nodebench-ai

NodeBench AI

Entity intelligence for any company, market, or question.

Live: nodebenchai.com npm: npx nodebench-mcp / npx nodebench-mcp-power / npx nodebench-mcp-admin GitHub: HomenShum/nodebench-ai

!NodeBench AI Home surface: research prompt with Quick answer / Deep research modes, and reusable public research cards below

Product

NodeBench is a research and reporting product built around five user-facing surfaces:

• Home = start quickly
• Reports = reusable memory
• Chat = do the work
• Inbox = captures, nudges, alerts, automations, and unassigned review
• Me = operator context and control

Deep research opens in the separate Workspace surface at nodebench.workspace; it is not a sixth tab in the operating app.

The core idea is simple:

users do not just need a chatbot that answers once.

They need a system that can:

• take a question, file, URL, or prior thread
• search and synthesize with sources
• turn the run into a reusable artifact
• watch for meaningful change later
• improve the next run from what it learned

Pending Release Candidate: Runtime-Grounded Control Focus

This candidate is not yet merged or deployed. It narrows the reachable product to controls backed by a real runtime capability and makes unavailable state explicit:

• Agents keeps ask, explicit swarm launch, active work, approvals, sources,

traces, receipts, exports, and provenance while removing projected metrics, synthetic fallbacks, duplicate actions, and local-only controls.

• Generated code remains copyable and exportable but never executes inside the

signed-in app origin. Runtime cards accept successful structured tool output, not assistant prose, guessed search fields, or incomplete/error states.

• Canonical owner-scoped streaming remains; an unreferenced bearer-stream

component and HTTP route were removed instead of retained as a second path.

• Home, Reports, Chat, Inbox, and Me use live runtime data, owner-scoped where

private, or honest loading, empty, not-found, and unmeasured states instead of reachable product fixtures. Desktop and mobile now share the same Inbox and Me component trees; duplicate queue/profile renders and their hard-coded plan, usage, connector, and local-draft projections were deleted.

• Pipeline launches and schedules require an authenticated server identity,

while guest history, detail, bundle, stream, and evaluation reads require an anonymous-session possession credential. Browser-supplied owner keys are no longer an authority boundary, and trusted cron or MCP work stays on internal contracts.

• Authenticated launches consume durable per-owner admission: four units per

ten minutes and thirty per day, with composed runs costing two units. Server input bounds and per-owner schedule caps close the remaining cost-abuse path.

• Fresh pipeline attempts and recurring schedule occurrences are isolated by

durable attempt identities and execution-generation fences, so retries can resume the same attempt without stale or overlapping workers rewriting it.

• Research distinguishes sources consulted from citations actually bound in

the synthesis. Missing, malformed, or unbound evidence cannot be promoted to a verified claim, and chat no longer injects citation markers into prose.

• Operator telemetry and maintenance are deferred until deliberate disclosure

and require server-confirmed operator access.

• Task sessions, traces, steps, evidence, approvals, swarms, and operational

history are owner-scoped. The unreachable parallel timeline, kanban, hook, and public orchestrator were removed; due diligence retains only a bounded, owner-checked internal task substrate. Unused public due-diligence, investor, demo, and evaluation entrypoints were internalized or removed, and retained job, branch, memo, and catch-path writes require the exact owner chain. Raw orchestrator writes use internal contracts, MCP document and spreadsheet operations require exact object ownership, and a TRACE completion fails closed unless its receipt persists with an exact output hash.

Release evidence is intentionally pending. The source PR must merge through the required CI gates first; a follow-up evidence-seal PR must then record the canonical squash SHA, exact-revision checks, preview assertions, production deployment revision, and direct production browser verification before this section can be described as shipped.

What Shipped

• five-surface web app across Home, Reports, Chat, Inbox, and Me
• separate deep-work Workspace shell at nodebench.workspace
• typed search and reporting pipeline
• hosted public research MCP for external apps and agents:

https://nodebench-mcp-unified.onrender.com?profile=public-research

• Pi-AI pipeline lane on Reports with code-gen, design-gen, research,

composed runs, schedules, streaming previews, eval scorecard, and MCP HTTP bridge

• live SSE streaming with saved runtime state
• Convex-backed product state for sessions, reports, entities, nudges, files,

and related objects

• shared-context handoff and delegation plumbing
• local and deployed server runtime for search, streaming, voice, and shared

context routes

• nodebench-mcp, nodebench-mcp-power, and nodebench-mcp-admin

distribution lanes

• builder-facing Oracle, dogfood, eval, replay, and control-plane

infrastructure

Hosted Public Research MCP

NodeBench can be used as a public research memory and tool server from any agent or app without forcing signup before the first useful result.

