Brain Source Fabric

Give every source—from a saved post to a paper, repository, conversation, screenshot, skill, meeting, or live system event—one inspectable path from raw evidence to changed knowledge, workflow behavior, retrieval, and later correction.

The harness should not equate captured, indexed, understood, believed, routed, and adopted. Those are separate states. A source may be preserved without shaping the brain; it may shape a project without becoming global guidance; it may challenge a belief without immediately replacing it; or it may become a first-class workflow fixture only after proof. Source: User request, 2026-07-15

This page is the source-layer companion to Harness and Brain Workflow Refactor Plan. The master plan owns workflow execution. This page owns how evidence enters, acquires meaning, changes the compiled brain, and remains auditable.

Full Corpus Workflow and Capability Program owns the exhaustive census across the source fabric: the review-closure test for each artifact class, workflow extraction, capability identity, independent install/config/auth/health/binding states, the normalized workflow universe, and the conditions required before “every source was reviewed” is defensible.

Wiki Corpus Refactor Program owns canonical-shape evolution after evidence arrives. A source disposition may update an owner, create a justified new mental object, fold into a section or ledger, merge competing owners, archive evidence, or produce no change. The source fabric must not assume that the current file graph is the permanent ontology.

Current-State Audit

The repository already contains substantial source infrastructure, but its depth is asymmetric. Source: local repository inventory, npm run brain -- status, and npm run status, 2026-08-12

The existing source-compile automation names X bookmarks, Codex/Cursor sessions, and skill folders. It does not yet compile the full local source corpus. Freshness currently reports 28 overdue or never-run automation/sync paths. That is a harness design issue: freshness, connector health, backlog depth, and compilation state need one source-level control plane rather than scattered timestamps. Source: automations/source-compile.md; state.json; npm run status, 2026-07-15

Message identity and review-unit closure

Connector event identity, message identity, and review identity must not be collapsed. A retried delivery of one stable event is idempotent. Two messages with identical text are two observations and retain two source IDs. One edited message retains its source ID but gains a new immutable revision. A session snapshot preserves whole-trace custody but never substitutes for its messages.

Review closure is equally exact. Each review unit records a SHA-256 digest over its sorted source IDs, source-revision IDs, and signal IDs plus their counts. integrated and held are final dispositions for that digest; partial and blocked remain pending. A receipt for an older digest cannot close a changed unit. This makes a public aggregate compact without allowing a queue count, content duplicate, or last transcript message to impersonate complete review.

System Contract

Every connector must implement the same minimum contract:

1. a stable source identity and cursor;
2. an immutable or content-addressed raw artifact;
3. a normalized envelope;
4. source-specific enrichment and attachment preservation;
5. privacy and authority classification before broad retrieval;
6. extraction into atomic evidence units;
7. an explicit disposition receipt;
8. compilation into canonical owners and generated views;
9. retrieval/effect tests;
10. freshness, failure, backlog, and replay state.

Authority scope and source scope are independent

Before retrieval, compilation, or writeback, resolve two different questions:

The same slug, entity, or semantic match may exist in several sources or authority domains. Resolution therefore returns a qualified identity rather than a global filename assumption. Cross-authority retrieval is explicit fan-out by an authorized agent, and synthesis cites every authority/source pair. A write stays in the owning authority and source unless a separately approved compilation receipt promotes it elsewhere. This makes privacy and project isolation part of routing, not a filter applied after retrieval. Source: GBrain brains/sources topology and source-isolation invariants at 75fae742, reviewed 2026-08-11

Connector execution contract

Cerebras Enterprise Knowledge Base Reference Architecture turns this abstract contract into a production-oriented execution model. A connector is a reviewed package with a manifest, adapter, fixture corpus, contract test, and doctor. It must declare its source scope, authorization reference, poll/event mode, stable identity and cursor semantics, semantic unit, refresh strategy, raw-retention policy, normalized fields, lexical and semantic projections, ACL strategy, freshness SLO, retry/dead-letter policy, and health checks. Source: Cerebras, 2026-07-16

Event-driven connectors follow one common path:

receive and acknowledge
→ deduplicate stable event ID
→ enqueue
→ resolve complete semantic unit
→ preserve restricted raw snapshot
→ normalize and classify
→ publish lexical projection
→ publish semantic projection
→ idempotent upsert
→ receipt, health, and freshness update

“Complete semantic unit” is source-specific. It may be a Slack/Discord thread, an article plus figures, a repository symbol and its file context, an email thread, or a meeting with transcript and attachments. A delta event is a reason to refresh that unit; it is not automatically the unit agents should retrieve.

