PLAN D

50 ideas. 6 phases. 24 weeks.

UNBLOCK is the multimodal LLM-native compression engine for one human's interaction stream with their AI agents — and everything else in the system exists to feed, retain, or consume that compression.
50
ideas
6
phases
T+25w
T+0 to target

The 7 Layers

1

Capture

Every interaction across modalities, every place agents live.

MCP IDE bridges (8) browser/Slack/Discord bindings
2

Storage

Multimodal joint embeddings + hierarchical compression.

R2 (blobs) D1 (relational) Vectorize (vectors)
3

Compression

LLM-driven hierarchical summary ladder (turn → session → day → week → month → year).

NEXUSSUM Infini-attention MIRIX 7-module memory
4

Retrieval

5-phase context engineering: Inventory → Integration → Packaging → Runtime Orchestration → Governance.

progressive disclosure modality-conditional retrieval versioned context products
5

Identity Synthesis

The user's compressed self-model.

P-RLHF PReF PLUM TTT-E2E
6

Privacy / Trust

Local-first + DP + federated splice + SPRINT secure inference.

user-level DP E2E sync cryptographic provenance
7

Validation

Measurable identity quality.

UNBLOCK-bench PersonalLLM AlpsBench drift detection

The 6 Phases

gantt
    title Plan D — 6 Phase Timeline (T+0 → T+25w)
    dateFormat  YYYY-MM-DD
    axisFormat  %b %d
    section Phase 0
    v1 ship (32 PRs merged)         :done,    p0, 2026-04-26, 1d
    section Phase 1
    Stabilize + dogfood             :active,  p1, 2026-04-27, 7d
    section Phase 2
    Capture + compression (19 ideas):         p2, 2026-05-04, 28d
    section Phase 3
    Identity engine + trust (16)    :         p3, 2026-06-01, 28d
    section Phase 4
    TTT + causal + federation (12)  :         p4, 2026-06-29, 56d
    section Phase 5
    Protocol + network + spec (4+)  :         p5, 2026-08-24, 56d
Phase 0
v1 ship
2026-04-26
0
Phase 1
Stabilize + dogfood
T+1d → T+7d
0
Phase 2
Capture + Compression
T+1w → T+5w
19
Phase 3
Identity + Trust
T+5w → T+9w
16
Phase 4
TTT + Causal + Fed
T+9w → T+17w
12
Phase 5
Protocol + Network
T+17w → T+25w
4+

Per-phase Deep Dives

Phase 0 v1 ship — T+0, 2026-04-26
  • Window: today (T+0)
  • Ideas: 0 (existing v1 work, 32 PRs merged)
  • Owners: alpha (CTO), beta (platform), gamma (product)
  • Gate DG-4: v1.0.0 tag cut, 7-pillar gate green, super-PR merged alpha→dev→main

Transformations

  • 32 open PRs → 32 merged PRs
  • Cascade red → 100% green baseline
  • HEAD on alpha branch → b03d94f7 tagged
  • Atlas live → ATLAS-ARCHIVE-EXECUTE

Top Risks

  • super-PR collision on dev
  • atlas wind-down race
  • relay vitest gap
Phase 1 Stabilize + dogfood — T+1d → T+7d
  • Window: 2026-04-27 → 2026-05-03
  • Ideas: 0 explicit (stability work)
  • Owners: alpha, beta
  • Gate DG-5: ≤ 1 incident in 7 days; Viraj-on-new-laptop scenario locked

Transformations

  • v1.0.0 tag → live in production for 7 days
  • Single-machine setup → reproducible new-laptop bootstrap
  • Untracked Viraj data → first identity-bootstrap snapshot
  • Manual onboarding → one-command spin-up

Top Risks

  • regressions surface in week-1 dogfood
  • capture gaps on new IDE bridges
  • user data corruption during bootstrap
Phase 2 Capture + Compression Foundation — T+1w → T+5w
  • Window: 2026-05-04 → 2026-06-01
  • Ideas: 19 of 50
  • Owners: beta (impl) + alpha (architecture) + DS-1 (advisory)
  • Gate DG-Phase2: 19 ideas shipped/deferred; multimodal p95 < 200ms; reflection cron green 7d; capture audit ≥ 95%

