CallScore — Agentic Crypto Creator Intelligence Platform

CallScore ranks crypto creators by evidence-backed prediction quality. It ingests creator/video/transcript data, extracts claims and market signals, matches them to outcomes, scores creator credibility/performance, and exposes a public product through call-score.com. The platform is commercially gated through Whop and operationally governed by Hermes/Workplane receipts, handovers, default-open owned public GTM policy, and fail-closed restricted-action gates.

• Live product: https://call-score.com
• Canonical deploy target: Netlify
• Canonical app repo: /opt/crypto-tuber-ranked on branch master
• Current readiness: CONTROLLED_FULL
• Latest canonical gate model: docs/ops/2026-06-14-full-system-live-canary-gate-decisions.md
• Canonical session startup: docs/ops/callscore-canonical-session-startup.md — standard docs/skills for future Hermes sessions, including /callscore-standard, headroom, and /agentmemory.

Executive summary

CallScore is an agentic crypto creator intelligence platform:

• ranks crypto creators by historical call performance and evidence quality;
• collects creator/video/transcript data through a bounded, cookie-local laptop transcript lane;
• extracts claims/signals with rule paths and local Ollama Gemma/Qwen shadow canaries;
• matches calls to symbols/outcomes and scores public-safe creator/call rows;
• serves a public website and API through Netlify + HH Read API;
• supports monetization and entitlements through Whop / Whop Auto;
• uses Hermes/Workplane for governed autonomous operations, receipts, and handovers;
• uses Art of War and Composio for default-open owned organic GTM, with restricted actions still fail-closed.

Current production status

Canonical architecture diagram

Source file: docs/architecture/callscore-agentic-platform.mmd

flowchart TD
  Visitor[Public visitor / buyer] --> Netlify[Netlify CallScore app\ncall-score.com]
  Netlify --> HHRead[HH Read API]
  HHRead --> PG[(Local HH PostgreSQL)]

  YT[YouTube / creator source data] --> Laptop[Omar laptop\nTailscale + cookies + yt-dlp]
  Laptop --> Ingest[Approved transcript ingest]
  Ingest --> PG
  YT -. diagnostic only .-> HHFallback[HH yt-dlp / ASR / vpn-ytdlp]
  HHFallback -. future canary .-> Ingest

  PG --> Extract[Extraction]
  Extract --> Ollama[Local Ollama\nGemma / Qwen shadow + diff]
  Ollama --> Extract
  Extract --> Match[Matching]
  Match --> Score[Scoring / freshness / audit]
  Score --> PublicRows[Public-safe ranked rows]
  PublicRows --> HHRead

  Hermes[Hermes] --> Workplane[Workplane\nCONTROLLED_FULL]
  Workplane --> Receipts[Workflow receipts]
  Workplane --> Handover[Docs / handovers / memory]
  Workplane -. monitors .-> PG
  Workplane -. monitors .-> Netlify

  WhopAuto[Whop Auto] --> Whop[Whop checkout / entitlement]
  Whop --> Netlify

  Art[Art of War] --> Packet[Private approval packet]
  Packet --> Owned[Owned-channel canary\napproval required]

  Composio[Composio MCP] --> Attio[Attio]
  Composio --> Gmail[Gmail]
  Composio --> X[Twitter/X]
  Composio --> PostHog[PostHog]
  Composio --> HF[Hugging Face]
  Composio --> LinkedIn[LinkedIn]
  Composio --> Discord[Discord]

  Receipts --> Monitored[Monitored gates\nbacklog / cooldown / manual-review deltas]
  Receipts --> FailClosed[Fail-closed gates\nspend / public sends / Whop mutation / destructive actions / secrets]

For the full subgraph version, open docs/architecture/callscore-agentic-platform.mmd. Narrative companion: docs/architecture/callscore-agentic-platform.md.

Repository and service map

Canonical environment

Canonical local runtime env source:

/opt/crypto-tuber-ranked/.env.hermes

Rules:

• .env.hermes is local-only, chmod 600, and gitignored.
• Do not print values, tokens, cookies, DB URLs, auth headers, private keys, or credential-bearing remotes.
• Hermes, Workplane, CallScore scripts, Whop Auto checks, Art of War workflows, Composio tooling, and owned public GTM tooling should source this file first.
• Compatibility env paths should be symlinks/loaders to this canonical file, not competing secret stores.
• Redacted manifest: docs/ops/callscore-canonical-env-manifest.md.
• Validator: node --import tsx scripts/validate-hermes-env.ts.

