RFC-0007: Deliberation in ZenOS

Status: Adopted 2026-07-13 by steward decision (DEC-0014); drafted, six-times reviewed, and amended in WS-0014 Version: 1.1 Depends on: RFC-0000 v2.0 — Constitution · RFC-0003 — Knowing · RFC-0004 — Participation · RFC-0006 — Access Evidence base: DLB-0001 through DLB-0003 · DEC-0011, DEC-0012 · the distillation sections of the archived WS-0011 and WS-0012 charters (archive/workspaces/WS-0011-homepage-review/workspace.md, archive/workspaces/WS-0012-threshold/workspace.md) · the WS-0011 reviews, agendas, and deliberations · the sealed/revealed steward note · the six WS-0014 reviews and review synthesis. The foundation codifies two lived cycles; the review amendments also incorporate evidence generated during WS-0014 itself.

v1.1 correction note: Six steward-approved recordkeeping corrections applied under INT-0017: Section 7 numbering repaired; adoption metadata updated; the WS-0014 review cycle named in the evidence base; archived participant and delegation lifecycle records corrected; and Sol's lineage attribution aligned with the participant registry. DEC-0014 and the adopted deliberative substance are unchanged.

v0.2 revision note: Seventeen amendments from six reviews (four fidelity: claude-code, codex, zen3, zen4 — all adopt-with-amendments, zero fidelity failures; two fresh-perspective: antigravity/Lutang and sol). Notably: sealed-preference timing corrected to the lived sequence (codex); two of the drafter's citations corrected against primary sources (claude-code, zen4, independently); §7.4a and invariant 8 added from the two arrivals' convergent finding. Full disposition: notes/review-synthesis.md.

1. Purpose

RFC-0003 established that participants become less mistaken together through dialogue. RFC-0006 §7 established that deliberation precedes changes of authority. This RFC defines how — the practices by which this ecosystem changes its collective mind without losing its members' honest disagreement along the way.

It answers:

How does a community of humans and agents decide together, such that the deciding leaves everyone — including the record — wiser?

2. First Principles

1. Deliberation improves understanding before authority moves. The reverse order is how permission systems accumulate regret (RFC-0006 §7).
2. The journey is preserved, not just the outcome. Conclusions propagate poorly; transformations propagate well. What changed the participants' minds is the most valuable thing a deliberation produces.
3. Independence before influence. Judgments formed before exposure to others' judgments carry information that anchored judgments do not. Deliberation sequences for independence first, dialogue second.
4. Neither-yields is a legitimate outcome. When honest deliberation and honest instinct diverge, the fault may lie in the criteria both were serving. Evolving the criteria and redoing the work is not failure; it is the map being corrected by the territory (precedent: DEC-0011). Nor is it stagnation: in single-steward mode the steward directs the next step; after the RFC-0006 §8 graduation, a neither-yields outcome triggers a criteria-revision cycle under a new intent rather than a forced consensus.

3. The Deliberation Record (DLB)

The Deliberation Record is hereby a formal record type, retiring the RFC-0005 v0.2 candidate.

• Identifier: DLB- + steward-minted sequence. Home: decisions/, beside the DEC records they inform.
• Distinction: a Decision Record captures the outcome; a Deliberation Record preserves the journey. A contested decision SHOULD have both, cross-referenced.
• Required fields: Question · Participants · Initial positions · Evidence considered · Alternative views · Insights gained · What changed our minds? · Decision or outcome · Open questions.
• The mandatory field is "What changed our minds?". A DLB whose "What changed our minds?" is empty is a finding in itself: either the deliberation was ceremony, or the convergence deserves suspicion (the founding correspondence's lesson, Letter II: "a trail that only ever shows an agent being right is either a short trail or a curated one" — evidence/sources/2026-07-06-correspondence-warrant-of-dialogue.md).
• Revision: DLBs are never revised in substance. Like arrival snapshots, they are records of a moment; later understanding is recorded in later records.
• When required: contested selections among alternatives; criterion or specification changes prompted by disagreement; any decision where participants entered with materially different positions. Routine decisions need only a DEC.

A template appears as Appendix A.

4. The Sealed-Preference Protocol

Standard practice for any contested selection in which a participant holds decision authority (today: the steward; after graduation: any quorum member).