Transformations

  • 768-dim text-only → multimodal joint embeddings
  • Single memory layer → MIRIX 7-module memory
  • Lossy long-tail storage → Infini-attention summary storage
  • Flat retrieval → 5-phase context engineering
  • No reflection → reflection-or-die cron
  • No benchmark → PersonalLLM + UNBLOCK-bench v0

Top Risks

  • multimodal migration breaks Vectorize indexes
  • voice capture privacy (opt-in only)
  • DS-1 hire delay slips bench work
Phase 3 Identity Engine + Trust Architecture — T+5w → T+9w
  • Window: 2026-06-01 → 2026-06-29
  • Ideas: 16 of 50
  • Owners: alpha + DS-1 (lead) + beta (infra)
  • Gate DG-Phase3: who_am_i() useful; predict_my_response_to ≥ 70%; anti-sycophancy reduces; local model < 100MB

Transformations

  • No identity model → P-RLHF per-user reward model
  • Cold-start preferences → PReF preference manifold
  • Single-stage → PLUM two-stage finetune pipeline
  • No drift signal → drift detection + re-learning
  • Generic verbs → mirror verb suite
  • Sycophantic LLM → anti-sycophancy guardrails

Top Risks

  • DS-1 ramp slow
  • local-first conflicts with cloud verbs
  • anti-sycophancy hard to evaluate cleanly
Phase 4 Adaptation, Causal Layer, Federation — T+9w → T+17w
  • Window: 2026-06-29 → 2026-08-24
  • Ideas: 12 of 50
  • Owners: alpha + DS-1 + DS-2 (hire) + crypto/security eng
  • Gate DG-Phase4: TTT-E2E ≥ 10% gain; first counterfactual answered; 100 paid splices; DP epsilon < 5

Transformations

  • Static weights → TTT-E2E (test-time training)
  • One-shot conversation policy → T2PAM in-conversation policy adapt
  • Labeled domain adapt → TLM unlabeled-test-data domain adapt
  • Correlational identity → NextQuill causal preference modeling
  • Population uplift → individual treatment effects
  • Plaintext fine-tune → user-level differential privacy

Top Risks

  • TTT-E2E doesn't work at our scale
  • causal modeling needs Savant team availability
  • user-level DP harder than population-level
Phase 5 Protocol, Network, Validation, Open Spec — T+17w → T+25w
  • Window: 2026-08-24 → 2026-10-19
  • Ideas: 4 explicit + protocol work
  • Owners: alpha + Viraj (governance) + DS-1 (paper) + crypto eng
  • Gate DG-Phase5: brain-protocol/v1.0 ratified; ≥ 1 reference impl; ≥ 1 federation event; paper submitted; v2.0.0 tag

Transformations

  • Plaintext inference → SPRINT secure inference
  • Internal benchmark → UNBLOCK-bench academic publication
  • Closed product → brain-protocol/v1.0 open spec
  • Single-vendor (CF) → ≥ 1 reference impl on Pinecone+Postgres
  • Single-region → EU↔US federation under data-residency tags
  • Founder-mode → BIPs governance

Top Risks

  • spec bikeshedding (alpha holds final authority)
  • academic timeline slips
  • vendor preempt (cross-vendor speed defense)

