Skip to content

Proof

Evidence, approach, and resources

Three kinds of material live here, each labeled for what it is: evidence you can check, explanations of how the systems are built, and reference and commercial documents.

failclosed: the Authority Gate, made runnable

failclosed applies the Authority Gate to the merge boundary. It runs an LLM reviewer, distrusts the verdict, and refuses to admit unparseable, schema-invalid, or self-contradictory output.

L1: Authority GatePublic. Runnable in two minutes. Receipt-backed.

Checkable Claims

Evidence. Every checkable claim below carries the command or source that confirms it. The limits are labelled as limits.

Independently validated

Someone else's system attests to these. None of them rests on my own word.

  • Proctor was named a finalist in UiPath AgentHack 2026, Track 3 (UiPath Test Cloud). 31 finalists were selected from the 203 solutions that reached judging, and Track 3 accounted for 11 of them.

    https://forum.uipath.com/t/this-years-uipath-agenthack-finalist-teams-are-here/5762660check it
  • A pull request authored by OrionArchitekton is merged into Arize-ai/openinference, a public repository with 1,238 stars, read 2026-09-29.

    gh api repos/Arize-ai/openinference/pulls/3238 --jq '{merged,author:.user.login}' && gh api repos/Arize-ai/openinference --jq .stargazers_countcheck it
  • A pull request authored by OrionArchitekton is merged into punkpeye/fastmcp, a public repository with 3,272 stars, read 2026-09-29.

    gh api repos/punkpeye/fastmcp/pulls/275 --jq '{merged,author:.user.login}' && gh api repos/punkpeye/fastmcp --jq .stargazers_countcheck it

Publicly reproducible

My own CI and releases. The claim is that you can re-run every check yourself, not that a third party audited it. Counts are dated readings; a check run later returns the current count.

  • 602 tests pass in CI across 7 public repositories: notary 159, reprise 105, orion-skills 98, schemafit 90, mcp-context-budget 56, standing-questions 55, fork-around-find-out 39, read 2026-09-29.

    for r in notary reprise schemafit mcp-context-budget standing-questions fork-around-find-out orion-skills; do id=$(gh run list -R OrionArchitekton/$r --branch main --status success --limit 1 --json databaseId --jq '.[0].databaseId'); echo "== $r"; gh run view $id -R OrionArchitekton/$r --log | grep -aE '[0-9]+ passed|Tests [0-9]+ passed|Ran [0-9]+ test'; donecheck it
  • failclosed contains exactly 59 test definitions across its 3 test files, matching the 59 tests its CI run executes.

    for f in test_graphql_helpers.py test_parse_codex_output.py test_resolve_bot_threads.py; do gh api repos/OrionArchitekton/failclosed/contents/$f --jq .content | base64 -d | grep -cE '^[[:space:]]*def test_'; donecheck it
  • schemafit publishes 5 released versions on PyPI, 0.1.0 through 0.5.0, MIT licensed.

    curl -s https://pypi.org/pypi/schemafit/json | python3 -c "import json,sys;d=json.load(sys.stdin);print(sorted(d['releases']), d['info']['license'])"check it
  • mcp-context-budget publishes 1 released version on PyPI, 0.4.0, MIT licensed.

    curl -s https://pypi.org/pypi/mcp-context-budget/json | python3 -c "import json,sys;d=json.load(sys.stdin);print(sorted(d['releases']), d['info']['license'])"check it
  • Both published packages release through PyPI Trusted Publishing. Both release workflows request an OIDC id-token and reference no API-token secret.

    for r in schemafit mcp-context-budget; do echo "== $r"; curl -s https://raw.githubusercontent.com/OrionArchitekton/$r/main/.github/workflows/release.yml | grep -nE 'id-token|secrets\.'; donesource 1source 2
  • The quality gates publicly fire rather than wave work through: reprise's CI history on main includes a push run where the Typecheck step failed the build, on 2026-07-26.