Runtime architecture

Public product

The public site runs on Netlify at call-score.com. Vercel is stale/defunct and must not be used as production. Public routes and Next.js API routes read from the canonical HH Read API / local HH PostgreSQL source. The public target-price monetization leak is fixed; public rows expose only public-safe fields.

Canonical data source

The production source of truth is local HH PostgreSQL plus the HH Read API. Neon is stale/defunct and must not be used as production. Broad writes, destructive SQL, broad backfills, and recomputes require bounded repo-approved jobs and receipts.

Transcript collection

Canonical transcript acquisition is Omar laptop over Tailscale using residential browser cookies and laptop-side yt-dlp. Cookies stay laptop-local; HH receives transcript JSON/results only through approved ingest paths. Provider HTTP 429 is a cooldown signal: stop, record a receipt, and resume only after cooldown with bounded batches.

HH direct yt-dlp, ASR, and VPN yt-dlp lanes are diagnostic/future fallback only unless future canaries promote them.

Intelligence pipeline

The pipeline flows through extraction, matching, scoring, freshness, and audit checks. Local Ollama Gemma/Qwen runs artifact-only shadow extraction and diff workflows. Bounded write-canaries are allowed only through explicit repo gates/receipts; broad promotion remains fail-closed.

Hermes / Workplane control plane

Hermes and Workplane provide governed automation:

• npm run workplane / npm run workplane:status for current machine-readable readiness;
• job registry and safe next action model;
• workflow receipts in .tmp/workflow-receipts/;
• handovers under docs/handovers/;
• durable memory through agentmemory / callscore-memory;
• gate decisions under docs/ops/.

Commercial / entitlement plane

Whop Auto and Whop provide checkout, Pro/Alpha entitlement checks, webhook handling, and revenue readiness. The zero-dollar/token-discount Pro renewal proof is valid checkout/payment authorization evidence. Future provider/customer/payment/pricing mutations remain fail-closed unless manifest + diff + rollback + approval receipt + local auth + explicit safe classification all pass.

Growth / marketing plane

Art of War lives under /srv/agents/repos/Claude_Code_Automations. It supports campaign generation, persona scoring, owned-channel public canaries, post-execution receipts, monitoring, and restricted-lane approval packets. Owned CallScore organic posts on owned/managed channels are allowed by default when zero-cost, safe, non-financial, non-secret, and within messaging policy. Email/DM/outreach, non-owned posting, paid actions, provider/customer/payment changes, and restricted claims remain fail-closed without approval receipt.

Connected apps

Composio is the MCP hub for connected app inventory and possible future governed actions:

• Attio CRM;
• Gmail/email;
• Twitter/X;
• PostHog;
• Hugging Face;
• LinkedIn;
• Discord.

Current read-only inventory found Attio, Gmail, Twitter/X, PostHog, Hugging Face, LinkedIn, and Discord active through Composio. Active connection does not override the GTM registry: sends, DMs, paid actions, CRM/analytics writes, provider mutations, and restricted claims remain gated.

Data flow

Creator/source discovery
→ laptop transcript collection over Tailscale
→ approved transcript ingest
→ extraction / local Gemma-Qwen shadow canaries
→ matching / symbol normalization
→ scoring / freshness / audit
→ HH Read API
→ Netlify public app
→ Whop entitlement / monetization
→ Art of War + Composio marketing/feedback loops
→ Workplane receipts, handovers, memory

Agentic control plane

Safety and gate model

Released / monitored gates

These are no longer hard blockers for current production readiness when evidence remains healthy:

• transcript audit backlog while bounded laptop cadence is proven;
• provider HTTP 429 cooldown when Workplane waits instead of retrying;
• Gemma manual-review deltas with zero missing/extra calls in the current diff;
• Art of War private approval packet readiness;
• Composio non-core connected-app gaps.

Fail-closed gates

These remain hard safety controls:

• paid spend, ads, enrichment, paid APIs, paid LLM calls;
• public posts, DMs, email sends, or outreach without approval receipt;
• Whop pricing/product/customer/payment mutations without all required gates;
• destructive SQL or destructive infrastructure;
• broad DB writes/backfills/recomputes without bounded receipt;
• credential rotation or secret exposure;
• cookie transfer to HH;
• hammer retry after provider HTTP 429.