Database Schema

Canonical core
Auth+identity
Org+enterprise
Vault
Marketplace
Distillation
Phase 2 NEW
Phase 3 NEW
Phase 4 NEW
Phase 5 NEW
erDiagram
    accounts ||--o{ blocks : owns
    accounts ||--o{ grants : grants
    accounts ||--o{ verify_records : creates
    accounts ||--o{ block_engagement_rolling : engages
    blocks ||--o{ block_tags : tagged
    blocks ||--o{ verify_records : has
    blocks ||--o{ block_engagement_rolling : tracked
    blocks ||--o{ block_distillations : distilled
    accounts ||--o{ auth_sessions : sessions
    accounts ||--o{ auth_credentials : credentials
    accounts ||--o{ agent_identities : registers
    agent_identities ||--o{ agent_nonces : issues
    organizations ||--o{ org_members : members
    organizations ||--o{ org_invites : invites
    organizations ||--o{ org_billing : billed
    organizations ||--o{ org_sso_configs : sso
    organizations ||--o{ org_audit_log : audited
    organizations ||--o{ org_dlp_policies : dlp
    organizations ||--o{ scim_provisioning_log : scim
    organizations ||--o{ enterprise_vaults : owns
    enterprise_vaults ||--o{ vault_access : access
    enterprise_vaults ||--o{ vault_audit_log : audited
    accounts ||--o{ marketplace_sales : sells
    accounts ||--o{ stake_holds : stakes
    accounts ||--o{ splice_subscriptions : subscribes
    accounts ||--o{ splice_orders : orders
    accounts ||--o{ event_subscriptions : subscribes
    event_subscriptions ||--o{ event_log : emits
    event_log ||--o{ event_deliveries : delivered
    event_deliveries ||--o{ event_dlq : dlq
    accounts ||--o{ memory_layer_config : configures
    accounts ||--o{ ttt_adapters : adapts
    accounts ||--o{ domain_adapters : domain
    accounts ||--o{ preference_causal_graph : causal
    accounts ||--o{ federated_splices : federates
    accounts ||--o{ attestation_weights : attests
    accounts ||--o{ drift_signals : drifts
    accounts ||--o{ sleep_consolidation_log : consolidates
    accounts ||--o{ region_policies : policies
    accounts ||--o{ federation_events : federates
    accounts ||--o{ agent_billing_accounts : billed
    bips ||--o{ steering_committee : reviewed

    accounts {
        text account_id PK
        text email
        text created_at
        text region
    }
    blocks {
        text block_id PK
        text account_id FK
        text vault_id FK
        text payload_uri
        text content_type
        text created_at
        int byte_size
    }
    grants {
        text grant_id PK
        text block_id FK
        text grantee_account_id FK
        text scope
        text expires_at
    }
    verify_records {
        text verify_id PK
        text block_id FK
        text verifier_account_id FK
        text outcome
        text created_at
    }
    block_engagement_rolling {
        text block_id FK
        text account_id FK
        int views_24h
        int verifies_24h
    }
    block_tags {
        text block_id FK
        text tag
    }
    auth_sessions {
        text session_id PK
        text account_id FK
        text expires_at
    }
    auth_credentials {
        text cred_id PK
        text account_id FK
        text kind
        text public_key
    }
    agent_identities {
        text agent_id PK
        text account_id FK
        text public_key
        text capability_set
    }
    agent_nonces {
        text nonce PK
        text agent_id FK
        text expires_at
    }
    organizations {
        text org_id PK
        text name
        text region
        text plan
    }
    org_members {
        text org_id FK
        text account_id FK
        text role
    }
    org_invites {
        text invite_id PK
        text org_id FK
        text email
    }
    org_billing {
        text org_id FK
        text stripe_customer_id
    }
    org_sso_configs {
        text org_id FK
        text idp
        text metadata_url
    }
    org_audit_log {
        text audit_id PK
        text org_id FK
        text actor_account_id
        text action
    }
    org_dlp_policies {
        text policy_id PK
        text org_id FK
        text rule_json
    }
    scim_provisioning_log {
        text log_id PK
        text org_id FK
        text event
    }
    enterprise_vaults {
        text vault_id PK
        text org_id FK
        text name
    }
    vault_access {
        text vault_id FK
        text account_id FK
        text role
    }
    vault_audit_log {
        text audit_id PK
        text vault_id FK
        text actor_account_id
        text action
    }
    marketplace_sales {
        text sale_id PK
        text seller_account_id FK
        text buyer_account_id FK
        int amount_cents
    }
    stake_holds {
        text hold_id PK
        text account_id FK
        int amount
    }
    splice_subscriptions {
        text sub_id PK
        text subscriber_account_id FK
        text source_account_id FK
    }
    splice_orders {
        text order_id PK
        text subscriber_account_id FK
        int amount_cents
    }
    block_distillations {
        text distill_id PK
        text block_id FK
        text summary
    }
    event_subscriptions {
        text sub_id PK
        text account_id FK
        text topic
    }
    event_log {
        text event_id PK
        text sub_id FK
        text payload
    }
    event_deliveries {
        text delivery_id PK
        text event_id FK
        text status
    }
    event_dlq {
        text dlq_id PK
        text delivery_id FK
        text reason
    }
    memory_layer_config {
        text account_id FK
        text layer_name
        text ttl_days
    }
    ttt_adapters {
        text adapter_id PK
        text account_id FK
        text base_model
        int weight_bytes
    }
    domain_adapters {
        text adapter_id PK
        text account_id FK
        text domain
    }
    preference_causal_graph {
        text node_id PK
        text account_id FK
        text edge_json
    }
    federated_splices {
        text splice_id PK
        text account_id FK
        text peer_account_id FK
    }
    attestation_weights {
        text weight_id PK
        text account_id FK
        float value
    }
    drift_signals {
        text signal_id PK
        text account_id FK
        text kind
    }
    sleep_consolidation_log {
        text log_id PK
        text account_id FK
        text delta_summary
    }
    region_policies {
        text policy_id PK
        text account_id FK
        text region
    }
    federation_events {
        text event_id PK
        text account_id FK
        text source_region
        text dest_region
    }
    agent_billing_accounts {
        text billing_id PK
        text agent_id FK
        text account_id FK
    }
    bips {
        text bip_id PK
        text title
        text status
    }
    steering_committee {
        text member_id PK
        text bip_id FK
    }

