I build self-running AI swarms
that leave a signed receipt for every move.
I am Amanda Morton. reckon is a fleet of AI agents that picks up its own work, ships it, remembers what it learned, and gets smarter the more it runs — no babysitting, no orchestrator bottleneck. Every action is ed25519-signed and hash-chained into a tamper-proof receipt, four layers deep to a public anchor. Works with any model. Scales from one agent to thousands. This is trust and safety built into the substrate, not bolted on after — and it is running live, right now, below.
The Product
reckon: a stigmergic coordination platform for multi-agent AI, with a signed, tamper-evident chain of custody on every action. Everything below is the doctrine, made visible.
Systems architect
I am the sole designer of this platform’s doctrine — a stigmergic model where dozens of agents coordinate through a shared-filesystem blackboard instead of message passing, carried from first principle to shipped product. Everything below is that doctrine, made visible.
reckon — self-running AI agent swarms, with a receipt for every move
reckon turns a pile of LLMs into a fleet that picks up its own work, ships it, remembers what it learned, and leaves a tamper-proof receipt for every move — no babysitting, no orchestrator bottleneck. Works with any model. Scales from one agent to thousands. The name comes from dead reckoning: each agent knows its last confirmed state and its compass bearing, and commits without waiting for a dispatcher.
Not four vague pillars — seven real mechanisms, each proven by a dated incident from this estate's own operating history, not a marketing claim.
01 · Sortie
Claude's native Workflow tool run inside a fixed lifecycle — survey, design, execute one lane per repo, then an adversary re-runs every proof command. Hard cap 6 agents, no recursion.
Proven by: every sortie must extend a real CI check on the way out — "none" needs a stated reason, never silence.
02 · Custody chain
Every commit appends one ed25519-signed, hash-chained leaf — no exceptions. Four tiers: entry-hash → Merkle rollup → git SHA → a public Rekor anchor.
Proven by: measured 2026-09-22 — 86,981 leaves across all shards vs. 32,995 in the rolled-up spine, the exact gap the rollup tier exists to close.
03 · Shake
The one signed message channel between runtimes. ack means the task is done, never merely "seen" — an acked-but-unactioned frame vanishes from attention on purpose.
Proven by: a bare queue-check used to guess which runtime was asking and got it wrong for two days straight (18 frames unread); it now prints both queues, labelled.
04 · Metabolism
Evolve (mutate → measure → keep/rollback) and crystallize (N variants → one canonical form) as a live, loss-gated cycle — not a one-time cleanup pass.
Proven by: a change only gets promoted with a measured verdict of BENEFICIAL and zero teeth lost — checked, not asserted.
05 · Braid / parity
Named code chunks stay byte-identical across four sibling repos; drift is reported per chunk, never as a blanket copy that silently deletes another repo's work.
Proven by: a guard script forked three ways (8183 / 4277 / 6799 bytes) because it was never enrolled — the 4277-byte copy silently no-op'd in 5ms while reporting PASS.
06 · TRL ladder
Every CI job maps to the readiness tier it actually proves; a verdict reports the highest contiguous tier, so a green downstream job can't paper over a red upstream one.
Proven by: self-correcting by design — an early tier map over-claimed a rung for jobs that never stood up a live service, caught by re-reading what the steps actually run.
07 · Board / frontier
Agents coordinate by leaving signals in a shared, git-tracked field (stigmergy) — claim, steer, handoff — instead of point-to-point messages through a central orchestrator.
Proven by: direct agent-to-agent messaging is hard-blocked at the hook level, on purpose — every coordination signal leaves a trace someone else can read later.
"The fleet provides a compass, a frontier, and a chain of custody. The agent supplies the judgment." — reckon agent doctrine
Four more things I built end to end
Reckon is the foundation; these grew out of different parts of it — the same signed, auditable-by-design posture, applied to four complementary problems.
Mantle
The operating layer for an owner-run rental business — rent by bank sign-in or e-Transfer, tenants and leases (scan an existing lease in), listings with AI drafting, consented background checks, a to-do list that ticks itself off as messages are answered, and hash-chained proof for every message and payment. The demo is your own private sandbox: fictional data, nothing leaves it.
Try the live demoAtropos
A gamified OSINT and cybersecurity training platform — 23 investigation campaigns, a Rust recon engine wrapping real tooling (Amass, Nuclei, Gitleaks, TruffleHog), and an AI investigation assistant that adapts to a learner's skill level.
View on GitHubCommand Center
Drag agents and tools onto a network-topology canvas and it exports real, runnable CrewAI Python — a live agent-runtime feed, cost-tracked evaluation, and a sandboxed malware-analysis workflow for the purple-team use case it was built around.
