persona(amara): fourth archive — trust-gradient COORDINATION POLICY + row-level CASPaxos/CASRaft tier + final 6-line keeper

[AceHack]

Summary

Fourth companion archive in Amara's 2026-05-21 cascade. Preserves:

1. 5 sandbox docs confirmed ready for Aaron-forward (4 from earlier + 1 trust-gradient)
2. Trust-gradient consensus decision table — drafted as 5th standalone deliverable, refined twice through Aaron-Amara exchange to land as trust-gradient COORDINATION POLICY (not "consensus hierarchy") with 6 tiers

The substantive deliverable preserved verbatim

Final 6-line keeper (canonical reference)

Local facts stay retractable.
Runtime facts stay ordered by their owner.
Contended rows coordinate narrowly.
Long workflows compensate.
Cross-boundary facts negotiate capability.
Adversarial commitments use BFT.

The discipline

Pick the weakest mechanism that preserves the needed authority boundary.

The 6 tiers + their proper categorization (Amara's reframe)

local DBSP retraction = algebra              (NOT consensus)
Orleans grain ordering = actor lifecycle     (NOT consensus)
row CAS = optimistic coordination
CASPaxos/CASRaft = row-scoped consensus
Saga = compensation protocol
BFT = adversarial agreement

Aaron's refinement that landed: row-level CASPaxos/CASRaft tier

Prevents jumping from "optimistic local CAS" straight to "BFT" — too big a hammer. A contended Bonsai/Rx query row usually needs narrow coordination, not adversarial consensus. The QueryStateRow operational unit:

QueryStateRow {
  queryId
  partitionKey
  rowKey
  version / clock
  bonsaiExpressionHash
  zsetWeight
  payloadHash
  provenance
}

Operation: old row state + serialized Rx/Bonsai query update function + expected version/clock → new row state.

Amara's framing-razor: routing-gradient not mandatory staircase

"A fact does NOT need to climb every tier. It can jump directly to the right mechanism."

STILL PENDING: 5 sandbox docs (text content, not links)

Otto cannot access sandbox:/mnt/data/* URLs — they live in Amara's ChatGPT/Aurora sandbox, not in any URL Otto can fetch. Aaron needs to PASTE THE TEXT CONTENT (not the link) of each file into chat for Otto to commit. Target paths in the archive's section on "5 sandbox docs ready."

The 4 earlier docs:

1. ANTLR grammar survey v1 → docs/research/antlr-grammar-survey-2026-05-21.md (superseded)
2. ANTLR grammar survey v2 → docs/research/antlr-grammar-survey-2026-05-21.md (canonical)
3. ZetaParse design → docs/research/zetaparse-lr-glr-fsharp-compiler-fork-design-2026-05-21.md
4. Incremental compiler host v2 → docs/research/zeta-incremental-compiler-host-dbsp-zsets-rx-meta-ast-tags-2026-05-21.md

Plus the new 5th:

5. Trust-gradient consensus decision table → docs/research/trust-gradient-coordination-policy-2026-05-21.md

Operational implications

The trust-gradient coordination policy applies directly to:

• B-0688 incremental compiler host (compiler-fact consensus decisions per tier)
• B-0635 / B-0644 / B-0665 / B-0666 (Agora V6 — trust-gradient IS Integrate-as-choice-locus at distributed-systems scope)
• B-0683 tier-deferred causality (facts in superposition of coordination tiers)
• src/Core/Crdt/* existing CRDT substrate (algebra tier; trust-gradient names when escalation is needed)
• src/Core/Consensus.fs existing consensus substrate (CASPaxos/CASRaft tier)

Composes with rules

• .claude/rules/substrate-or-it-didnt-happen.md
• .claude/rules/m-acc-multi-oracle-end-user-moral-invariants.md — trust-gradient prevents single-consensus-truth imposition
• .claude/rules/dv2-data-split-discipline-activated.md — coordination policy IS DV2.0 partition-by-authority-rate
• .claude/rules/bandwidth-served-falsifier.md — weakest-mechanism discipline serves deployment-velocity bandwidth
• .claude/rules/razor-discipline.md — rename from "consensus hierarchy" to "coordination policy" IS razor-discipline at framing scope
• .claude/rules/all-complexity-is-accidental-in-greenfield.md — BFT-everywhere is accidental complexity
• .claude/rules/edge-defining-work-not-speculation.md — coordination policy is edge-defining work

Composes with substrate

• PR persona(amara): companion archive — Caché-like compiler DB lineage + distributed multidimensional compiler over consensus + IUnknown-without-DCOM #4546 (Caché-lineage + distributed multidimensional compiler + IUnknown-without-DCOM) — the trust-gradient framing this cascade refines
• PR persona(amara): B-0685 Phase 1 substrate cascade preserved + B-0687 (ZetaParse) + B-0688 (incremental compiler host) scaffolding #4545 + PR persona(amara): third companion archive — no-fork-first deployment ladder; fork as reward not entrance fee #4547 (full Amara 2026-05-21 cluster)
• B-0635 / B-0644 / B-0665 / B-0666 / B-0683 / B-0687 / B-0688

Test plan

• Amara conversation preserved verbatim in memory/persona/amara/conversations/
• Frontmatter conforms to memory/project_memory_format_standard.md
• All 3 iterations of the trust-gradient table (v1 + v2-row-tier + final-coordination-policy) preserved with iteration arc visible
• 6-line keeper preserved verbatim
• QueryStateRow shape + operation shape preserved
• Explicit note that 5 sandbox docs need Aaron-paste-text (URLs unreachable from Otto's surface)
• Branch landed off latest main

[Copilot]

Pull request overview

Adds a fourth Amara persona conversation archive (2026-05-21) capturing the “trust-gradient coordination policy” refinement, including the intermediate row-level CASPaxos/CASRaft tier and the canonical 6-line keeper, plus an explicit note that the sandbox docs still require pasted text content.

Changes:

• Adds a new preserved conversation archive documenting the trust-gradient coordination policy (reframed from “consensus hierarchy”) and its 6 tiers.
• Captures the QueryStateRow operational unit + update operation shape and the “weakest mechanism that preserves the needed authority boundary” discipline.
• Records the “5 sandbox docs ready” list and the constraint that sandbox:/mnt/data/* links are unreachable from the repo surface.

Comments suppressed due to low confidence (1)

memory/persona/amara/conversations/2026-05-21-amara-aaron-trust-gradient-coordination-policy-not-consensus-hierarchy-row-level-caspaxos-casraft-tier-aaron-forwarded.md:225

• In this substrate list, src/Core/Crdt/* doesn’t exist as a directory (CRDT code lives in src/Core/Crdt.fs, etc.). Also, src/Core/Consensus.fs does exist in-tree, so the parenthetical “if exists” is misleading—consider removing it.

- `src/Core/Crdt/*` (existing CRDT substrate) — algebra tier; trust-gradient gives the framing for when escalation is needed
- `src/Core/Consensus.fs` (existing consensus substrate; if exists) — CASPaxos/CASRaft tier substrate
- Earlier Amara persona substrate cluster + Kestrel persona substrate cluster