    curl -s https://api.github.com/repos/OrionArchitekton/reprise/actions/runs/30221340528 | python3 -c "import json,sys;r=json.load(sys.stdin);print(r['head_branch'],r['event'],r['created_at'][:10],r['conclusion'])"; curl -s https://api.github.com/repos/OrionArchitekton/reprise/actions/runs/30221340528/jobs | python3 -c "import json,sys;j=json.load(sys.stdin)['jobs'][0];print(j['name'],j['conclusion'],[s['name'] for s in j['steps'] if s['conclusion']=='failure'])"check it
  • failclosed's enforcement contract is runnable by any stranger: a fresh clone passes its 59-test suite, which includes tests asserting that unparseable and schema-invalid reviewer output blocks a MERGE_READY verdict, read 2026-09-29.

    git clone --depth 1 https://github.com/OrionArchitekton/failclosed && cd failclosed && python3 -m venv .venv && .venv/bin/pip install -q pytest && .venv/bin/python -m pytest -o addopts='' -qcheck it
  • The orion-skills library publicly catalogs 27 agent skills as individually readable SKILL.md files, read 2026-09-29.

    curl -s "https://api.github.com/repos/OrionArchitekton/orion-skills/git/trees/main?recursive=1" | grep -oE '"skills/[^"]*/SKILL.md"' | wc -lcheck it
  • Three repositories pin every third-party GitHub Action to a full commit SHA rather than a floating tag: notary, reprise, and fork-around-find-out. Each has exactly one workflow file, so the check covers all of them.

    for r in notary reprise fork-around-find-out; do echo "== $r"; gh api repos/OrionArchitekton/$r/contents/.github/workflows --jq '.[].name'; gh api repos/OrionArchitekton/$r/contents/.github/workflows/ci.yml --jq .content | base64 -d | grep 'uses:'; donesource 1source 2source 3

Published work and live surfaces

Self-published and self-hosted. These establish volume, recency, and that the systems serve live. None of it is reach, endorsement, or third-party review. Counts are dated readings; a check run later returns the current count.

  • 51 individually addressable articles are self-published on danmercede.com: 42 essays under /thoughts and 9 long-form guides under /guides, each enumerated in the public sitemap, read 2026-09-29.

    curl -s https://www.danmercede.com/sitemap.xml | grep -oE 'danmercede.com/(thoughts|guides)/' | sort | uniq -ccheck it
  • 53 articles are published on dev.to under the danmercede account, dated 2026-06-20 through 2026-08-03, and 50 of the 53 declare a rel=canonical URL pointing back to danmercede.com or danmercede.online, read 2026-09-29.

    curl -s "https://dev.to/api/articles?username=danmercede&per_page=200" | python3 -c "import json,sys;a=json.load(sys.stdin);print(len(a),min(x['published_at'][:10] for x in a),max(x['published_at'][:10] for x in a),sum(1 for x in a if str(x.get('canonical_url')).startswith(('https://www.danmercede.com/','https://www.danmercede.online/'))))"check it
  • The Hashnode publication danmercede.hashnode.dev carries 36 published posts, and all 36 declare a canonical URL pointing back to danmercede.com or danmercede.online, read 2026-09-29.

    curl -s -X POST https://gql-beta.hashnode.com -H 'Content-Type: application/json' -d '{"query":"query{publication(host:\"danmercede.hashnode.dev\"){posts(first:50){totalDocuments edges{node{canonicalUrl}}}}}"}' | python3 -c "import json,sys;p=json.load(sys.stdin)['data']['publication']['posts'];print(p['totalDocuments'],sum(1 for e in p['edges'] if str(e['node']['canonicalUrl']).startswith(('https://www.danmercede.com/','https://www.danmercede.online/'))))"check it
  • A public speaker profile exists on Sessionize, listing speaking topics across AI governance and agentic systems. A self-managed profile, not third-party recognition.

    https://sessionize.com/dan-mercede/check it
  • Each of the 19 product pages under danmercede.com/works serves live with its own product title, asserted against an expected-title map rather than a status code.