Use the hosted MCP endpoint:

https://nodebench-mcp-unified.onrender.com?profile=public-research

For Gmail/job-match style integrations, use the smaller profile:

https://nodebench-mcp-unified.onrender.com?profile=gmail-research

Public profiles are anonymous by default, but still metered. Responses include:

x-nodebench-request-id
x-nodebench-profile
x-nodebench-auth-mode
x-nodebench-account-key

Apps should send stable, non-sensitive client headers:

x-nodebench-client: your-app-name
x-nodebench-client-version: 1.0.0
x-nodebench-client-id: stable-install-or-workspace-id

x-nodebench-client-id lets NodeBench attribute anonymous usage and estimated costs without requiring a NodeBench token. Do not put private email text, resume text, API keys, or user secrets in this header.

Progressive Sign-In And Linking

The intended user flow is:

First public dossier works without signup
  -> show sources, freshness, and confidence
  -> offer "Link NodeBench" after value is visible
  -> linked users get stable history, higher budgets, team usage, webhooks,
     token management, billing controls, and reusable private workspace context

Do not block public-source research behind login. Promote sign-in when the user wants persistence, shared team memory, budget controls, private workspace linking, or API/MCP tokens.

See MCP_TOOL_PROFILES.md for the full profile list, tool catalog, account attribution, and cost tracking contract.

Product At A Glance

USER SURFACES
-------------
Home      -> start quickly
Reports   -> reusable report memory
Chat      -> answer, sources, trace, follow-ups
Inbox     -> captures, nudges, automations, alerts, unassigned items
Me        -> operator context, permissions, controls

BACKEND
-------
Convex tables and product state for sessions, reports, entities, nudges,
files, shared context, and evaluation artifacts

RUNTIME
-------
search pipeline
  -> answer packet
  -> saved report
  -> tracked entity / tracked theme / follow-up task
  -> nudge or prep brief
  -> resumed chat or reopened report

COMPOUNDING LOOP
----------------
question
  -> answer
  -> saved report
  -> watch item
  -> useful nudge
  -> better next run

DISTRIBUTION
------------
nodebenchai.com
nodebench.workspace
nodebench-mcp
nodebench-mcp-power
nodebench-mcp-admin

Event Intelligence Serving Model

Event serving extends the same budgeted search route used by the main app and MCP runtime. NodeBench treats search as a memory-building operation with a budget. For events, NodeBench checks the event corpus and workspace memory before live search, then persists useful results as entities, claims, sources, and workspace context.

ScratchNode is the lightweight live-event sidecar for this model: a disposable room that turns public chat and sourced /ask answers into a public wiki while keeping attendee notes private. It complements Luma, Slack, Eventbrite, and other event surfaces instead of replacing them; the explicit handoff to NodeBench opens https://nodebenchai.com/events/:eventSlug/private with private-note continuation context and no URL-borne ownerKey.

The event flow is:

Before event
  -> build event corpus

During event
  -> capture messy notes instantly

After event
  -> turn captures into report, cards, follow-ups, and reusable memory

The product model is:

ScratchNode sidecar room + public event corpus + private NodeBench continuation

Event corpus and capture data stay separated:

• Shared event corpus = public event info, speakers, sponsors, company pages,

sessions, and public source cache.

• Private captures = what a user personally heard, wrote, recorded, or

photographed; ScratchNode private notes never enter the public feed, public wiki, or public /ask cache.

• Team/org memory = shared only inside the fund, company, or workspace.
• Event aggregate insights = opt-in or anonymized only.

