EXECUTIVE LOOP
PRIVATE INTELLIGENCE LAYER
Overview
CURRENT STATE
A durable brain, built to remember without pretending.
Verified knowledge stays separate from agent claims. Every project is isolated. Model output remains a candidate until you decide otherwise.
ATTENTION
Needs a look
Brain answers do not approve work, send messages, create verified memory, or authorize external action. They are advice for Chance to review.
BOUNDED COGNITION
Ask Chance Brain
EVIDENCE-BACKED ROUTING
Expert Mesh
Describe one task and let the Talent Router assemble an evidence-backed advisory team from reviewed skills and agent resumes.
No agent is started, no model is called, no prompt is sent, no memory is written, and no work or external action is approved.
TALENT ROUTER
Build a council plan
OWNER LAUNCH PAD
Agent Runner
Queue one bounded Claude review from the exact current Expert Mesh plan. Codex receipts remain visible, but Codex execution is hard locked in v0.33.0 pending genuine Windows deny-read isolation.
Owner-only · one read-only advisory run · repository read/search only. Claude is the sole executable Phase 1 lane and is limited to Read / Glob / Grep with Bash and PowerShell denied. Codex is disabled because the native Windows read-only sandbox cannot deny reads outside the detached repository snapshot. No unrestricted or mutating shell, edit/write, web/network, MCP, external actions, fanout, child launches, or automatic retry.
EXACT PLAN LAUNCH
Queue one specialist review
DURABLE RUN RECEIPT
Current or latest run
Status refresh uses GET only. Polling can observe a queued run; it never creates, starts, retries, or fans out work.
VERIFIED CONTEXT
Memory
EPISTEMIC REVIEW
Working Memory
Stored hypotheses are not facts and never enter trusted context automatically. Evidence can support or challenge a belief; only Chance can review what deserves further verification.
This area preserves what a model proposed and what still needs research. It does not silently teach the Brain that a claim is true.
MODEL HYPOTHESES
Cognitive beliefs
OTHER PROPOSALS
Candidate memories
CROSS-AGENT BRIDGE
Messages and handoffs
DURABLE RECEIPTS
Work episodes
POINT-IN-TIME RESEARCH
Swing Trading Lab
Build source-backed research packets for 5–15 session U.S. stock and ETF hypotheses, then keep observations, theses, experiments, and paper outcomes separate.
No Robinhood connection, account access, positions, balances, order placement, live trading, recommendation authority, or automatic memory promotion.
RESEARCH INTAKE
Build a bounded packet
HYPOTHESES
Open theses
TESTS
Open experiments
POINT-IN-TIME INPUTS
Recent observations
PAPER EVIDENCE
Candidate outcomes and reviews
PUBLIC INTEREST RESEARCH
Case Desk
Build a case from specific claims, original sources, counterevidence, and responses. Every addition becomes part of its recorded history.
POPULATION-LEVEL OVERSIGHT
Program Audit Lab
Audit a complete public-data population with the same explainable rules, preserve the exact source snapshot, and move only the strongest exceptions into confidential case work.
Priority scores identify records worth verifying. They do not measure guilt, damages, recoverability, or tip eligibility. Nothing here contacts a recipient, publishes a claim, or files a report.
EVIDENCE FEDERATION
Recipient coverage and corroboration
Corroborated research priorities
0Unresolved identity matches
0Exact UEI, EIN, FAIN, award ID, or grant-number links can join automatically. Similar names never do. A corroborated tier measures independent public-source coverage, not guilt, intent, loss, or recoverability.
ORIGINALS AND SOURCE-BOUND FACTS
Evidence archive
Original files are checked against their SHA-256 hash. Extractions remain unverified until reviewed. Relationships, reported resolutions, and calculations do not change rankings or establish fraud.
Recent original documents
Recent extracted assertions
Reproducible calculations
Import selected public documents or IRS filings
For public documents, prepare a source manifest with exact identifiers, reporting periods, page or field locators, and extracted records. Validate locally first:
npm run audits:evidence -- --input manifest.jsonFor selected full Form 990 XML filings, the trusted-computer importer extracts financial metrics, explicit governance disclosures, and reported relationships:
npm run audits:irs -- --input irs-manifest.jsonAdd --apply only after review. IRS imports preserve originals and create a selected-return snapshot; they do not run federation or change existing rankings. Missing fields are not zero or false. Automatic PDF/OCR extraction is not enabled.
Resolved-case playbookWhat actually made prior Head Start matters provable—
Case-derived investigation patterns
These are manual source-join targets, not automatic fraud rules. Each one includes the decisive evidence and the innocent explanations that must be tested.
Evidence gate for a credible referral
No percentage is calculated. One missing element can defeat an otherwise impressive packet.
Official resolved-case sources
PROVENANCE