    set -euo pipefail; while IFS='|' read -r s want; do got=$(curl -sL --fail --show-error https://www.danmercede.com/works/$s/ | grep -o 'title>[^<]*' | head -1); case "$got" in *"$want"*) echo "OK $s";; *) echo "FAIL $s: $got"; exit 1;; esac; done <<'EOF' agent-demo-video|automated, narrated, captioned demo videos algorithm-reviews|a review whose reviewing is reviewable codex-rule-ledger|evidence-bound audits for Codex runs engram|a memory engine for AI agents failclosed|Fail-closed merge admission control fork-around-find-out|speculative execution for agent safety invisible-hand|an agent economy where on-chain settlement localfiscal|Local-first private receipt mcp-context-budget|enforce MCP tool-surface budgets notary|the context lie detector orion-skills|skills for Claude Code plainspeak|AI that reads the fine print proctor|behavioral regression testing for AI agents quorum|the decision memory for your Slack workspace rekindle|your abandoned side project misses you reprise|check what you already generated schemafit|Lint LLM structured-output standing-questions|ask a live stream once whisperways|noise-aware eVTOL flight corridors EOFcheck it
  • danmercede.online, the public raw working log, lists 149 dated entries in its sitemap, read 2026-09-29.

    curl -sL https://www.danmercede.online/sitemap.xml | grep -cE '/[0-9]{4}-[0-9]{2}-[0-9]{2}-'check it
  • Six of the shipped tools serve live public demos, read 2026-07-27: Reprise, Fork Around & Find Out, Standing Questions, Quorum, Notary, and Engram.

    set -euo pipefail; while IFS='|' read -r u want; do got=$(curl -sL --fail --show-error https://$u | grep -o 'title>[^<]*' | head -1); case "$got" in *"$want"*) echo "OK $u";; *) echo "FAIL $u: $got"; exit 1;; esac; done <<'EOF' reprise-murex.vercel.app|Reprise fork-around-find-out.vercel.app|speculative execution for agent safety standing-questions.vercel.app|Standing Questions quorum-slack-agent.vercel.app|Quorum notary-replay.vercel.app|Notary engram.orionbot.online|a memory engine for AI agents EOFsource 1source 2source 3source 4source 5source 6
  • Orion Intelligence Agency, the consulting lane I operate, publishes 23 individually addressable insight articles on its live site, each listed in the public sitemap, read 2026-09-29.

    curl -s https://www.orionintelligenceagency.com/sitemap-0.xml | grep -c 'insights/'source 1source 2
  • OIA's case-studies page publishes anonymized engagement shapes under an explicit disclaimer: Not invented guarantees. They show how we work, not promised metrics.

    curl -s https://www.orionintelligenceagency.com/case-studies | sed 's/<[^>]*>/ /g' | tr -s ' ' | grep -oE 'Not invented guarantees|These are anonymized engagement shapes|They show how we work, not promised metrics' | sort -ucheck it
  • Five primary pages of the OIA site (home, about, services, contact, and case-studies) each disclose the operating entity: Orion Intelligence Agency, LLC, founded 2025 in Ohio. Disclosure on my own site, not an independent registry record.

    for p in "" about services contact case-studies; do printf '/%s: ' "$p"; curl -s "https://www.orionintelligenceagency.com/$p" | grep -c 'Orion Intelligence Agency, LLC, founded 2025 in Ohio'; donesource 1source 2source 3source 4source 5

What this page does not claim

The limits are part of the evidence. A proof page that lists only wins is marketing.

  • Self-run CI is not third-party review. The test counts above come from my own pipelines. Reproducibility is the claim; independent audit is not.

    Stated limitation
  • 6 of the 14 tool repositories ship no CI at all: engram, proctor, agent-demo-video, invisible-hand, plainspeak, and whisperways.

    for r in failclosed orion-skills schemafit mcp-context-budget engram notary proctor reprise agent-demo-video invisible-hand plainspeak standing-questions whisperways fork-around-find-out; do printf "%s " "$r"; gh api repos/OrionArchitekton/$r/contents/.github/workflows --jq '[.[].name]|join(",")' 2>/dev/null || echo "NO CI"; donecheck it
  • Fork Around & Find Out was submitted to Daytona HackSprint #5 and won no prize. It appears in the public project gallery with no winner badge.

    https://devpost.com/software/fork-around-find-outcheck it
  • Invisible Hand is listed on the SwarmHack gallery with judging locked and no award, prize, or placement.

    https://tokensand.com/p/invisible-handcheck it
  • No paying-customer outcomes are published here. The reference architectures on this page are illustrative patterns, not engagement records, and each download opens by saying so.