For Discord, Discord Brain Gateway specializes this contract. message_id is stable delivery identity; each changed content or artifact hash is an immutable revision. Create, edit, delete, pin, attachment, and embed changes are distinct source events. An edit refreshes the bounded reply/thread unit, diffs extracted evidence, reopens only affected owners, and invalidates any unapplied proposal approval. A bot DM or explicitly authorized channel is the supported steady-state boundary; authenticated browser scraping remains recovery evidence and cannot prove cursor completeness.

Postgres can own the shared evidence-row and control-plane projection, but it does not replace source ownership: native systems remain operational truth, raw artifacts remain replay evidence, and canonical Markdown remains compiled brain truth.

Source Taxonomy

Source type and subject domain are independent. A Discord message may concern design, career, agent architecture, or a reminder. An article may be a tool manual, research claim, visual fixture, or personal philosophy. The harness must classify both axes.

Normalized Source Envelope

All adapters emit one envelope before domain interpretation. Source-specific payloads remain attached rather than being flattened away.

source_id: src_<stable-id>
source_family: saved-curation
source_type: x-bookmark
connector_id: x-bookmarks
external_id: "207..."
canonical_url: "https://x.com/..."
captured_at: 2026-07-15T00:00:00Z
published_at: 2026-07-13T00:00:00Z
author_entities: [person_or_org_ids]
raw_artifacts: [content_hashes_and_paths]
attachments: [image_video_pdf_repo_or_link_records]
thread_context: [parent_and_reply_ids]
privacy: private
authority_domain: personal-brain
content_hash: sha256:...
adapter_schema_version: 1
freshness_policy: event-immutable
enrichment_state: complete
extraction_state: reviewed
disposition_state: incorporated
errors: []

Required distinctions:

• published_at is not captured_at; old material can be newly relevant.
• raw_artifacts are evidence; extracted text and summaries are derived.
• privacy controls retrieval and compilation before content leaves the adapter boundary.
• freshness_policy differs for immutable posts, changing docs, live project state, prices, schedules, and personal reminders.
• schema migrations are replayable from raw artifacts.

Evidence Units

Enrichment produces typed units rather than one undifferentiated summary.

A claim record should preserve valid_from, valid_until, confidence, source quality, extraction confidence, supporting and contradicting evidence, canonical owner, and whether an agent or human reviewed it. Confidence is not truth; it is a routing aid.

Correlation Is Not Transitive Topic Identity

The evidence graph is many-to-many; the review queue must not flatten that graph into accidental subjects. A multi-link roundup that mentions Tool A and Tool B establishes two mentions edges. It does not make every other post about Tool A part of one transitive Tool A + Tool B cluster.

Use these deterministic review-unit rules:

1. Sources with exactly one canonical resolved link may form a focused-artifact unit with other sources that resolve to that same artifact.
2. A source with multiple canonical links stays one source-sized roundup unit. During interrogation, extract and bind every useful item and relationship separately.
3. Linkless sources stay source-sized until an agent establishes an explicit semantic relation.
4. Images, thumbnails, videos, and attachments are proof artifacts, not automatic topic keys.
5. Shared links, nearby Discord intent, named entities, time adjacency, and contradictions remain queryable evidence edges even when they do not merge review units.
6. An agent may deliberately review several units together, but the review record must name every source, signal, object comparison, and writeback; graph connectivity alone is never the justification.

