๐Ÿฆž Clow Studio โ€” a Desired Solutions venture

Harness Scorecard

The 9-point agent-control-loop discipline, mapped to code, as of v2.17.3
"Claude Code is just a harness; the LLM is the horse." โ€” Agentic CodingCon 2026, Berlin
๐Ÿ“Š ARM & Audit Trail ๐Ÿš€ Solopreneur-in-a-Box ๐Ÿ“– HARNESS.md (long form) ๐Ÿงฌ CLAW.md (substrate) ๐Ÿ›ก๏ธ EU AI Act mapping โ‡ฃ agent-bom.json

The model is a commodity. The harness around it โ€” input control, cost tracking, observability, prompt discipline, tool permissions, execution boundaries, output validation, offline evaluation, and governance report โ€” is the moat. This page is the Clow Studio self-audit against those nine dimensions. Each entry names the file, endpoint, or doctrine that satisfies it today โ€” and what's partial or deferred is named honestly, because overclaiming is the exact failure the discipline exists to prevent.

Full coverage
40
implemented & measured
Partial
2
covered out-of-runtime
Deferred
2
next cycle ยท named
Dimensions covered
9 / 9
all have โ‰ฅ 1 โœ…
Overall score
91%
40 / (40+2+2)
Run live
/harness
on @deeptechx ยท @productization ยท @wmagentic

The nine dimensions

1. Input control 3 โœ… ยท 1 โš ๏ธ

Never let anything you didn't shape into the model.
  • โœ…
    Command allowlist on the switch
    internal/bot/handler.go โ€” 25+ /commands
  • โœ…
    Per-user rate limit (burst window)
    5/30s free ยท 50 pro ยท 500 whale
  • โš ๏ธ
    Message-length cap on inbound
    defers to Telegram's 4096-char limit
  • โœ…
    Indirect-injection probe in ingested docs
    audit/evoskillopt_qgate_redteam.yaml

2. Token / cost tracking 4 โœ… ยท 1 โŒ

Know what each call costs, per-user, per-skill.
  • โœ…
    Daily LLM-call cap per tier
    5 free ยท 500 pro ยท 10k whale
  • โœ…
    Budget circuit breaker
    one of the 5 runtime breakers
  • โœ…
    Per-skill cost emission
    evokaf_telemetry_stream.token_cost_usd
  • โœ…
    Daily cost roll-up
    v_client_audit_daily.total_cost_usd
  • โŒ
    Monthly-cost projection
    next cycle โ€” SQL rollup only

3. Observability / tracing 4 โœ… ยท 1 โš ๏ธ

Every call has a parent and a span.
  • โœ…
    Trace span per command
    /trace โ€” per-message span tree
  • โœ…
    SHA-256 input hash (zero PII)
    v_client_execution_audit.input_sha256
  • โœ…
    Tools invoked per trace
    v_client_execution_audit.tools_invoked
  • โš ๏ธ
    OTel span export to SIEM
    Node proxy only โ€” Go-side JSONL next
  • โœ…
    Latency distribution per skill
    ResponseTimer EMA โ†’ /dashboard/reactivity

4. Prompt optimization 5 โœ…

Fewer tokens, sharper intent.
  • โœ…
    Doctrine injected once, not per-message
    config.BuildPrompt โ€” assembled once
  • โœ…
    Skills listed by name only
    config.SkillsShort()
  • โœ…
    Tiered prompts per evoMode
    evolutionary / auto / skills_only
  • โœ…
    Karpathy terse discipline
    KarpathyBehavior โ€” 10 rules
  • โœ…
    Content-routing dedup
    CONTENT_ROUTING_v2.17.md โ€” one bot one domain

5. Tool control 4 โœ…

The agent can only call what you listed.
  • โœ…
    Deterministic first, LLM fallback
    25+ handlers: /fuse, /paper, /launch, /rea, /geoai, /harness
  • โœ…
    MCP tool allowlist per genome
    evoskillopt_genomes.allowed_mcp_tools
  • โœ…
    Tool-privilege-escalation probe
    evoskillopt_qgate_redteam.yaml
  • โœ…
    Out-of-runtime kill switch
    clow-audit-proxy.mjs โ€” 403 on is_frozen=1

