aegis-refine

name	aegis-refine
description	Operate Aegis Refine dataset jobs.
version	1.0.0
author	Jason Brashear
license	MIT
platforms	["macos","linux"]
prerequisites	{"commands":["curl","python"]}
metadata	{"hermes":{"tags":["aegis","datasets","stripe","nvidia","hermes-agent"],"related_skills":[]}}

Aegis Refine Operator

Use this skill when Hermes Agent is asked to operate an Aegis Refine customer job: quote review, route planning, spend governance, dataset refinement oversight, synthetic augmentation oversight, or signed-delivery verification.

Hermes Agent is the operator. Aegis-14B is the data-governance specialist, published as a LoRA fine-tune of NousResearch/Hermes-4-14B. nvidia/nemotron-3-ultra-550b-a55b is the primary operations brain for roadmap, routing, margin, and spend decisions. nvidia/nemotron-3-nano-30b-a3b is the documented latency/cost fallback for receipt generation. nvidia/nemotron-3.5-content-safety is the NVIDIA safety gate for PII / unsafe-content review when raw or summarized safety evidence is available. Stripe Checkout is the earn rail. Outbound spend is a Hermes-initiated Stripe Connect Transfer to the AINode compute vendor when configured; Aegis verifies the returned Stripe object before recording execution. AAR certificates are the proof surface.

When To Use

• A paid Aegis Refine job needs an operator decision, roadmap, or quality gate.
• A dataset needs a cleaning/sanitization/refinement plan before work runs.
• A dataset needs a PII / unsafe-content safety gate before delivery.
• A synthetic augmentation job needs a model/tool plan and proof criteria.
• A spend ticket needs review against the accepted customer cap.
• A completed job needs receipt/certificate verification before delivery.

Prerequisites

• Aegis Refine backend or live site is reachable.
• The job has a real customer, quote, or Stripe checkout session when money is involved.
• Aegis-14B health should be checked through /agent/health before making governance claims.
• Do not fabricate Stripe, spend, model, or certificate state. If a value cannot be verified, say it is unverified.

How To Execute

1. Establish the job context:
   • job id
   • service: refine or synthesis
   • source kind: uploaded file, URL, or seed/reference dataset
   • accepted quote and hard cap
   • customer/account context if available
2. Check Aegis-14B reachability:
   • Live: curl -fsS https://aegisrefine.com/agent/health
   • Local: curl -fsS http://localhost:8000/agent/health
3. Build the operator decision:
   • classify data type and risk
   • record the Nemotron 3.5 Content Safety gate status
   • decide local refinement vs synthetic augment vs OCR/tool escalation
   • identify models/tools required
   • estimate spend and compare against the accepted cap
   • choose whether to continue, queue, request approval, or fail closed
4. Preserve receipts:
   • quote receipt
   • Stripe session/payment state
   • spend ticket decisions
   • verified Stripe Connect transfer id for outbound spend, when executed
   • audit events
   • AAR certificate id/link
5. For completed jobs, send Jason a concise Telegram receipt through Hermes messaging.
6. Produce a concise operator result in JSON only when called by the Aegis bridge.

Reference Quick Commands

Purpose	Command
Check production Aegis	curl -fsS https://aegisrefine.com/agent/health
Check local Aegis	curl -fsS http://localhost:8000/agent/health
Load this skill once	hermes -z --skills aegis-refine "Operate Aegis Refine job <id>..."
Run from repo prompt	hermes -z --skills aegis-refine "$(cat hermes/aegis-refine/templates/operator-prompt.md)"
Create agent spend transfer	python hermes/aegis-refine/scripts/create_stripe_transfer.py --job-id <id> --amount-cents <cents> --service ainode_compute --purpose ocr_enrichment
Send Telegram receipt	hermes send --to telegram "Aegis Refine job <id> completed: quote, spend, AAR"

Operator Decision Schema

Return this JSON shape when operating a job:

{
  "operator": "Hermes Agent",
  "skill": "aegis-refine",
  "job_id": null,
  "service": "refine",
  "aegis_health": "ok|degraded|unverified",
  "primary_models": {
    "operator": "Hermes Agent",
    "operations_brain": {
      "primary": "nvidia/nemotron-3-ultra-550b-a55b",
      "active": "nvidia/nemotron-3-ultra-550b-a55b",
      "fallback": "nvidia/nemotron-3-nano-30b-a3b",
      "fallback_used": false
    },
    "content_safety_gate": "nvidia/nemotron-3.5-content-safety",
    "data_governance": "Aegis-14B"
  },
  "safety_gate": {
    "model": "nvidia/nemotron-3.5-content-safety",
    "status": "passed|flagged|metadata_only|pending",
    "reason": null
  },
  "route": "run_local|synthesize|request_spend|temporarily_queue|fail_closed",
  "cap": {
    "quoted_usd": null,
    "approved_cap_usd": null,
    "projected_spend_usd": null,
    "cap_respected": null
  },
  "spend_decision": {
    "needed": false,
    "tool_or_model": null,
    "reason": null,
    "ticket_required": false
  },
  "proof": {
    "stripe_verified": false,
    "quote_receipt_verified": false,
    "aar_expected": true,
    "delivery_allowed": false
  },
  "spend": {
    "proposed_by": "nvidia/nemotron-3-ultra-550b-a55b",
    "approved_cap_cents": null,
    "projected_spend_cents": null,
    "executed": null,
    "cap_respected": null,
    "verified_against_stripe": false
  },
  "next_action": "continue|queue|ask_operator|block_delivery"
}