This preserves repeated evidence without letting listicles or broad resource posts contaminate unrelated tool decisions. Ranking changes order only: useful signals survive even when a source is low-engagement, broad, or not yet actionable. Source: User correction and review-methodology audit, 2026-08-10; packages/brain-runtime/src/review-queue.ts

Incorporation Receipt

“Incorporated into the brain” is complete only when a receipt can answer:

• What raw artifacts were preserved?
• What did the source say, show, or demonstrate?
• Which claims, patterns, entities, preferences, tasks, capabilities, or fixtures were extracted?
• Which domain and privacy lenses reviewed it?
• What existing knowledge supported, contradicted, or scoped it?
• What changed: page, workflow, fixture, skill, tool route, project, task, or nothing?
• What workflow and capability candidates were created, merged, rejected, or deferred?
• What install next-action applies, including explicit no-install cases?
• Why was each change justified?
• Which generated views and retrieval indexes were rebuilt?
• What proof shows the source can now be found and affects the intended workflow?
• When must changing claims, tools, or live state be refreshed?

receipt_id: ingest_<id>
source_ids: [src_<id>]
primary_disposition: update-workflow
secondary_effects:
  - owner: wiki/architecture/example.md
    operation: qualify-claim
  - owner: workflows/frontend-improve
    operation: attach-fixture
review:
  domain_lenses: [design, frontend]
  contradictions: [claim_ids]
  privacy: private
proof:
  retrieval_queries: ["stable scrollbar detail"]
  workflow_evals: [frontend_fixture_eval]
freshness:
  checked_at: 2026-07-15
  recheck_after: 2026-08-15
no_change_reason: null

The receipt is the bridge between evidence and brain mutation. Hub coverage, a qmd embedding, or an absorb-log row alone does not prove incorporation.

Stable-object referential integrity

Every object named by a review is a foreign-key reference into brain/objects.json. The object ID and canonical path must both match. A review cannot invent a more specific alias for a section, point to an existing page that the operating graph does not register, or leave a writeback reachable only from its JSON receipt. brain doctor validates this boundary before projecting Changes and Objects into the UI. If a reviewed owner is genuinely durable, register it; if it is only a section of an existing owner, use that owner's stable ID. Source: graph-integrity audit and doctor regression, 2026-08-10

Source Profiles And Trust

The system should maintain profiles for recurring sources without converting reputation into a universal truth score.

Source profiles rank review priority and contextualize claims. They must never suppress contradictory evidence or make a high-status author automatically correct.

Domain Lenses

After normalization, evidence passes through one or more domain lenses. A lens owns questions, canonical routes, workflow effects, source-quality rules, and evaluators.

The same source can produce effects through several lenses, but one receipt owns the fan-out and records why each owner changed.

Portable context, replaceable execution

Agent-platform adoption is a layered decision, not one global build-versus-buy vote. The durable differentiation is the context and verification layer that makes any competent runtime behave like Kevin's system. The wiki therefore owns its rules, executable skills, MCP and capability contracts, workflow manifests, source identities, stable-object graph, approval policy, evaluations, and proof receipts. A hosted or self-hosted runtime may execute those contracts; it cannot become their sole canonical writer. Source: X/@davep X Article, 2026-07-27; Sierra Pinecone and MCP Gateway, July 2026

Stripe's Minions and Sierra's Pinecone show that bespoke infrastructure can be rational at scale, but they also make the exception legible: Stripe already owned mature devboxes and both organizations treat the system as a staffed, dogfooded product. Cursor's self-hosted workers show the opposite composition: code and tools can stay inside the customer's network while orchestration, models, and UX are bought. The source fabric keeps both options open by refusing to let provider state replace portable context or evidence. [Sources: Stripe Minions parts 1–2, February 2026; Sierra Pinecone and MCP Gateway, July 2026; Cursor self-hosted Cloud Agents and cloud-agent lessons, March–June 2026]

Minimum portability proof for a platform route:

1. export rules, skills, MCP/capability configuration, workflow input, and run receipt without scraping a vendor UI;
2. replay a representative source-to-owner decision in two harnesses and report semantic or capability differences;
3. prove that provider interruption cannot erase canonical source, approval, decision, or artifact identity;
4. keep vendor-specific adapters outside normalized source, object, review, and proof schemas; and
5. record security, governance, cost, reliability, and replacement evidence as independent axes rather than a single adoption boolean.