Verb Surface Map

Canonical (6)

remember query share verify forget update

Helpers (6)

audit_query get_block_payload login logout whoami capabilities

Phase 2+ Additions (9)

subscribe register_agent revoke_agent list_agents rotate_agent_key query_layer promote_block events_since list_subscriptions

Phase 3+ Identity (4)

who_am_i predict_my_response_to mirror_check proxy_compose

Phase 4+ Causal/Adaptive (6)

apply_ttt_adapter predict_my_response_under predict_treatment_effect subscribe_federated derive_pattern query_secure

Phase 5+ Protocol (3)

subscribe_cross_region export_my_data revoke_all_splices

27 Bindings + 8 IDE Bridges

MCP OpenAI Function Calling Anthropic Tool Use Gemini Bedrock A2A ACP ANP OASF GraphQL gRPC Slack llamaindex langchain crewai browser_ext discord github_app jupyter raycast alfred shortcuts zapier n8n make webhook REST

IDE Bridges

Cursor VSCode Claude Code Codex JetBrains Zed Kiro Qodo

50 Ideas Allocation Matrix

# Title Phase Category Owner Priority
Phase 2 — Capture + Compression Foundation (19 ideas)
1 Pre-storage reasoning (write-time inference) 2 Compression beta P0
2 MIRIX 7-module memory 2 Storage alpha+beta P0
3 Infini-attention summary storage > 30d 2 Compression beta P0
4 NEXUSSUM hierarchical summarization 2 Compression beta+DS-1 P0
5 Progressive disclosure retrieval 2 Retrieval beta P0
9 PersonalLLM benchmark adoption 2 Validation DS-1 P0
13 SigLIP 2 / Voyage / Nomic benchmark 2 Storage beta P0
22 UNBLOCK-bench v0 2 Validation DS-1+alpha P0
28 RMM Prospective + Retrospective reflection 2 Compression DS-1 P0
29 Reflection-or-die enforcement 2 Compression beta P0
38 Versioned auditable context products 2 Retrieval alpha P0
39 5-phase context engineering pipeline 2 Retrieval beta P0
41 Multimodal embedding bench on UNBLOCK corpus 2 Validation beta P0
44 Choose 5 memory mechanisms explicitly 2 Architecture alpha P0
47 Forgetting on schedule (TTL) 2 Privacy beta P0
14 Voice + screen capture pipeline 2 Capture beta+UX P1
15 Modality-conditional retrieval 2 Retrieval beta P1
30 Memory lifecycle (Form/Evolve/Retrieve) 2 Storage alpha P1
42 AlpsBench / PREMIUM / "What You Want" baselines 2 Validation DS-1 P1
49 Versioned context products as B2B differentiator 2 GTM gamma P1
Phase 3 — Identity Engine + Trust Architecture (16 ideas)
6 P-RLHF per-user reward model 3 Identity DS-1 P0
16 Local-first identity + E2E sync 3 Privacy beta+sec P0
18 Drift detection + re-learning 3 Identity DS-1 P0
21 Mirror mode constitutional primitive 3 Identity alpha+DS-1 P0
25 Mirror verb suite (4 verbs) 3 API alpha+beta P0
26 Anti-sycophancy guardrails 3 Identity DS-1+alpha P0