Validation commands

Run from the canonical repo:

cd /opt/crypto-tuber-ranked
git diff --check
npm run typecheck
npm run lint
npm run build
npm run hygiene
npm run workplane:status || npm run workplane
npm run freshness:check
npm run audit:pipeline -- --summary --allow-partial-shadow
npm run verify:public
npm run verify:public -- --source live --base-url https://call-score.com
node --import tsx --test $(find tests -name '*.test.ts' | sort)

Operating runbooks

Check Workplane

cd /opt/crypto-tuber-ranked
npm run workplane:status || npm run workplane

Expected current state: status=OK, automation_readiness=CONTROLLED_FULL. If any core domain becomes BLOCKED, PARTIAL, or NOT_CONNECTED, inspect before claiming readiness.

Check public website

cd /opt/crypto-tuber-ranked
npm run verify:public -- --source live --base-url https://call-score.com

Expected: live health OK with HH Read API source and homepage/leaderboard checks passing.

Resume transcript cadence after cooldown

Use the canonical laptop lane, not HH-only fallback:

ssh -i ~/.ssh/bridge_hetzner_to_wsl -o IdentitiesOnly=yes omar@100.118.20.40 \
'hostname && /mnt/c/Windows/System32/WindowsPowerShell/v1.0/powershell.exe -NoProfile -ExecutionPolicy Bypass -File C:/Users/albak/run-transcript-collector-fixed-wslssh.ps1 -Limit 5 -Browser firefox -GapSeconds 60 -SinceDays 365 -HhHost hermes-agent-box -Write'

If HTTP 429 appears: stop, record cooldown receipt, do not retry hammer.

Review audit backlog

cd /opt/crypto-tuber-ranked
npm run audit:pipeline -- --summary --allow-partial-shadow

missing_transcripts_or_terminal_reasons is monitored backlog while the transcript mechanism remains healthy.

Review Gemma diff

cd /opt/crypto-tuber-ranked
ls -1t .tmp/shadow-extraction/*diff*.jsonl | head
npm run shadow:diff -- --help

Classify manual-review rows. Do not broad-promote. Production writes require receipt-gated canary paths.

Whop safe checks

cd /opt/crypto-tuber-ranked
node --import tsx --test tests/whop-certification-pack.test.ts tests/infrastructure-canonical.test.ts tests/whop*.test.ts

No live pricing/product/customer/payment mutation unless manifest + diff + rollback + approval receipt + local auth + explicit safe classification all pass.

Art of War dry-run / preflight

cd /srv/agents/repos/Claude_Code_Automations
python3 scripts/art_of_war.py campaign-loop --dry-run --campaign-id callscore-readiness-check --output /tmp/callscore-art-of-war-readiness-check.json

Owned CallScore organic public posting is default-open under READY_PUBLIC_OWNED after messaging-policy pass and must write a post-execution receipt. Spend, email/DM/outreach, non-owned posting, Whop/payment/customer/provider mutation, DB/deploy/infra mutation, and secrets remain fail-closed without approval receipt.

Composio inventory

codex mcp list
codex mcp get composio --json

Redact headers/tokens in any shared output. Use read-only inventory before any CRM/email/social/analytics action.

Update handover

Append operational changes to:

docs/handovers/2026-06-14-hermes-agent-callscore-activation.md

Include command evidence, receipts, changed files, what Hermes should read first, and what not to repeat.

Deployment model

• Netlify is canonical production hosting.
• Vercel is not production.
• Deploy only from a clean validated commit.
• Do not deploy for docs-only changes.
• After runtime/app deploys, verify live with:

npm run verify:public -- --source live --base-url https://call-score.com

Handover and memory

Read these before any new production activation run:

1. This README.
2. docs/architecture/callscore-agentic-platform.mmd.
3. docs/architecture/callscore-agentic-platform.md.
4. docs/ops/2026-06-14-full-system-live-canary-gate-decisions.md.
5. docs/handovers/2026-06-14-hermes-agent-callscore-activation.md.
6. agentmemory / callscore-memory.

Known non-core operational backlog

• Continue bounded transcript backlog clearing after cooldown.
• Optional: complete Hugging Face Composio auth if a Composio-backed HF lane becomes required.
• Next GTM action: run one owned X/Twitter public canary under READY_PUBLIC_OWNED, then write execution receipt and monitor read-only metrics.
• Optional: stale mirror/archive/secret quarantine cleanup with separate explicit approval.

Canonical GTM agent registry

GTM/channel ownership lives in docs/ops/callscore-gtm-agent-registry.json, with human summary at docs/ops/callscore-gtm-agent-registry.md and diagram at docs/architecture/callscore-gtm-agent-registry.mmd. Read it before any marketing, Whop, Composio, CRM, social, email, community, analytics, or provider action. It now marks owned organic X/LinkedIn/owned-community/SEO/Whop-copy lanes as READY_PUBLIC_OWNED; restricted lanes remain gated.