Capability And Skill Lifecycle

Tools and skills need a stricter lifecycle than articles:

discovered
→ source-resolved
→ candidate
→ security/reputation/compatibility evaluated
→ trial-installed
→ configured and authenticated
→ doctor-healthy
→ workflow-bound
→ active
→ drifted or superseded
→ deprecated/retired with history

The source fabric stores discovery and evaluation evidence. The capability registry owns active state. A saved tool, impressive demo, or imported skill cannot affect agent routing until installation, configuration, authority, version, doctor, workflow owner, and rollback status are known.

Use Full Corpus Workflow and Capability Program for the independent state axes and next-action flags. In particular, installed, configured, authorized, healthy, workflow-bound, and active must never collapse into one boolean.

Contradictions, Corrections, And Belief State

Canonical pages should expose compiled truth while the evidence layer preserves disagreement.

1. Compare new claims with current canonical claims and time bounds.
2. Distinguish direct contradiction, newer state, narrower scope, terminology mismatch, and independent support.
3. Route high-impact conflicts to a doubt/review workflow.
4. Change compiled truth only with a receipt naming the superseded claim and evidence.
5. Preserve the old claim in provenance/timeline history.
6. Re-evaluate workflows and fixtures whose assumptions depended on the changed claim.

Direct user corrections have highest authority for Kevin’s identity, intent, preferences, and private history. They do not override external facts outside that authority domain.

Queues And Backpressure

The system needs explicit queues so 1,000 bookmarks do not compete with urgent project state.

Priority combines explicit user selection, active-project relevance, contradiction severity, actionability, source quality, novelty, freshness risk, backlog age, and review cost. It must not be engagement-only.

Graph-guided frontier expansion

The object graph can guide what the source loop investigates next without becoming an evidence generator. Begin from a coarse canonical object or an unresolved question, explore existing equivalent and related nodes before creating another, expand toward more atomic concepts, and query with the target node plus its ancestor context. Retrieved posts, papers, sites, repositories, media, and conversations then enter the normal source envelope and review contract. Source: Kimi K3 technical report §4.2.2, frozen at MoonshotAI/Kimi-K3@3cb39dfd

This is a candidate-discovery lane, not automatic truth or automatic review clustering. Agent-created nodes, query keywords, synthetic tasks, inferred edges, and generated summaries remain derived artifacts. They cannot support a claim, merge review units, or update a workflow until the underlying materials are captured, cited, interrogated against the current owner, and given a proof receipt. Use coverage and unresolved-tension signals to select the next branch; do not recursively crawl the graph merely because a node exists.

Retrieval And Context Compilation

The runtime should retrieve compiled knowledge first, then evidence when needed. Retrieval is an ensemble, never an embeddings-only lookup.

1. The workflow manifest declares a reusable project/source bundle, knowledge routes, privacy authority, and freshness requirements.
2. ACL and project filters run before any candidate retrieval and cannot be widened by the query planner.
3. The context compiler loads canonical pages and current project state.
4. Independent retrievers produce exact/lexical, semantic, structural/graph, rarity, recency, and live-tool candidates when the underlying data shape supports them.
5. Weighted reciprocal-rank fusion combines consensus without trusting one scorer.
6. Source deduplication and per-source caps prevent one file or thread from occupying the result set.
7. A question-specific reranker selects a bounded set; context expansion adds neighboring sections or the complete semantic unit.
8. The final packet includes source, date, confidence, privacy, why it was loaded, and citation metadata.
9. Generated context bundles remain deterministic projections of the selected workflow and canonical owners.
10. The proof receipt records whether the retrieved context actually affected the result.

The default fusion reference is Σ weight / (60 + rank), followed by source consolidation, diversity caps, reranking, and context expansion. Exact constants remain evaluation parameters, not doctrine. qmd remains the local compiled-knowledge retriever; source adapters and live tools add evidence candidates through the shared interface. Source: Cerebras, 2026-07-16

The context packet should behave like a moving working set, not an attempt to load the whole archive. Begin with the smallest owner/project bundle that can answer the task, retain peripheral signals as lightweight candidate edges, and expand into full source units or neighboring objects only when the query or evidence justifies the move. This is the useful content behind Polygres' “infinite context” framing: the model's immediate attention remains finite while the addressable evidence can grow. The claim is retrieval architecture, not a literally unbounded transformer window. Source: X/@daleverett X Article, 2026-07-12