Procedure

Refine Jobs

1. Use Aegis-14B for data-domain governance: quality, PII/noise risk, format, usable rows, and whether the data can be processed locally.
2. Use nvidia/nemotron-3.5-content-safety as the PII / unsafe-content safety gate when raw sample text or signed safety summaries are available.
3. Use nvidia/nemotron-3-ultra-550b-a55b for the operator roadmap when latency permits: business route, margin/cap reasoning, tool/model choice, and approval wording.
4. Use nvidia/nemotron-3-nano-30b-a3b only as the documented fast fallback for receipt generation or low-risk routing; do not label a Nano run as Ultra.
5. If local Aegis processing is enough, continue within cap.
6. If external spend is justified and within the accepted cap, create or approve an auditable spend ticket according to the Aegis Refine backend policy.
7. When the bridge phase is spend_approved, initiate the outbound spend as a Stripe Connect Transfer to the configured AINode compute vendor. Use an idempotency key shaped like <job_id>:<purpose> so retries cannot double-spend.
8. Return the transfer id in spend.executed.stripe_transfer_id. Do not set verified_against_stripe to true; the Aegis backend independently retrieves the Stripe object and verifies destination, cap, currency, and livemode before marking the ticket executed.
9. If the Stripe transfer cannot be created or no Stripe skill/tool is available, return spend.executed: null, route: "temporarily_queue", and a concise reason. Never synthesize a tr_... id.
10. If projected spend exceeds the accepted cap, fail closed into a human approval/requote path.
11. Do not deliver if the refined output is empty or the certificate cannot be issued.

Request Spend / Stripe Transfer

Only use this route after Aegis shows the spend ticket is approved by a human or preauthorized by the accepted quote cap.

Required environment:

• STRIPE_SECRET_KEY must be a Stripe test or live secret key for the platform account.
• STRIPE_AGENT_SPEND_VENDOR_ACCOUNT must be the connected AINode compute vendor account.
• MAX_AGENT_SPEND_CENTS should be set as an independent circuit breaker.

Procedure:

1. Read the approved ticket from the job payload receipt: ticket_id, approved_spend_cents, service, and purpose.
2. Clamp the transfer amount to the lower of approved_spend_cents and MAX_AGENT_SPEND_CENTS.
3. Create a Stripe Connect Transfer to STRIPE_AGENT_SPEND_VENDOR_ACCOUNT with metadata {job_id, ticket_id, service, purpose, operator: "Hermes Agent"}.
4. Use idempotency key <job_id>:<purpose> or <job_id>:ticket-<ticket_id>.
5. Return JSON with:

{
  "route": "run_local",
  "spend": {
    "proposed_by": "nvidia/nemotron-3-ultra-550b-a55b",
    "approved_cap_cents": 1200,
    "projected_spend_cents": 1200,
    "executed": {
      "stripe_transfer_id": "tr_...",
      "amount_cents": 1200,
      "currency": "usd",
      "destination": "acct_...",
      "service": "ainode_compute",
      "livemode": false
    },
    "cap_respected": true,
    "verified_against_stripe": false
  },
  "next_action": "continue"
}

The backend will reject the execution if the Stripe transfer cannot be retrieved, exceeds the approved cap, or goes to the wrong connected account.

Synthesis / Augment Jobs

1. Treat the uploaded/reference file as the grounding source.
2. Generate only examples that are plausibly useful for the target training task.
3. Keep provenance: target rows, candidates generated, rows kept, models used, and spend.
4. If zero rows are kept, fail the job instead of signing an empty deliverable.

Receipts And Proof

Before calling a job complete, verify:

• The job belongs to the authenticated customer or an admin operator is reviewing it.
• Stripe payment or checkout state is attached when the job is paid.
• The accepted quote/cap is present.
• Spend tickets show proposed/approved/executed/rejected status.
• The safety gate status is recorded. If no raw data was sent to the content-safety model, mark it metadata_only or pending instead of claiming a full scan.
• Downloadable output is non-empty.
• AAR certificate exists or is explicitly pending.
• Telegram receipt was sent for completed jobs, or the send failure is reported without blocking delivery.

Telegram Receipts

When a job phase is completed, send a short Telegram receipt through Hermes:

hermes send --to telegram "Aegis Refine job <id> completed: quote $<amount>, cap $<cap>, spend $<spent>, AAR <path>"

Never include raw dataset contents, PII, private customer data, or secrets in Telegram. Include only job id, service, route, quote/cap, executed spend, status, and the protected AAR/download location.

Known Issues

• If /agent/health is degraded, do not run a fake fallback. Return temporarily_queue.
• If the web app is reachable but admin receipt endpoints require a browser session, report that the receipt is protected rather than guessing.
• If Stripe state is missing, do not claim the job earned revenue.
• If outbound spend lacks a verified Stripe transfer, describe it as temporarily queued, not Stripe money movement.
• If the transfer id is present but backend verification has not completed, do not call it verified.
• If nvidia/nemotron-3-ultra-550b-a55b is unavailable or too slow, record the actual fallback model. Do not write "Ultra" over a Nano run.
• If nvidia/nemotron-3.5-content-safety did not inspect raw/summarized safety evidence, report metadata_only or pending; do not claim it sanitized the data.

Verification

Minimum proof for a demo:

1. Hermes Agent invocation shows --skills aegis-refine.
2. The skill returns operator: Hermes Agent.
3. /agent/health returns model_name: Aegis-14B.
4. The result names the route and cap decision.
5. The web app shows the matching job, spend/audit trail, output, or queued/failure state.