Approved, Approved+Reinforcement, or Redirected+Reinforcement... every major session ends with a human writing the WHY that compounds the Brain's judgment.
Without the loop, you have a logger. With it, you have an apprentice that gets more capable every week.
We built it this way because we know our own weaknesses well. If you build something different, it'll be because you know yours.
Phase 0 is the foundation. The Brain itself runs as a future Phase 2 deliverable. Today the page is honest about what is shipped and what is coming. The reinforcement loop is the load-bearing mechanism; the runtime is downstream of it.
Every decision the Brain makes carries a WHY at the moment of choice. Not retroactive. Not assembled later. The reasoning is captured in the same breath as the choice it explains.
At the end of every major session, a decision making member or team reads the report. Each call gets one of three verdicts. The corpus updates from each one differently, and the trust meter for that decision category moves accordingly.
Reviewer reads the WHY at moment of decision and accepts. The choice and the reasoning matched the established pattern.
Reviewer accepts AND adds context that strengthens the pattern. The reasoning was right; now it is also more durable.
Reviewer disagrees and explains why. The redirect is itself reinforcement. Teaching, not punishment.
The 'plus Reinforcement' on both Approved and Redirected is intentional. A redirect with a written WHY is reinforcement; the corpus learns from the correction the same way it learns from a strengthened approval. The reviewer is the teacher, not the gate.
Retroactive reasoning lets a system rationalize bad choices. WHY captured before the choice is committed binds the reasoning to the moment. It is the discipline that makes the corpus honest.
Global agreement-rate hides the categories drifting away from the reviewer's actual preferences. Per-category subdivision keeps the calibration honest. A category that loses trust re-escalates on its next decision until it rebuilds.
The Operating Standard clarifies WHO an organization is. The Brain clarifies HOW it decides. The Wisdom Library, a separate artifact, holds what the organization knows.
Conflating those three flattens the architecture and the corpus loses precision. The distinction is small to read and load-bearing in practice. The Brain reads from the Operating Standard for constitutional rules; it reads from the Wisdom Library for relevant context; it writes to its own ledger and corpus.
The Brain started as a ledger: it recorded choices after the fact. The Deliberation Council lets it actively deliberate. A convener seats a constitutional core plus the specialist lenses a question calls for, they argue from our real values, voice, and system corpus, and a Judge synthesizes one verdict. What you see here is the organ's shape and its discipline, never the content of any real verdict.
Three constitutional seats are always at the table, whatever the question. They are roles, not people.
Guards the values and the people the work serves. Holds a veto: a hard-rule breach forces a kill, no matter how the rest of the room scores it.
Assumes the idea fails and argues the strongest honest case against it, so the best version of the objection is always heard.
Speaks as the donor-partner, the family, the person on the receiving end, and asks whether the choice makes them feel loved or processed.
Beyond the constitutional core, a specialist faculty seats the lenses a given question needs and benches the rest. The faculty is disclosed here by category and count only. Some of these lenses speak for the people we serve, so they are never named on a public page.
A council verdict lands NOVEL and pending. It rides the same three-state reinforcement loop every Brain decision does, and dissent travels with the verdict, verbatim and attributed. See How it works, above, for the loop itself.
The council never promotes its own verdict to trusted precedent. Every verdict stays pending until a human ratifies it on the review card. That gate is separate from, and stricter than, the question of whether a verdict is correct: nothing reaches this public page until a second review clears its text as safe to publish.
The council is operational, and it deliberates real questions. It is also young. We will not dress a small sample as maturity.
3 ratified decisions across 3 categories so far. The live verdict view and the per-category trust meter open once the corpus is large enough to be honest about, and once the second publication review is in place. Until then, this page shows the organ's shape, not its records.
The Brain map follows the same JSON-source-of-truth, deterministic-renderer, drift-validator pattern the Operating Standard uses for its own development map. Reuse, not fork. 20 of 32 nodes shipped (62%).
Shipped 20 of 32 (62%)
Planned 12
What the Brain stands on before code. Master capture, OS+Brain companion framing locked in memory, three-verdict reinforcement loop, role-based reviewer design call, share-not-sell voice discipline, public-site launch plan. Phase 0 is the architectural ground truth.
p0.brain.master-capture Master Architectural CaptureCaptures the full Brain architecture: Strategy C autonomous-overnight orchestrator, three-verdict reinforcement loop, decision-ledger schema sketch, anti-patterns, success metrics, integration points with existing systems, phased build plan.
p0.brain.os-companion OS + Brain Companion FramingLocks the Brain as the operational HOW-layer that complements QOS. QOS clarifies WHO an organization is (schemas); Brain clarifies HOW it decides (choices + WHY). Wisdom Library is a separate artifact handling what the org KNOWS, not part of the Brain's territory.
