retry.	SKILL.md:247-249. AGENTS.md:250-277.
Scale evidence	1 migration: 114 subagents; 29/32 Soul passes found defects (about 75).	2,269 commits; about 65k tracked Rust lines; 9 bins; no accepted current-package end-to-end capstone.	postmortem-cultcache.md:10-22. git log; Cargo.toml; map.yaml:7, 114-118.
What each proposes but lacks	Nothing typed: all artifacts are prose and habits.	Lessons doc proposes typed Verdict, Rulings, landed-names digest, edit queues, steward triggers, and a Self budget, matching the skill's habits; none are implemented.	faculty-workflow-lessons-2026-09-04.md:146-176.