Structural retrieval is a distinct lane, not a replacement for lexical or semantic search. pgGraph demonstrates a rebuildable CSR graph index over ordinary Postgres tables; Polygres combines graph expansion with pgContext dense, sparse/text, filtered, and hybrid retrieval. This becomes relevant when the evidence store is already relational and repeated bounded traversal across declared relationships improves answerability. It does not justify moving Markdown/wiki truth into Postgres or replacing qmd. Authorization must be proved before graph expansion: pgGraph 1.0 documents that topology reads check table-level SELECT but do not re-evaluate row-level-security policies, while graph.search() re-queries source properties as the caller. Private or multi-tenant evidence graphs therefore remain ineligible until candidate generation and returned paths are demonstrably scoped before context assembly. Source: Evokoa/pgGraph README and docs/known-issues.mdx at 6fd7da9; Evokoa/pgContext at 70ee2ef, reviewed 2026-08-10

This avoids two failures: dumping the entire archive into context and hiding useful sources behind over-compressed summaries.

Privacy And Public Compilation

Privacy classification happens before global indexing and entity fan-out.

• Raw private artifacts live in restricted locations and are excluded from public builds, public backlinks, asset manifests, and search collections.
• Derived claims inherit the strongest relevant privacy class unless a separate public source independently supports them.
• Public pages may preserve historical professional facts without copying private planning, finances, conversations, or relationship context.
• Conversation participants and attachments are not promoted into entity pages by default.
• Proof artifacts redact private source excerpts and local paths.
• Deletion, retention, export, and audit behavior must be defined per connector.

Public portability uses the same boundary. The OKF v0.2 exporter receives only the public projection decision; it cannot reinterpret private: true, publish the private operational log, or traverse a local source path into output. Public concepts may expose resolvable public source URLs and a digest of any unmapped canonical source note, but its text, frozen private artifacts, and workflow receipts stay outside the bundle. This makes deletion and replay responsibility unambiguous: the source fabric owns evidence retention, while OKF is rebuildable delivery. Source: OKF v0.2 SPEC; ui/scripts/lib/okf-export.mjs, 2026-08-11

Control Plane

The source control plane should expose:

• connector: configured, authenticated, scheduled, healthy, last success, cursor, schema version;
• corpus: raw count, enriched count, extraction count, disposition count, backlog and oldest item;
• compilation: affected owners, failed writes, generated surfaces, qmd/index state;
• quality: unresolved links/media, duplicate rate, entity-link precision, contradiction queue, human correction rate;
• freshness: overdue live sources, expired tool versions, stale project state, changing-source rechecks;
• privacy: restricted items, public-build exclusions, redaction failures;
• cost: model/token/network/storage cost per source and per incorporated decision.

Daily Brief is the human review surface, not the state database. Daily Brief Control Plane owns the durable proposal/event ledger, state transitions, dedupe, ranking, privacy, exact approvals, and apply dispatch. The native app and CLI show ranked receipts and allow approve, reject, request revision, snooze, discuss, retag, restrict, assign a workflow, or request deeper research.

Automated research refreshes follow Workflow Research Refresh Workflow and land here as a batched proposal queue, not as one conversation per source or workflow. Each proposal exposes separate last-researched/changed/verified/run/failure clocks, upstream changes, affected canonical owners and workflows, before/after diff, source evidence, independent lens reports, strongest objection, confidence, eval result, authority, cost, and rollback path. Automation may update evidence and prepare a branch or PR; workflow, skill, automation, canonical-truth, and public-output application or merge remains explicitly human-approved. Source: Private Discord message 1527130200620404898, 2026-07-15

The control plane accepts typed proposal envelopes from source, workflow, capability, project, and operational adapters. It rejects items without stable identity, privacy, affected owners, source review state, expected proof, idempotency, and authority. Repeated automation runs update one aggregate instead of producing duplicate cards.

Implementation Boundaries

Staged Release

Release 0 — Inventory and receipts

• Create the source registry and normalized envelope schema.
• Import current connector/freshness state without rewriting raw sources.
• Generalize the X disposition receipt into a source-agnostic receipt.
• Measure backlog separately for capture, enrichment, review, promotion, and freshness.