p0.brain.role-based-reviewer Role-Based Reviewer DesignForward-looking design call: decisions.v1 schema must support multiple reviewers and role-based routing from day one. A future supporter adopting the Brain might have a small board, not a single founder. Schema needs reviewer (role + identity) and verdict.signed_by from the start.
p0.brain.share-not-sell Share-Not-Sell Voice DisciplineNamed corpus preference enforcing no-competitive-framing on all public-facing QWF content. Frame work as 'what fits our weaknesses and strengths,' never 'better than what others do.' Underneath: invitation, not argument. Reader's first job is Know Thyself.
p0.brain.site-launch-plan brain.quietlyos.org Launch PlanPhased launch plan for the public Brain home. Separate-repo decision (QuietlyWorking/quietly-brain), CF Pages + SvelteKit, subdomain CNAME, six Phase 0 sections, seven new components, day-one node tree, three-sweep anonymity discipline, mission-thesis hero copy locked.
p0.brain.site-phase-0 brain.quietlyos.org Phase 0 Public LaunchBuild and ship the public Phase 0 surface of brain.quietlyos.org per the launch plan. Six sections (hero, how-it-works, architecture, build map, what-goes-here-next, open standard), seven components, supporter sweep + em-dash scrub at build, CF Pages deploy, DNS CNAME, SSL, cross-link with quietlyos.org footer.
Depends on: p0.brain.site-launch-plan
p0.brain.three-verdict-loop Three-Verdict Reinforcement LoopLocks the three verdict states (Approved / Approved+Reinforcement / Redirected+Reinforcement) and their effects on corpus and trust meter. Names the redirect itself as a reinforcement event when paired with WHY.
decisions.v1 schema, the report-format template, the append-decision helper script. The JSON contract for every Brain entry, plus the writer.
p1.brain.append-script append_decision.py HelperPython helper script that appends a decisions.v1 entry to the ledger with WHY, trust weight, and trace. Validates against decisions.v1 schema before write. Returns structured result.
Depends on: p1.brain.decisions-v1
p1.brain.decisions-v1 decisions.v1 SchemaJSON Schema (Draft 2020-12) defining the decision-ledger entry shape: id, context, call, WHY-at-time-of-decision (required), expected outcome, trust weight (HARD-RULE / SOFT-PREFERENCE / NOVEL), verdict (pending / approved / approved-with-reinforcement / redirected), feedback (TIG's WHY on redirect), trace (commits, files, HQ items, dev-map nodes), reviewer (role + identity), verdict.signed_by.
p1.brain.report-format Report Format TemplateLocked template for end-of-session Brain reports. Required: Non-technical description first per call, top-N call detail, action table at end, open-questions section, cross-build observations. Format requirement no-silent-approval-during-formation enforced until trust meters stabilize.
Depends on: p1.brain.decisions-v1
The headless Claude Agent SDK runner that reads a build plan, queries the corpus at decision time, logs WHY, pauses on novel decisions and escalates to HQ, stages commits without pushing. The Brain becomes load-bearing here.
p2.brain.commit-staging Commit Staging (No Push)Brain stages commits but never pushes to main. TIG green-lights pushes during morning review. The zero-autonomous-commits gate from CLAUDE.md applies absolutely.
Depends on: p2.brain.headless-runner
p2.brain.corpus-query Corpus Query at Decision TimeDecision-time lookup against memory/feedback_*.md files plus prior decision-ledger entries. Returns relevant SOFT PREFERENCES and precedent decisions for the Brain to reason against.
Depends on: p2.brain.headless-runner
p2.brain.first-overnight-run First Successful Overnight RunInaugural autonomous overnight run executes a real build plan (likely a future schema checkpoint), produces ledger entries, queues HQ escalations as needed, stages commits. Morning review verifies the run via the report format.
Depends on: p2.brain.corpus-query,p2.brain.surprise-pause,p2.brain.commit-staging
p2.brain.headless-runner Headless Claude Agent SDK RunnerLong-running Claude Agent SDK process that reads a build plan, executes checkpoints, calls append_decision.py at every decision point, surfaces escalations to HQ, stages commits without pushing.
Depends on: p1.brain.decisions-v1,p1.brain.append-script
p2.brain.surprise-pause Surprise Pause + HQ EscalationWhen the Brain hits a NOVEL decision (not in corpus, not a HARD RULE), it pauses the run, files an HQ item with full context, and waits for TIG's verdict. No autonomous resolution of novel calls. Refined during Phase 3 build: rather than literally pausing the run on NOVEL decisions, the runner logs them as pending verdicts (via append_decision dual-write) and continues. The reviewer verdicts asynchronously on the Brain page. Functionally equivalent to the original spec... no autonomous resolution of novel calls, every NOVEL gets a card, the corpus only grows via reviewer-blessed verdicts. Aligns with the Phase 2A escalation-contract refinement (chat-report+card pattern replaces blocking pause).