During the event, most captures should hit the event corpus first and avoid paid search:

voice memo / text / screenshot
  -> captureRouter
  -> active event corpus
  -> entity and claim extraction
  -> active event session attachment
  -> budget policy
  -> ack + next action

Example mobile ack:

Saved to Ship Demo Day session
Detected 1 person | 1 company | 2 claims | 1 follow-up
Using event corpus | 0 paid calls

After the event, the report opens in nodebench.workspace:

Brief      -> post-event memo
Cards      -> people, companies, products, themes
Notebook   -> raw notes, transcripts, screenshot OCR, cleaned notes
Sources    -> field notes, public evidence, verification status
Chat       -> follow-up questions and deeper refreshes
Map        -> graph view later

The canonical spec lives in EVENT_INTELLIGENCE_SERVING_MODEL.md. The ScratchNode/NodeBench privacy boundary lives in SCRATCHNODE_NODEBENCH_BOUNDARY.md.

Why This Design

NodeBench is designed around a few product realities:

1. A useful answer should not disappear after one chat turn.
2. Saved work should become reusable memory, not a dead archive row.
3. The product should bring the user back only when something meaningful

changes.

1. The system should gradually learn how the user works without forcing a heavy

onboarding flow.

1. Operator context should improve future runs without turning the system into

corporate-speak or fake-agreeable sludge.

That drives the current design:

• answer-first execution
• advisor mode by design via dynamic routing:
• fast executive lane for routine work
• deeper advisor lane for ambiguity, planning, and harder reasoning
• similar in spirit to Claude Code's official opusplan split:

stronger planning lane, cheaper execution lane

• saved artifacts as first-class objects
• visible sources and traceability
• a five-page loop instead of five unrelated tabs
• future Harness v2 work focused on specification, operator context, and

compounding behavior

Plain English:

NodeBench should not spend the most expensive reasoning path on every request.
It should move fast by default, then go deeper when the task, evidence, or user
request justifies it.

The detailed implementation, verification, and evaluation plan for this mode lives in:

• HARNESS_V2_PROPOSAL.md (archived)
• HARNESS_V2_BUILD_PLAN.md (archived)

How The Five Pages Compound

NodeBench should not feel like five separate destinations.

The intended product behavior is:

Home
  -> start quickly

Reports
  -> turn that artifact into reusable memory

Chat
  -> do the work
  -> create the first useful artifact

Inbox
  -> triage captures, nudges, automations, alerts, and unassigned items

Me
  -> improve how the next run is handled

Workspace
  -> open deep Brief / Cards / Notebook / Sources / Chat / Map work
  -> lives at nodebench.workspace, not in the operating tab bar

Next Home or Chat run
  -> starts with more context than before

The shortest version of the compounding loop is:

question
  -> answer
  -> saved report
  -> watch item
  -> useful nudge
  -> better next run

Plain-English artifact flow:

input
  -> answer packet
  -> saved report
  -> tracked entity / tracked theme / follow-up task
  -> nudge or prep brief
  -> resumed report or resumed chat
  -> user correction or confirmation
  -> updated operator context
  -> better next run

What each page contributes:

• Home starts the run with the least friction possible
• Reports turns those into a durable report the user can reopen, refresh, and

reuse

• Chat creates the answer, sources, trace, entities, and next actions
• Inbox collects nudges, captures, automations, alerts, and unassigned items
• Me stores the operator context that improves the next answer
• Workspace owns recursive cards, notebook editing, source verification, and

long-lived intelligence memory

Current Legacy Infrastructure

NodeBench is not starting from zero. The repo already contains a substantial legacy stack that works today.

Current legacy foundation:

• five-surface web product
• Convex-backed canonical data layer
• local and deployed server runtime
• harness v1 planning and execution path
• shared-context handoff and delegation support
• MCP distribution lanes
• builder-facing evaluation and control-plane systems

What that means:

• the problem is not missing architecture
• the problem is product behavior, workflow compression, and clearer

cross-surface compounding

Roadmap

The near-term goal is:

keep the working legacy foundation
remove accidental complexity
add specification-aware operator context
ship one clear compounding workflow

Main tasks still to finish:

• [ ] make Home -> Reports -> Chat -> Inbox -> Me behave like one continuous

workflow instead of five adjacent surfaces

• [ ] turn harness v1 into the clearer v2 shape described in

HARNESS_V2_PROPOSAL

• [ ] ship Layer 0 operator context so the system can learn useful workflow

patterns without forcing a heavy onboarding flow

• [ ] support permissioned transcript ingestion from NodeBench chats first, then

optional external logs such as Claude Code JSONL transcripts for nodebench-mcp

• [ ] add style-drift guardrails so the system learns judgment and workflow

without overfitting to corporate voice, filler, or sycophancy

• [ ] add anticipatory prep behavior so the system can prepare the user before

important interactions, not only answer after the fact

• [ ] make saved reports behave like reusable memory, not storage
• [ ] keep Nudges as an Inbox section with at least one working daily trigger
• [ ] make Me clearly improve future runs by exposing what context is being

used and why

• [ ] finish the nodebench-mcp v3 cut-and-split plan so default runtime,

power runtime, and admin runtime are clearly separated

• [ ] instrument real latency, real cost, real artifact completion, and real

reuse across both web and MCP flows

• [ ] keep README, runtime behavior, and exposed tool counts in sync so the

public story matches the actual system

• [ ] keep dogfood, eval, and builder-control infrastructure as internal

leverage instead of letting it leak into the main user-facing product

Quick Start

Web app

Open nodebenchai.com and start in Home.

MCP

# Claude Code
claude mcp add nodebench -- npx -y nodebench-mcp

# Claude Code power lane
claude mcp add nodebench-power -- npx -y nodebench-mcp-power

# Claude Code admin lane
claude mcp add nodebench-admin -- npx -y nodebench-mcp-admin

# Cursor
npx nodebench-mcp --preset cursor

# Generic MCP client
npx nodebench-mcp

Local development

git clone https://github.com/HomenShum/nodebench-ai.git
cd nodebench-ai
npm install
cp .env.example .env.local

# Frontend + Convex + voice server
npm run dev

# Production build
npm run build

Architecture

nodebenchai.com (React + Vite + Tailwind)
    |
Convex Cloud (sessions, reports, entities, nudges, files, product state)
    |
server runtime + search pipeline + SSE
    |
answer packet
    |
saved report
    |
tracked entities / watch conditions / nudges
    |
future runs with better operator context

Student Learning Lessons

The notebook and diligence stack in this repo are a good example of a common product engineering tradeoff:

• the best user experience is one notebook that feels continuous
• the safest current runtime is still layered and block-addressable underneath

For NodeBench, that means:

• founder is a trait and diligence block, not a permanent sixth tab
• diligence should use one generic pipeline, not many narrow *Identify.ts

features

• the runtime should stay `scratchpad-first -> structuring pass ->

deterministic merge`

• user-owned prose should feel local-first and calm while typing
• live agent output should arrive as overlays or decorations first, not as

direct document mutations

• accepted agent output should become frozen, user-owned notebook content
• provenance should stay available, but secondary to the reading and writing

flow

Why the notebook does not use one giant live editor model yet:

• collaboration is more reliable when the system can address bounded sections
• provenance, evidence, and contribution logs need stable attachment points
• background agent updates should not compete with user keystrokes
• deterministic section-level merge is easier to reason about than whole-page

mutation churn

The practical rule in this repo is:

UX should feel monolithic.
Runtime should stay layered.
Typing should be local-first.
Agent output should be overlay-first.
Accepted output should become owned prose.

Current notebook refactor lessons:

• hide the block machinery from the reading path
• keep chrome quiet and move metadata to hover or focus
• isolate the notebook surface from page-level re-render churn
• favor one memoized notebook boundary over many inline object props
• treat live diligence as read-only reference overlay until the user accepts it
• when accepted, materialize a frozen notebook snapshot with explicit provenance
• anchor live overlays at the notebook surface, not inside the first editable row
• let Convex projection rows carry real source metadata so the UI is not forced to reconstruct trust state from prose alone
• use one generic projection producer for overlays: report save writes the same structured rows that page-load backfill and manual refresh re-run
• when moving beyond report-backed overlays, stream raw scratchpad only in a secondary rail and emit structured projection rows on checkpoint rather than dumping scratchpad prose into the notebook body
• if checkpoint structure comes from an LLM, keep it block-scoped and schema-bound: scratchpad checkpoint -> JSON -> validation/repair -> deterministic fallback -> projection row
• let the model structure intermediate JSON, but keep merge, persistence, and notebook ownership deterministic
• ship generic diligence primitives first, then block-specific renderers

For students reading the code, the most relevant docs are:

• AGENT_PIPELINE
• DILIGENCE_BLOCKS
• SCRATCHPAD_PATTERN
• PROSEMIRROR_DECORATIONS
• SESSION_ARTIFACTS

The live notebook refactor is deliberately incremental:

• current shipped slices make the notebook feel more continuous and reduce

per-keystroke render churn

• current shipped slices also move live diligence into notebook-surface

overlays instead of seeded block-like records and freeze accepted snapshots

• the end state is one notebook experience with layered internals, not a raw

block UI and not a brittle giant document runtime

Key tech

• Frontend: React, Vite, TypeScript, Tailwind CSS
• Backend: Convex
• Search: Linkup + Gemini extraction + grounding pipeline
• MCP server: Node.js + TypeScript
• Realtime runtime: SSE + Convex-backed persistence

API Keys

Set these in .env.local for local work or in Convex / Vercel for deployed environments.

| Key | Required | Purpose | | --- | --- | --- | | GEMINI_API_KEY | Yes | classification, extraction, synthesis | | LINKUP_API_KEY | Recommended | web search and sourced answers | | VITE_CONVEX_URL | Yes | Convex deployment URL |

Codebase map

Top-3 levels, annotated. See docs/ARCHITECTURE.md for the pipeline diagram and docs/architecture/README.md for the 13 canonical architecture docs.

nodebench-ai/
├── README.md                   ← you are here
├── CONTRIBUTING.md             ← contribution bar
├── CLAUDE.md                   ← Claude Code conventions for this repo
├── AGENTS.md                   ← agent methodology + eval bench
├── LICENSE                     ← MIT
│
├── apps/
│   └── web/src/                ← React frontend (Vite)
│       ├── features/           ← feature-first (Home · Reports · Chat · Inbox · Me · Workspace · entities · agents · …)
│       │   └── <feature>/      ← views · components · hooks · lib · __tests__ (colocated)
│       ├── shared/             ← shared UI primitives, hooks, utils
│       ├── lib/                ← registry, analytics, error reporting
│       └── layouts/            ← shell + cockpit + public
│
├── workers/node/               ← Node runtime (Express + MCP gateway)
│   ├── pipeline/               ← agent harness runtime + diligence blocks
│   ├── routes/                 ← HTTP routes (search, harness, founder episodes)
│   ├── mcpGateway.ts           ← WebSocket MCP gateway
│   └── services/               ← shared services
│
├── backend/convex/             ← Convex backend
│   ├── domains/                ← domain folders (agents · product · research · founder · search · …)
│   ├── schema.ts               ← database schema (includes agentScratchpads)
│   └── crons/                  ← scheduled jobs
│
├── packages/
│   ├── mcp-local/              ← the published nodebench-mcp npm package (MIT)
│   ├── mcp-client/             ← typed client SDK
│   └── convex-mcp-nodebench/   ← Convex-side MCP auditor
│
├── .claude/
│   ├── README.md               ← map of the .claude/ layout
│   ├── rules/                  ← 31 modular rules with related_ cross-refs
│   ├── skills/                 ← reusable how-to procedures
│   ├── agents/                 ← subagent configs
│   └── commands/               ← custom slash commands
│
├── docs/
│   ├── README.md               ← docs tree map
│   ├── ARCHITECTURE.md         ← top-level pipeline diagram
│   ├── ONBOARDING.md           ← 30-minute new-contributor path
│   ├── architecture/           ← 13 canonical specs + plans/ + README index
│   ├── agents/                 ← agent docs + bootstrap configs
│   ├── guides/                 ← how-to for builders
│   ├── decisions/              ← ADRs
│   ├── changelog/              ← release notes
│   ├── product/                ← product decisions
│   ├── qa/                     ← QA protocols
│   └── archive/                ← superseded content, provenance-only
│
├── evals/
│   └── e2e/                    ← end-to-end eval suites
│
├── scripts/                    ← dogfood, eval harness, one-offs
└── public/                     ← static assets served by Vite + Vercel

Related Docs

Start here: docs/ONBOARDING.md · docs/ARCHITECTURE.md · docs/architecture/README.md

The 13 canonical architecture docs are organized in 4 tiers. See docs/architecture/README.md for the indexed map:

• Tier 1 (core pipeline): AGENT_PIPELINE · DILIGENCE_BLOCKS · USER_FEEDBACK_SECURITY
• Tier 2 (sub-patterns): SCRATCHPAD_PATTERN · PROSEMIRROR_DECORATIONS · AGENT_OBSERVABILITY · SESSION_ARTIFACTS
• Tier 3 (features): FOUNDER_FEATURE · REPORTS_AND_ENTITIES · AUTH_AND_SHARING
• Tier 4 (cross-cutting): MCP_INTEGRATION · EVAL_AND_FLYWHEEL · DESIGN_SYSTEM

Active architecture addenda: GRAPH_SEARCH_AGENT_CONTEXT captures the graph/search/agent-context strategy, exact product questions, scale projection, node attention model, and human-vs-agent retrieval split.