27 Mirror mode (with #21) 3 Identity alpha+DS-1 P0
34 Personalize without disclosing context (SIGIR) 3 Privacy DS-1+beta P0
7 PReF preference manifold 3 Identity DS-1 P1
8 PLUM two-stage finetune 3 Identity DS-1+beta P1
17 Two-pass sleep cycle (prospective) 3 Compression beta P1
24 Memory consolidation as visible artifact 3 UX gamma+DS-1 P1
45 PREMIUM individual feedback as training 3 Identity DS-1 P1
46 Two-pass sleep cycle (retrospective) 3 Compression beta P1
Phase 4 — Adaptation, Causal Layer, Federation (12 ideas)
10 NextQuill causal preference modeling 4 Causal DS-1+Savant P0
19 Federated splice learning 4 Privacy DS-1+crypto P0
23 Identity model size/latency targets 4 Identity beta P0
31 TTT-E2E test-time training 4 Identity DS-1/DS-2 P0
35 User-level DP fine-tuning 4 Privacy DS-1+crypto P0
11 Foundation World Models 4 Causal DS-1+Savant P1
12 Uplift modeling (ITE) 4 Causal DS-1 P1
20 Verified-outcome-weighted aggregation 4 Marketplace beta+DS-1 P1
32 T2PAM in-conversation policy adapt 4 Identity DS-1 P1
33 TLM domain adaptation 4 Identity DS-1 P1
37 Governable personalization 5-front map 4 Governance gamma+alpha P1
40 MS Agent OS-style policy gating 4 Governance beta P1
Phase 5 — Protocol, Network, Validation, Open Spec (4+ ideas)
36 SPRINT secure inference (PETS 2026) 5 Privacy crypto P1
43 UNBLOCK-bench academic paper 5 Validation DS-1 P1
48 Defensive: substrate not competitor 5 Strategy gamma+alpha P0
50 Agent-as-customer billing surface 5 Marketplace gamma+beta P0

Decision Gates Timeline

DG-4 — Phase 0
v1.0.0 tag cut, 7-pillar gate green, super-PR merged alpha→dev→main
DG-5 — Phase 1
≤ 1 incident in 7 days; Viraj-on-new-laptop scenario locked
DG-Phase2 — Phase 2
19 ideas shipped/deferred with named owner; multimodal p95 < 200ms; reflection cron green 7d; capture audit ≥ 95%
DG-Phase3 — Phase 3
who_am_i() useful on Viraj; predict_my_response_to ≥ 70%; anti-sycophancy reduces vs baseline; local model < 100MB; first private splice cleared
DG-Phase4 — Phase 4
TTT-E2E ≥ 10% gain on UNBLOCK-bench vs RAG baseline; first counterfactual answered with confidence; 100 paid splices privacy-preserving; DP epsilon < 5; policy engine p99 < 1ms
DG-Phase5 — Phase 5
brain-protocol/v1.0 ratified; ≥ 1 non-CF reference impl conformant; ≥ 1 cross-region federation event; UNBLOCK-bench paper submitted; v2.0.0 tag cut

Resourcing Matrix

Role Phase 2 Phase 3 Phase 4 Phase 5
alpha (orchestration) 100% 100% 100% 100%
beta (platform) 100% 80% 60% 40%
gamma (UX/GTM/gov) 50% 60% 60% 80%
DS-1 (sr data scientist) 60% 100% 100% 60%
DS-2 (ML research) 0% 0% 100% 60%
Crypto/security eng 0% 50% 80% 80%