Depends on: p2.brain.headless-runner
p2.council.m1-proof Deliberation Council M1: Proof SliceStands up the council.v1 persona schema and the deterministic engine: a constitutional Core Quorum (Compass Keeper with veto, the Contrarian, the Voice of the Served) plus hand-authored specialists deliberate one real question, and a Judge synthesizes a single verdict. Dry-run only at this milestone; no ledger writes.
Depends on: p1.brain.decisions-v1
p2.council.m2-faculty Deliberation Council M2: Faculty GeneratorAdds the auto-derivation generator that expands the faculty across QWF systems and brands, plus structured verdict types (binary, selection, ranking, open) that carry structured fields rather than prose. The faculty is derived from QWF system and brand profiles, never hand-curated per run; the Core Quorum is never generated.
Depends on: p2.council.m1-proof
p2.council.m3-live-write Deliberation Council M3: Live Write + Ratification GateDrops the dry-run guard behind a live gate: the council writes a real pending decisions.v1 entry and an HQ card, and a ratification write-back syncs the canonical verdict state only after the human ratifies that card. Agent proposes; human ratifies. Every verdict lands novel and pending until then.
Depends on: p2.council.m2-faculty,p1.brain.append-script
p2.council.m4-calibration Deliberation Council M4: Calibration MeterAdds the per-category calibration aggregator (how often the council's lean matched the human's actual verdict) plus a retrieval-miss log, so any move toward autonomy is gated on accumulated agreement data rather than on confidence. Accumulation is ongoing and human-triggered; the agent never convenes itself.
Depends on: p2.council.m3-live-write
The interface that closes the loop. Batch review surface, three-verdict mechanism (Approved / Approved+Reinforcement / Redirected+Reinforcement), trust-meter calculation, corpus update on redirect.
p3.brain.batch-review Batch Review Surface (HQ Filter View)HQ Command Center filter view that surfaces every Brain ledger entry awaiting verdict, grouped by session, with the report format pre-rendered. TIG (or designated reviewer) verdicts in 5 to 10 minutes.
Depends on: p2.brain.first-overnight-run
p3.brain.corpus-update-on-redirect Corpus Update on RedirectWhen a verdict is Redirected+Reinforcement, the writer updates the relevant memory file with TIG's WHY. Append for new categories, rewrite for existing rules. Decrement trust meter.
Depends on: p3.brain.verdict-mechanism
p3.brain.trust-meter-calc Trust Meter CalculationPer-decision-category agreement-rate over the last 20 decisions. High (90%+) auto-handles, medium (70-89%) flags for spot-check, low (<70%) re-escalates. Subcategorization prevents over-generalization.
Depends on: p3.brain.verdict-mechanism
p3.brain.verdict-mechanism Three-Verdict MechanismReviewer interface that captures Approved / Approved+Reinforcement (with WHY) / Redirected+Reinforcement (with WHY). TIG-only signature. Append-only ledger; no edits to past verdicts.
Depends on: p3.brain.batch-review
Per-category trust meters, stale-corpus detection, random-sample spot-check on high-trust decisions, prompt-injection integrity protections.
p4.brain.sr-reinforcement-capture Senior Reviewer Reinforcement Capture (v1)Capture corpus-grade APPROVED + REINFORCEMENT verdicts issued by Senior Reviewer agents during normal QWU work, so the gold compounds in the Brain corpus rather than evaporating into chat history. v1 shipped: two new fields on hq_brain_verdicts (detection_rule, issued_by) via migration 03; write_brain_verdict.py v1.1.0 accepts them and touches last_reinforced on cited memory files as a free-win side effect; Senior Reviewer prompt-template instruction block appended to feedback_senior_reviewer_multi_agent_pattern.md; Lovable prompt 146 added an issued_by filter chip to the Brain page alongside existing trust-weight chips plus provenance surfacing in the card detail drawer. Threshold gate: only write a card when the reinforcement names a NEW pattern OR calibrates an existing pattern at a NEW altitude. Seven inaugural Senior Reviewer cards landed on the live system 2026-05-04 covering surface-design, migration-discipline, verification-discipline, scope-discipline, ui-affordance-discipline, persistence-discipline, and validation-discipline categories.
Depends on: p2.brain.headless-runner,p3.brain.verdict-mechanism
p4.brain.per-category-trust Per-Category Trust SubdivisionSubdivide trust meter into specific decision shapes (e.g., 'cascade naming for forbidden patterns' vs 'cascade naming for vocabulary alternatives') so high agreement in one shape does not inflate trust in adjacent shapes.
Depends on: p3.brain.trust-meter-calc
p4.brain.prompt-injection-integrity Prompt Injection Integrity ChecksSigned ledger entries (agent-id + commit hash). Append-only ledger. TIG-only verdict signatures. Random-sample integrity checks against known-clean baseline.