Historical specs are preserved in docs/archive/2026-q1/.

Production Readiness & Evaluation

NodeBench ships with a comprehensive evaluation harness that proves correctness across 32+ scenarios, 9 user personas, and 9 feature categories. This is not hand-wavy "it works" — it is measured, versioned, and reproducible.

Latest Published Run Results

Pi-AI pipeline cascade: merged to main on 2026-04-30 at 2a541037874c0f8c675ab393d5c08f50123cf6d2.

| Lane | Result | |------|--------| | PR chain | #211 -> #212 -> #213 -> #214 -> #215 -> #216 all merged | | Production surface | https://www.nodebenchai.com/?surface=packets | | MCP bridge | https://agile-caribou-964.convex.site/mcp/pipeline/* behind MCP_SECRET | | Code-gen run | pipeline_mokobe4y_6n23be succeeded, verified, 6 files, 32.9s, about $0.001 | | Research streaming | pipeline_mokpvi1b_yoj8ot completed with 4,317 streamed characters | | Linkup research | pipeline_mol2wj2j_2lgx2u succeeded with 18 snippets across 5 sub-questions | | Composed pipeline | research_then_code completed stage 1 research and stage 2 code-gen | | Schedule workflow | once schedule swept by cron/manual sweep and auto-disabled after run | | Design output | design-gen produced a PNG stored in Convex storage | | UI launcher | DOM-submitted composed run updated the reactive run list | | Pipeline scorecard | 41.7% verified, Brier 0.135 across 12 runs |

The implementation is mounted on the Reports surface:

• PipelineLauncher
• PipelineSchedulesPanel
• PipelineEvalScorecard
• PipelineRunsPanel
• EntityFindingsPanel

The detailed handoff is in docs/handoff/PI_AI_PIPELINES_HANDOFF.md.

Workflow-loop eval bank: added on 2026-04-30 to test the full product loop, not just answer text.

query / capture
  -> memory search
  -> entity resolution
  -> report update
  -> notebook update
  -> graph edges
  -> sources / claims
  -> follow-up / export

| Eval bank | Result | |-----------|--------| | Total workflow cases | 124 | | Minimum P0 suite | 30 cases | | Coverage categories | 11 | | Score dimensions | 12 | | Validator | apps/web/src/features/evaluation/data/nodebenchWorkflowEvalBank.test.ts | | Latest local check | npx vitest run apps/web/src/features/evaluation/data/nodebenchWorkflowEvalBank.test.ts -> 4/4 passed |

The eval bank lives in apps/web/src/features/evaluation/data/nodebenchWorkflowEvalBank.ts.

Two-Layer Judge Architecture

Every production run is evaluated by two independent systems:

Layer 1: Deterministic Boolean Gates (server/pipeline/diligenceJudge.ts)

• 10 strict pass/fail checks: tier validity, latency budget, token tracking,

source capture, terminal status

• Verdicts: verified | provisionally_verified | needs_review | failed
• Zero LLM involvement — pure deterministic validation

Layer 2: LLM Semantic Scoring (server/pipeline/diligenceLlmJudge.ts)

• 5 dimensions scored [0,1]: prose quality, citation coherence, source

credibility, tier appropriateness, overall semantic fit

• Prompt version tracking (llmjudge-v1) for cohort separation
• Bounded: 30s timeout, 512KB response cap, honest error reporting

This dual-layer approach means hallucinations and quality regressions are caught by two independent systems before they reach users.