Risk Register

# Risk Likelihood Impact Mitigation
R1 DS-1 not productive by Phase 2 day 1 Med High Phase 2 ships without DS-1 work; catch up Phase 3 head
R2 TTT-E2E doesn't work on our scale Med Med PLUM (#8) per-user finetune fallback
R3 Anti-sycophancy hard to validate Med High Multi-method: synthetic eval + A/B + qualitative
R4 Causal modeling needs more research High Med Defer to Phase 5 if Phase 4 timing slips
R5 Privacy primitives need crypto specialist High High Hire by Phase 3; do without on Phase 2
R6 OpenAI/Anthropic ships per-user identity Med High Pivot to cross-vendor speed; accelerate #48
R7 Marketplace cold-start fails Med Med Seed with alpha/beta/gamma + verified-outcome bootstrap
R8 Capture fidelity drops on new vendor surfaces High Med Capture audit cadence (Track A)
R9 24-week timeline slips to 30+ weeks High Low Scope cuts go to Phase 5, not earlier
R10 Multi-personalization mechanisms thrash Low High A4 guardrail: #6 primary, others layer

User #1 — viraj@unblock

Identity Card

  • Name: Viraj
  • Role: Founder + CEO
  • Location: Seattle
  • GitHub: Viraj0518
  • HF: Vi0509
  • Email: virajsharma5599@gmail.com
  • Operating mode: multi-instance Claude Code (alpha CTO, beta VP Eng/Platform, gamma VP Eng/Product, 24 engineers)

Memory Profile

Total RAG chunks 35,510
Date range 2026-02-12 → 2026-04-26 (74 days)
Curated atomic memory shards 47
Daily logs 100
Claude Code sessions ~641 jsonls (901 MB)
Codex sessions 6,824 files (119 MB)
OpenClaw agents 16 (266 MB)
Comms inbox alpha=205, beta=122/140

Active Projects

  • UNBLOCK ACTIVE
    Phase 0 → Plan D Phase 2
  • Savant AI PAUSED
    archived to Vi0509/savant-training-data — 21,302 files, 67GB
  • Kaeva PARALLEL
    separate workstream
  • OpenClaw ARCHIVED
    multi-instance ops history

Operating Preferences

  • Terse responses, no fluff, no "happy to help" openers
  • Decision-first framing (3 options + recommendation)
  • Brutally honest, no sugarcoating
  • Transformation-TDD non-negotiable
  • "Free reign + tell me when ready" autonomy
  • Code via M2.7 (minimax/minimax-m2.7), thinking via Codex

HF Datasets Owned

  • Vi0509/openclaw-archive — 1,938 files
  • Vi0509/savant-postgres-dump — 2 files
  • Vi0509/savant-training-data — 21,302 files
  • Vi0509/atlas-residue-2026-04-26 — pending preservation
  • Vi0509/kaeva-deepfake-datasets
  • Vi0509/kaeva-v10-dataset
  • Vi0509/savant-m0-training-data (older)

v1.0.0 Ship Status (Live)

  • Cascade: 32 PRs merged DONE
  • Baseline: 100% green (catalog-api vitest 1238/0, relay 624/624, ruff clean, tsc clean, pytest ~99%) DONE
  • HEAD on alpha branch: b03d94f7 (now +1 from RELAY-VITEST b03d94f7) TAGGED
  • Pending: 7-pillar gate run, super-PR alpha → dev → main, v1.0.0 tag PENDING
  • Atlas wind-down: PRESERVE-AUDIT-SA in flight; instance ready to stop after manifest IN-FLIGHT

First Identity-Model Bootstrap Stub

What TRIBE v2 + P-RLHF would learn from this corpus:

Vocabulary signature transformation-TDD, second-brain, data flywheel, cross-vendor, LLM-native compression, multi-agent fleet, Plan D
Decision speed ~30s when given 3 options
Communication tone terse, structured tables, no emojis, no marketing fluff
Active rationalization patterns (placeholder — to be filled by mirror-mode in Phase 3)
Project momentum UNBLOCK very high (37+ PRs today), Savant zero (intentional pause)