Depends on: p1.brain.decisions-v1
p4.brain.spot-check-high-trust Random-Sample Spot-Check on High-Trust DecisionsEvery batch report random-samples a few HIGH-TRUST decisions for explicit verdict ('the Brain says this was 95% confidence... do you actually agree?'). Track over-approval as its own metric.
Depends on: p3.brain.verdict-mechanism
p4.brain.stale-corpus-detection Stale Corpus Detection (90-Day Rule)Every memory file gets a last-reinforced timestamp. Decisions in categories not reinforced for 90 days re-escalate to TIG. Stale rules expire safely instead of compounding silently.
Depends on: p3.brain.corpus-update-on-redirect
The Brain pattern published as a public QOS reference implementation. Documentation, governance license, adopt-page with fork instructions, first supporter adopter.
p5.brain.adopt-page Adopt-The-Brain PagePublic page on brain.quietlyos.org with concrete fork-it instructions: clone the schema, scaffold a corpus, plug into your decision rights structure, run your first batch review. Includes worked examples from QWF (anonymized) and any consenting supporter adopters.
Depends on: p5.brain.public-pattern-doc,p5.brain.governance-license
p5.brain.first-supporter-fork First Supporter ForkFirst non-QWF organization adopts the Brain pattern using the published schema + adopt page. Their corpus stays private; HARD RULES inherited from QOS public standard; SOFT PREFERENCES are theirs.
Depends on: p5.brain.adopt-page
p5.brain.governance-license Governance LicenseOpen-source license under which the Brain pattern (schema + reference implementation) is published. Aligns with the broader QOS license decision (currently '[pending governance]').
Depends on: p5.brain.public-pattern-doc
p5.brain.public-pattern-doc Brain Pattern DocumentationPublic documentation of the Brain pattern as a QOS reference implementation. Lives at brain.quietlyos.org; covers schema, runtime, reinforcement loop, anti-patterns, success metrics, fork-it instructions.
Depends on: p2.brain.first-overnight-run,p3.brain.corpus-update-on-redirect
The map is authored at 005 Operations/Standards/roadmap/BRAIN.development-map.json in the QWU backoffice and copied into this repo on each update. Schema reuse:
validates against qos_development_map.v1.schema.json (no Brain-specific fork in Phase 0). Identity differentiation via identity.shortName: "BRAIN". Map version v1.1.0.
Public render with supporter-anonymity sweep, youth sweep, and em-dash scrub
applied at build time.
Phase 0 ships the foundation. The Brain runtime, the live ledger, the trust meters, and the named-preferences corpus are all downstream. Each lights up at its own phase and gets its own surface here.
The placeholders below name the coming surfaces and the phase that ships them. No mocked data. The page tells the truth about today.
This is where anonymized decision-ledger entries will render. Each entry leads with a non-technical description, then surfaces the choice, the WHY at the moment of decision, the verdict, and the date. Three sweeps (supporter, youth, TIG-personal) run at render time.
Today there are 3 ratified decisions across 3 categories. That is a true count, not a sample we will dress as more. The public verdict view opens once the corpus is large enough to be honest about, and once a second publication review clears each entry's text as safe to publish. Until then this slot stays empty rather than thin.
Per-category agreement-rate bars: how often the council's lean matched the human's actual verdict. High-trust categories handle autonomously; medium flag for spot-check; low re-escalate. Subcategorization prevents trust inflation.
Today the calibration log holds 3 records across 3 categories, every one still at the human-reviewed stage. A meter drawn from a sample that small would read as maturity that has not been earned, so it stays closed. It opens once enough ratified decisions accumulate to make a per-category rate honest.
Named preferences with their reinforcement history. Each entry lists the sessions where the preference was reinforced or redirected, plus its last-reinforced timestamp for stale-corpus detection at the 90-day mark. Lights up in Phase 3, alongside the verdict mechanism.
The Brain pattern is one road through the same country many builders cross. We chose it because we know our own weaknesses well. If you build something different, it'll be because you know yours.
The schema for ledger entries (decisions.v1) ships in the next phase and will live in the Operating Standard schema gallery. Once it ships, the contract is portable: clone the schema, plug it into your decision-rights structure, run your first batch review with your own corpus.
The Operating Standard is the public home for QWF's machine-readable identity layer. The Brain pattern lands there as a reference implementation when the schema, runtime, and reinforcement mechanism stabilize.
Visit the Operating StandardThe repo for this site, the build map JSON, and the build-time sweeps are public. Read what we did and why we did it. Use what fits. Leave what does not.
github.com/QuietlyWorking/quietly-brainLicense pending governance. The QWF Advisor Brain pattern, schema, and reference implementation will publish under the same license as the broader Operating Standard once that decision lands.