    Stated limitation
  • No third-party editorial placement exists to date. Every article above is self-published or syndicated under my own canonical URLs, and no conference talk has been delivered to date.

    Stated limitation

How the systems are built

Approach. These diagrams are conceptual. They explain the design and are not measured outcomes. For the platform concept, see Cosmocrat OS: Platform.

See fail-closed execution in action

Interactive example. Compare advisory and gated decisions and inspect the simulated receipt.

Advisory vs. fail-closed

Advisory only

Unconstrained driftMerge can proceed with open findings

Fail-closed

Authority gateHalted with receipt

Merge admission control for AI-written code

L1 · Authority Gate before merge

Contradictory reviewer output

Reviewer JSON payload

{
  "verdict": "MERGE_READY",
  "blocking_findings": 2,
  "tests_verified": false,
  "notes": "Auth check skipped on retry"
}
PASSStrict JSON parse
JSON structure is syntactically valid
PASSSchema validation
Required keys are present
FAILSelf-consistency
Claims MERGE_READY with 2 blocking findings
Gate verdictFAIL_CLOSED

Instead of trusting the optimistic label, the deterministic gate halts merge and records the contradiction.

Emitted decision receipt

{
  "gate": "merge-admission",
  "decision": "FAIL_CLOSED",
  "admitted": false,
  "reason": "contradictory_verdict_with_blocking_findings",
  "blocking_findings": 2
}

Run the real engine locally: Regression-test an agent after a prompt or model change.

Runtime Execution Control Plane Architecture, a conceptual diagram: Authority Gate, Immutable Receipts, Drift Guard, Gated Substrate

The design places governance at four boundaries. Authority. Attestation. Behavioral Constraint. Physical Isolation.

Evaluated before mutation.
Ambiguity defaults to halt.
No receipt, no commit.

Interactive Architecture Explorer · Select a layerBoundary L1 active

L1 · Authority Gate · Pre-mutation authorization check

Verifies parsed schema, structural contract, and verdict consistency before any merge or state-changing action runs.

Inspect failclosed

Gated Execution Pipeline

Conceptual. PLAN → EXECUTE → REVIEW → APPROVE: the design checks each boundary.

In this design, each state mutation passes three gates.

Gated Execution Pipeline, a conceptual diagram: PLAN, EXECUTE, REVIEW, APPROVE with enforcement points at each boundary

Gate 1: Authority → Gate 2: Attestation → Gate 3: Behavioral Constraint.

How Governance Could Change the Economics of Risk

Economics (conceptual). Cost per successful task. Escalation rate. Cycle-time compression. Audit defensibility.

The framework describes how governance could turn operational risk into predictable unit economics.

Conceptual economics framework: cost per successful task, escalation rate reduction, cycle-time compression, audit defensibility metrics

Reference Enforcement Architectures

Reference enforcement architectures by industry and regulatory surface. Illustrative patterns, not measured client outcomes.

Case Study Template

Blank template · DOCX

Structured engagement template mapping findings to enforcement stack layers. Captures enforcement points deployed, behavioral drift metrics, containment thresholds, and quantified risk reduction tied to operational blast radius.

Download

Working together

Commercial and speaking. The speaker one-sheet describes talks, not results. For hiring, collaboration, or implementation work, use the Contact page.

Speaking One-Sheet

Speaker one-sheet · PDF

One page for event organizers: bio, the upcoming OpsWorld 2027 workshop, the NODES 2026 talk, the UiPath AgentHack finalist result, four talk topics, and a direct speaking contact.

Download

Start with one workflow and clear ownership.

Workflow readiness map. Failure-mode heatmap. Ownership and handoff plan. 30/60/90 rollout roadmap.

Work with OIA on one workflow

Follow new field notes

Get new essays and technical guides in your feed reader.