Current Production Status

Latest Full-Stack Eval: 2026-04-23T06:46:53Z

Overall Pass Rate:     100% ✅
LLM Judge Average:     9.6/10 (target: ≥7 for production)
Dogfood Score:         100/100 (0 real issues)
Entity Resolution:     100% ✅
Factual Accuracy:      90.6% ✅
No Hallucinations:     90.6% ✅
Actionable Output:     100% ✅
Answer Control:        100% ✅ (all 8 dimensions)
Feature Breadth:       100% ✅ (31 scenarios)
Retention/Continuity:  4/4 passed ✅

All production gates passing:

• ✅ Expanded Feature Coverage Production Gate
• ✅ Answer Control Production Gate
• ✅ Dogfood Production Gate
• ✅ Notebook Capacity Production Gate
• ✅ History Soak Production Gate

Note: The only outstanding item is p95 latency optimization (174s vs 90s target) — a performance enhancement, not a correctness blocker. The system is production-ready for all quality scenarios.

Evaluation Coverage

Capability Eval — 32 Persona Scenarios

| Persona | Example Query | Status | |---------|---------------|--------| | JPM Startup Banker | "DISCO — worth reaching out? Fastest debrief" | ✅ 100% | | Early Stage VC | "OpenAutoGLM — what's the wedge?" | ✅ 100% | | CTO Tech Lead | "QuickJS — do I have exposure?" | ✅ 100% | | Enterprise Exec | "Gemini 3 — procurement next step?" | ✅ 100% | | Ecosystem Partner | "SoundCloud VPN — who benefits?" | ✅ 100% | | Founder Strategy | "Salesforce Agentforce — counter-positioning?" | ✅ 100% | | Academic R&D | "RyR2/Alzheimer's — literature anchor?" | ✅ 100% | | Quant Analyst | "DISCO — extract funding signal" | ✅ 100% | | Product Designer | "DISCO — schema-dense UI card JSON" | ✅ 100% | | Sales Engineer | "DISCO — share-ready outbound summary" | ✅ 100% |

Expanded Feature Breadth — 31 Scenarios

| Category | Count | Pass Rate | |----------|-------|-----------| | Calendar | 3 | 100% ✅ | | Disclosure | 4 | 100% ✅ | | Document | 3 | 100% ✅ | | Hybrid | 4 | 100% ✅ | | Media | 3 | 100% ✅ | | Skills | 4 | 100% ✅ | | Spreadsheet | 3 | 100% ✅ | | Tools | 4 | 100% ✅ | | Web | 3 | 100% ✅ |

Answer Control — 8 Dimensions

• Entity resolution: 100% ✅
• Retrieval relevance: 100% ✅
• Claim support: 100% ✅
• Final response quality: 100% ✅
• Trajectory quality: 100% ✅
• Actionability: 100% ✅
• Artifact decision quality: 100% ✅
• Ambiguity recovery: 100% ✅

How to Verify

Run the full production evaluation suite:

# Full 8-phase evaluation (typecheck → build → capability → expanded →
# answer-control → dogfood → notebook → history)
npm run eval

# Quick verification (3 scenarios)
npm run eval:quick-slice

# Individual lanes
npm run eval:capability      # 32 persona scenarios
npm run eval:feature-breadth # 31 feature scenarios  
npm run eval:retention       # Wiki continuity suite

All artifacts are versioned in docs/architecture/benchmarks/:

• full-stack-eval-latest.md — aggregate summary
• comprehensive-eval-*.md — capability results
• expanded-eval-*.md — feature breadth results
• product-answer-control-eval-*.md — answer control results

What "Production Ready" Means Here

1. Deterministic gates pass — no regressions in core correctness
2. LLM judge scores ≥7 — semantic quality validated by independent LLM
3. Dogfood score ≥85 — internal usage shows no real issues
4. All 32 persona scenarios pass — diverse user types handled correctly
5. All 31 feature scenarios pass — broad surface area covered
6. Retention/continuity passes — long-term memory works
7. Answer control 100% — artifact decisions, ambiguity recovery solid

The system meets all of these. The only remaining work is latency optimization — making fast answers even faster, not making broken answers work.

Model Strategy

• Primary: moonshotai/kimi-k2.6 (OpenRouter) — 100% capability pass
• Fallback: gpt-5.4 — automatic retry on empty/missing debrief
• Judge: kimi-k2.6 — 9.6/10 average across all scenarios

Kimi is the primary lane. GPT-5.4 remains the safety fallback until Kimi's first-attempt stability improves, but both paths are production-tested.

Product Suite

NodeBench AI   = flagship user surface
nodebench-mcp  = workflow lane
Attrition.sh   = measured replay + optimization lane

Attrition is not a third flagship. It is the measurable optimization lane for the same NodeBench workflow.

License

MIT