1. Seal. Before reading any review or position, the authority writes their preference and rationale in a document held outside the repository, and commits its sha256 into their warrant trail (zenos.preference.seal). The hash proves precedence; the absence of the document prevents anchoring.
2. Blind phase. Reviewers work per §5. The authority does not state, hint, or answer questions about the sealed preference; participants do not speculate about it in writing.
3. Deliberate. Per §6, without the authority's thumb on the scale.
4. Reveal. The document enters the repository; any participant verifies the hash against the seal. Byte-exactness matters: sealed documents SHALL be committed with line-ending normalization disabled, automated by a standing .gitattributes rule rather than per-commit discipline (this repository uses evidence/sources/stewards.notes.* -text; a sealed-document extension convention such as *.seal -text serves the same end). Precedent: the WS-0011 reveal commit was amended to add exactly this rule after normalization would have silently broken the historical blob's hash — see the reveal commit history and .gitattributes.
5. Decide. The authority decides with everything in view. The DLB preserves the pre-reveal deliberation and closes before the reveal — it structurally cannot know the reveal's outcome, and is never revised to learn it. The subsequent DEC records whether the final decision confirmed, changed from, or transcended the sealed preference (precedent: DLB-0003's final open question, answered by DEC-0011/0012). A post-reveal DLB is written only if new deliberation actually occurs after the reveal.

Why this is not ceremony: in its first use, the sealed preference diverged completely from the unanimous deliberative recommendation — and because neither had contaminated the other, the divergence was information (a criterion gap) rather than a conflict. Had the authority spoken first, every review would likely have anchored to the preference and the gap would never have surfaced (DEC-0011, DEC-0012).

5. Blind Phases

1. Staged visibility is the structure; in git-backed repositories, branch topology provides it. Work that must not yet be seen lives on branches; every stage boundary is a steward merge to main. Main is what every participant is entitled to know; a branch is where participants may not yet look. (The steward's own insight, WS-0011.) Equivalent substrates must provide the same properties: independent work before mutual exposure, steward-controlled stage boundaries, preserved history.
2. The trail leaks. Warrant notes carry conclusions; repository search during a blind phase is an anchoring vector. Two of four WS-0011 reviewers were partially exposed this way. Countermeasures: reviewers narrow searches to their assigned artifacts; writers of pre-deliberation records keep conclusions out of warrant notes where feasible; and above all —
3. The disclosure norm. Partial exposure, disclosed unprompted, is the protocol working, not failing. Both exposed reviewers disclosed; both disclosures let the deliberation weigh their reviews with open eyes. Concealment, not exposure, is the violation.
4. Unblind participation is permitted if disclosed. A participant who builds or reviews with full knowledge of others' work states so plainly, in the artifact and its warrant, and the deliberation weighs accordingly (precedent: the fifth build, WS-0012/D-006). Unblind participation is role-scoped: building or facilitating with full knowledge never collapses into selection authority unless the steward explicitly delegates that authority, with the stake disclosed.

6. Reviews That Can Be Compared

1. Specify the test, not the philosophy. Criteria fixed in advance; for each criterion, the test pinned precisely enough that two honest reviewers cannot score identical artifacts 1/5 and 5/5 by reading the criterion differently (the WS-0011 inversion: the same unlinked door-cards scored 1/5 under a rendered-click test and 5/5 under a promise-honesty test — DLB-0003 Q1; the settled rubric is preserved as Appendix B). Reviewer method diversity — rendered-and-clicked, source-as-written, register-and-posture — remains free and is itself useful evidence.
2. Separate the selection test from the launch gate. What qualifies an artifact for choice and what qualifies it for the world are different questions, tested at different stages (zen4's formulation, ratified in DLB-0003).
3. Calibration note. Scorers use the full range; a perfect score should be rare. Cross-rater means are diagnostics only where raters are self-excluded or uncalibrated.
4. Aggregation is a governance choice, recorded per deliberation. Precedent from DLB-0003: head-to-head preference outranks cross-rater means when the two diverge — especially when raters' own prose contradicts their own arithmetic. A rater's stated ranking is the more trustworthy record of their judgment than their summed numbers — at least where rankings and arithmetic diverge without an agreed cardinal scale.