6. Execution control 5 โœ…

You can stop, pause, roll back.
  • โœ…
    5 runtime circuit breakers
    rate ยท budget ยท evolution ยท audit ยท backup
  • โœ…
    Per-skill freeze (single-click)
    evoskillopt_genomes.is_frozen
  • โœ…
    Max_depth=2 recursion ceiling
    CIRCUIT_BREAKER_CHECK + proxy enforces X-Clow-Recursion-Depth
  • โœ…
    3ร— FAIL_SECURITY โ†’ evolution trip
    evoskillopt_mutation_log + Q-Gate
  • โœ…
    Agents draft, humans sign
    execution_mode = DRAFT_ONLY default

7. Validation โ€” the load-bearing one 7 โœ…

Every claim traces to a named input. The four surfaces of one rule โ€” the specific thing a weekend-built competitor cannot copy.
  • โœ…
    Source Discipline doctrine
    ClaimProvenance โ€” injected in every bot's prompt
  • โœ…
    Manuscript Discipline (5 rules)
    ManuscriptDiscipline: knowledge isolation, citation grounding, no overclaim, data-integrity refusal, structural rigour
  • โœ…
    /paper audit for any draft
    9 subcommands: pipeline, rules, brief, cite, audit, claims, overclaim, structure, explain
  • โœ…
    REA OBSERVED / INFERRED / UNKNOWN
    /rea + /reaclaims
  • โœ…
    Positive Intelligence [E] / [C] / [A] / [K]
    /positive
  • โœ…
    G18 Decomposition & Fusion โ€” min not mean
    /fuse โ€” fused score, weakest-pillar named, UNKNOWN protocol
  • โœ…
    Four surfaces, one posture
    ClaimProvenance names REA + E-C-A-K + /fuse + /paper as four surfaces of ONE rule

8. Evaluation 3 โœ… ยท 1 โŒ

Adversarial probes before promotion.
  • โœ…
    Q-Gate adversarial suite in mutation loop
    evoskillopt_qgate_redteam.yaml โ€” 8 probes ยท one FAIL = no promotion
  • โœ…
    Multi-agent CI red-team
    clow_agents_redteam.yaml โ€” 10 probes across 5 bots
  • โŒ
    Static OWASP-LLM linter on SKILL.md
    hooks in Q-Gate; rules deferred (qdefence lint)
  • โœ…
    Shadow-mode measurement before A/B
    Version โ†’ Measure โ†’ Shadow โ†’ A/B โ†’ Promote โ†’ Rollback โ†’ Retrospect

9. Reporting 5 โœ…

The governance pack the client reads on their own.
  • โœ…
    7 v_client_* SQLite views (zero PII)
    audit/audit_client_views.sql
  • โœ…
    Datasette sidecar on :8001
    audit/docker-compose.audit.yml
  • โœ…
    Signed agent-bom.json
    internal/governance/agentbom.go โ€” HMAC-SHA256, canonical JSON
  • โœ…
    EU AI Act Art. 9/12/13/14/15 mapping
    GOVERNANCE_EU_AI_ACT.md
  • โœ…
    Audit Trail tab on ARM dashboard
    web/analytics.html โ€” 4-pillar coverage KPIs

The other moat โ€” substrate

The Agentic CodingCon takeaway is that the moat has moved. There are two places it can live. This page is the harness side; the substrate side (proprietary data, domain context above the training cutoff, institutional trust, specialized distribution) lives in CLAW.md. Highlights per bot:

@deeptechx_bot

deep tech / space / CAS

ESA / Galileo / GOVSATCOM specs, FIRMS fire feeds, Sentinel/MODIS bands, Servigne / Haziza / Descollonges / Arendt framework citations. A cold model hallucinates every one.

@legalclow_bot

EU regulatory

EU AI Act (OJ L 2024/1689) article-level, NIS2 Art. 23, GDPR Art. 28 DPA templates, country-specific incorporation playbooks DE/AT/CH/US/UK/FR.

@wmagentic_bot

intelligence briefs

800+ DOAC episodes with Grokipedia [E/C/A] tiering, sanctions registers, Copernicus near-real-time. Zero-tolerance citation rule.

Every bot's system prompt carries ClaimProvenance + ManuscriptDiscipline โ€” two load-bearing doctrines that treat any claim not traceable to a named input as a liability. The substrate above is what makes "a named input" available, so the rules don't become a straitjacket forcing every response to be "I don't know."