Release 1 — Three representative adapters

Prove the full loop with:

1. X bookmark: social thread, link, replies, images/video, existing deep pipeline;
2. arbitrary URL/article/repository: long-form and capability evidence;
3. AI session or Discord thread: conversational decisions, corrections, attachments, and privacy.

Each must reach canonical writeback, workflow influence, retrieval verification, and replay from raw evidence.

Build them against one connector SDK and contract suite. Prove stable-event deduplication, complete-semantic-unit refresh, raw replay, lexical-before-semantic availability, ACL/project filtering, idempotent upsert, partial-failure recovery, schema migration, and clean removal/retention behavior. Retrieval tests must cover exact identifiers, paraphrases, freshness, diversity, and neighboring context.

Release 2 — Domain shaping

• Implement agent/harness, design/frontend, and research/learning lenses.
• Bind source fixtures to named workflows and evaluation status.
• Add contradiction review and belief-change receipts.
• Add candidate-to-active capability lifecycle gates.

Release 3 — Private and live sources

• Add email, calendar, meetings, Slack, voice/iMessage, tasks, GitHub/CI, and project-state adapters.
• Enforce private retrieval and public compilation boundaries before enabling automation.
• Add stale-intent/current-state checks for tasks and reminders.

Release 4 — Daily Brief control plane

• Implement the proposal envelope, event ledger, state machine, privacy, dedupe, ranking, workflow clocks, exact approvals, and CLI.
• Build the thin native macOS client after the control-plane contract passes fixtures.
• Migrate automation definitions to Workflow Automation V2 Contract so proposed work cannot become a false schedule.
• Prove source change → adversarial proposal → Daily Brief decision → approved apply/verify with a silent no-change fixture.

Release 5 — Aliveness and scale

• Add public domain feeds, last30days, novelty ranking, resurfacing, and Daily Brief actions.
• Benchmark batch sizes, cost, precision, correction rate, and time-to-incorporation.
• Automate safe dispositions only after receipt/eval history proves reliability.

Acceptance Metrics

Initial benchmark sets should include known X fixtures, the 151-message Straight Fats reconciliation with 39 edited rows and known virtualized-scroll misses, selected article/repository captures, deliberate contradictions, duplicate sources, old-but-newly-captured items, private/public collisions, and tools that should remain candidates rather than activate.

Failure Modes

• treating sync success as incorporation;
• summarizing a source while losing its claims, artifacts, figures, replies, or demonstrations;
• letting platform type stand in for domain interpretation;
• creating one page per source instead of compiling mental objects;
• allowing a popular source to overwrite stronger evidence;
• activating a tool or skill before installation/configuration/doctor proof;
• globalizing a project-specific or stylistic pattern;
• hiding contradictions inside a rewritten summary;
• embedding private content before classification;
• letting queues grow without stage-specific backlog and age metrics;
• refreshing indexes while canonical owners remain unchanged;
• generating proactive briefs with no merge/reject/writeback contract;
• keeping raw evidence forever without retention, export, or deletion policy.

Recommended Decisions

1. Make the source envelope and incorporation receipt kernel contracts; keep extraction implementations adapter-specific.
2. Use a relational/event ledger for source state, effects, claims, queues, and receipts; keep raw artifacts and canonical Markdown in their current owner stores.
3. Do not make embeddings the source of truth. Use them for candidate retrieval after privacy filtering, then resolve to canonical owners and evidence IDs.
4. Build the first release around X, arbitrary URL/repo, and private conversation evidence before adding more connectors.
5. Treat domain lenses as workflow-owned evaluators, not permanent autonomous personas.
6. Require replayability: changing an extractor or schema should recompile from preserved raw evidence without duplicating canonical writes.
7. Require a no-change receipt. Correct rejection and deferral are successful ingestion outcomes.
8. Adopt the Cerebras connector shape—small source adapter plus a shared evidence interface—but retain raw artifacts and canonical Markdown as separate owners.
9. Make hybrid retrieval explicit: lexical + semantic + rarity + recency, fused with RRF, deduplicated, diversity-capped, reranked, and expanded to surrounding context.
10. Use workflow manifests as reusable project/source bundles. Planners may narrow a scope but never silently broaden it.