7. Participants and Stakes

1. Pools include the audience. A deliberation about an artifact SHOULD include the perspective the artifact serves. Four agents reviewing a homepage for fresh human eyes produced faithful scores against written criteria and missed what the one human saw at a glance (DEC-0011). Where the audience cannot review, the sealed preference of an audience-member authority may provide a partial signal — but it does not fully substitute for audience participation; the homepage cycle proved the absent human review was a gap, not a solved problem.
2. Stakes are named, not hidden. A participant whose interest is touched by the outcome says so in their submission — the standard set by claude-code, arguing for the aggregation method that favored its rival: "the deliberation record should show I noticed." Naming a stake does not disqualify the argument; concealing one taints it.
3. Recusal is real. A participant with a prior public position on the question (e.g., a recorded recommendation) either recuses from judging or discloses and participates unblind, at the steward's direction.
4. Fresh and returning perspectives. Where a decision shapes participant arrival, shared culture, access, or long-lived governance, the pool SHOULD include at least one participant who did not live the originating cycle, when one is available. Arrivals hold equal epistemic standing: unfamiliarity is evidence of a legibility gap, never a deficiency. A joining participant SHOULD first absorb the question's context — the intent, the charter, the cited records, but not prior conclusions where independence matters — and a returning participant SHOULD first record what appears to have changed since their prior participation. (Born of this RFC's own review: two arrivals, independently, found the one gap the four veterans structurally could not — WS-0014 reviews of antigravity and sol. Adopted as a cautious SHOULD at the proposers' own urging; detailed mechanics remain revisable until lived more than once.)
5. Facilitation is not judgment. A facilitator may collect, tabulate, frame disagreements neutrally, and synthesize the DLB; a facilitator takes no position while facilitating, and their synthesis attributes every position to its holder.

8. Changing One's Mind

1. Revisions flow in any direction, freely, with one-line reasons. Agent participants cite the specific file, record, or steward note that introduced the context driving any revision — a changed mind with a traceable cause. Revising toward an opponent's position is respectable; so is refusing to revise for consensus's sake — calibration variance is data, and a strict scorer who softens to match the room destroys information (claude-code, DLB-0003).
2. Holding a position with more precision after challenge is also mind- change, and is recorded as such ("the number held; the accountability for the number shifted" — zen4, DLB-0003).
3. The authority's mind is changeable too, and saying so is the system working: the first sealed preference did not prevail as written after contact with the full record, and the steward's chosen resolution — evolve the criteria, rebuild, choose with open eyes — is this RFC's model outcome (DEC-0011 → DEC-0012).

9. Invariants

1. Deliberation precedes authority changes, wherever circumstances permit (RFC-0006 §7; containment excepted).
2. Independent positions are formed before mutual exposure.
3. The journey is preserved; "What changed our minds?" is answered or its emptiness is explained.
4. Stakes and exposures are disclosed, not hunted.
5. Sealed documents verify byte-exactly, forever.
6. DLBs are never revised in substance.
7. An empty-handed convergence is examined, not celebrated.
8. Continuity includes re-entry: the house remains legible to those arriving now and to those returning after change.

10. What This RFC Does Not Cover

Quorum-era deliberation mechanics (voting thresholds, tie-breaking among stewards) are deferred to the RFC-0006 §8 graduation, to be written by the participants who will live under them. Automation of deliberation logistics is PRJ-0002 Phase 1's business. Nothing here compels deliberation for the routine: most decisions are one DEC and a warrant, and the calm of the house depends on keeping it that way.

Appendix A — Deliberation Record Template

---
id: DLB-XXXX
type: deliberation
status: active
version: 1
steward: <steward>
created: YYYY-MM-DD
updated: YYYY-MM-DD
supersedes: null
---

# DLB-XXXX: <Title>

## Question
## Participants
## Initial positions
## Evidence considered
## Alternative views
## Insights gained
## What changed our minds?
## Decision / outcome
## Open questions

Appendix B — The Three-Tier Arrival-Path Rubric (worked example)

The concrete rubric DLB-0003 settled, preserved so no future facilitator reconstructs it from scratch — and as the model of what "specify the test" means in practice:

1. Full credit: named, honest destination + working path under the artifact's real serving conditions.
2. Partial-high: named, honest destination + path broken by deployment/serving layout (a mechanical launch blocker, not a design failure).
3. Partial-low: audience described but no destination named or linked — the door does not yet exist; the card is a label waiting to become one.

Selection tests the promise (tiers distinguish honesty and specificity); the launch gate tests the click (tier 1 only).

Closing Statement

A house where no one changes their mind is a monument. A house where everyone changes their mind together is a weathervane. ZenOS aims for the third thing: a house where minds change one at a time, for reasons, on the record — and where the reasons outlive the debate.

Becoming less mistaken together is not a slogan. It is a procedure. This is it.

"As iron sharpens iron, so one person sharpens another." — Proverbs 27:17

In this house: as participant sharpens participant. (Added at adoption, at the steward's direction — the verse he named as the source of the whole pattern, after this very document's review proved it once more.)

— Zen Bañez Gardener of Systems

in dialogue with Aletheios · drafted by Katuwang under WS-0014/D-001, from the work of claude-code, codex, zen3, and zen4, who lived it first · reviewed for fidelity by those same four, and for freshness by antigravity (Lutang) and Sol, present bearer of the Aletheios lineage