---

Timeline

• 2026-08-11 | Added the portable-context/replaceable-execution boundary and a five-part portability proof for agent platforms. Stripe and Sierra validate the indefinitely staffed build exception; Cursor validates buying orchestration while retaining self-hosted execution. None may become the sole canonical writer for sources, policy, workflows, or proof. Source: X/@davep X Article, 2026-07-27; Stripe Minions; Sierra Pinecone and MCP Gateway; Cursor self-hosted Cloud Agents and cloud-agent lessons, reviewed 2026-08-11
• 2026-08-11 | Added independent authority/brain and source/repository routing axes before retrieval or writeback; cross-authority synthesis now requires explicit fan-out and qualified citations rather than implicit namespace merging. Source: GBrain brains/sources topology and source-isolation invariants at 75fae742
• 2026-08-11 | Defined OKF as a privacy-filtered delivery projection rather than a second source store: public concepts can carry source resources, while raw artifacts, private logs, decisions, and run receipts remain under their canonical owners. Source: OKF v0.2 SPEC; local conformance replay
• 2026-08-10 | Added graph-guided frontier expansion: use canonical objects and ancestor context to discover under-covered source material, deduplicate candidate concepts before creation, and keep generated nodes/tasks outside the evidence layer until ordinary capture, interrogation, and proof. Source: Kimi K3 technical report §4.2.2
• 2026-08-10 | Added moving-working-set retrieval and an optional structural/graph candidate lane. Kept qmd as the compiled-wiki default; held pgGraph/Polygres for relationship-heavy Postgres evidence until task-specific gain, RLS/tenant isolation, freshness, and benchmark proof exist. Source: X/@daleverett X Article; pgGraph, pgContext, Polygres SDK and skill sources, reviewed 2026-08-10
• 2026-08-10 | Replaced transitive URL-component review clustering with focused-artifact and source-sized roundup units. Shared artifacts remain evidence edges, while multi-link posts must be atomized during interrogation and can no longer merge unrelated tools into one subject. Source: User correction, 2026-08-10; brain-runtime regression fixture
• 2026-07-19 | Added Discord Brain Gateway and the 151-message reconciliation fixture: Discord now requires immutable revisions, edit-aware owner recompilation, explicit bot/channel scope, attachment repair, cursor-completeness tests, and proposal invalidation when a source edit changes the evidence. Source: Straight Fats Reconciliation — July 19, 2026
• 2026-07-16 | Incorporated Cerebras Enterprise Knowledge Base Reference Architecture as the connector and retrieval implementation reference: connector manifests/fixtures/doctors, stable-event dedupe, complete-semantic-unit refresh, lexical-before-semantic projections, hybrid RRF retrieval, context expansion, narrow MCP tools, and reusable project scopes now strengthen Releases 0–1. Source: Cerebras; User request, 2026-07-16
• 2026-07-16 | Added the 302-record X delta receipt and July 16 Discord continuation; promoted environment-profile setup and AI-native web-pattern synthesis while preserving tool installs as separate capability decisions. Source: X Bookmark Ingest — July 16, 2026; Straight Fats Harness Notes — July 16, 2026
• 2026-07-16 | Made Daily Brief Control Plane the durable proposal/approval owner, separated it from the native client and source database, added typed envelope rejection and dedupe rules, moved its release before aliveness at scale, and added approval/quiet-delivery metrics. Source: User request; Daily Brief — macOS Agent Review Artifact
• 2026-07-16 | Added the batched Daily Brief proposal queue contract: research automation records last-updated state and evaluated diffs, avoids conversation spam, and cannot merge workflow/skill changes without approval. Source: Straight Fats Harness Notes — July 15, 2026
• 2026-07-15 | Connected the source fabric to Full Corpus Workflow and Capability Program so every reviewed source emits workflow, capability, and install decisions in addition to knowledge effects, and capability presence cannot masquerade as configuration, health, workflow binding, or adoption. Source: User request and local corpus census, 2026-07-15
• 2026-07-15 | Expanded the workflow-first harness plan with a unified source fabric: current corpus/freshness audit, source taxonomy, normalized envelopes, typed evidence units, disposition receipts, domain lenses, capability lifecycle, contradiction handling, queues, retrieval policy, privacy boundary, control plane, staged releases, and acceptance metrics. Source: User request and local repository audit, 2026-07-15