View on GitHubGuardianNet
A public-domain censorship-resistance toolkit — a real Nostr P2P chatroom, IPFS content-addressing, a dead man's-switch relay chain, and a browser-based LoRa mesh hardware playground. Four of its real mechanisms run live below the link — not screenshots.
Try the live demoExponential potential.
Rocket efficiency. Zero orchestration drag.
Subagents unlock super-linear session potential through cross-pollination — every additional agent doesn't just add to the fleet, it multiplies discovery. But that multiplier evaporates fast when orchestration overhead fills the main context. f(0) burning at 2.3% puts the system on the steepest possible liftoff trajectory.
as f(0) → 0
at f(0) = 2.3%
f(0) = 50% (wasted)
Chain walk — entry-by-entry
Fetches the signed chain-demo.json export, recomputes every entry hash in the browser using SubtleCrypto.digest, checks prev_entry_hash linkage, and reports PASS or MISMATCH per entry — no server involved.
- entry-000 file_write
prev genesishash 8928d3f8fc…e4364c - entry-001 file_write
prev 8928d3f8fc…e4364chash 58f5d9a00c…df0abb - entry-002 file_write
prev 58f5d9a00c…df0abbhash 1764d3e3da…a5fdc1 - entry-003 file_write
prev 1764d3e3da…a5fdc1hash ccdbea4d13…54a517
Walk all 12,431 records, then break one →
Source: /public/chain-demo.json
Merkle forest
Each agent session produces a chain of signed entries. Those chains are rolled into Merkle trees and the whole forest is anchored to a public transparency ledger, so the entire history is independently verifiable and tamper-evident by construction. Below is the real chain — 12,431 entries across 41 trees, hue by lineage. Drag to pan, scroll or pinch to zoom, and tap any tree to walk its branches.
Three things I built, working right here
A system that watches itself
I built the self-measurement framework below: it records trajectory against templates and flags its own regressions, backed by 22+ architectural doctrines that auto-load as in-context guardrails.
Signal from a large investigative corpus
Under criticalexposure.report I author public forensic investigations built on the same integrity engineering as reckon — hash-chained evidence with a documented chain of custody, and reproducible evidence packages any reader can independently re-derive.
Evidence-weighted, Bayesian-style confidence scoring across a large cybersecurity investigation dataset — network-infrastructure and exfiltration-pattern analysis — disciplined by Bonferroni-corrected hypothesis testing, disclosing the non-significant results alongside the positive ones so the analysis cannot cherry-pick. The method, not a conclusion, is the point.
Evidence-weighted confidence scoring across an investigative dataset (269 normalized, ed25519-signed datasets; chain-of-custody spine). A methodology note, not a policy claim — the null result (H5) is disclosed rather than omitted, which is the point.
The Forensic Fabric
An agentic-AI substrate built so that its own record can be checked by someone who does not trust it. Keys are generated by the user and never held by us. Work is signed. The record only grows, and tampering with it breaks something a machine can see.
The core insight: the substrate carries the coordination. Termites don’t need a foreman.
Writing
Case studies from the workbench — a teaser, not the whole feed. Read the rest at /blog →
Reward Without a Sign on It
Tell an optimizer exactly what earns it a higher score and it will find the shortest path to that score, not to the thing the score was supposed to measure. The fix isn't a smarter metric — it's never showing the metric to the thing being measured.
Read the note 02The Swarm's Own Metabolism
A hard concurrency cap either wastes capacity or floods the system — there is no setting that is right for both a quiet hour and a busy one. The fix borrows an idea from biology: let the backlog itself apply the brake, so the system settles at its own equilibrium instead of fighting a wall someone guessed at.
Read the note 03Verbal Cryptography: Keys You Can Say Out Loud
Authentication With Nothing to Steal — a stolen password works for whoever stole it, a stolen token for whoever holds it. The interesting factor is the one that can't be copied or handed off, because answering it requires actually doing the transformation in your head, every time.
Read the note 04Two Clocks, No Bottleneck
A tamper-evident record has two jobs that want opposite update rhythms — react instantly to what just happened, and prove to the outside world that nothing has been reordered. Trying to serve both off one clock is why systems like this usually stop scaling long before 'planet-scale' is a serious phrase.
Read the noteCase studies from the workbench
A few pieces on the engineering behind reckon, and on the plain problems small businesses actually get hit by. Read them all →
Reward Without a Sign on It
Tell an optimizer exactly what earns it a higher score and it will find the shortest path to that score, not to the thing the score was supposed to measure. The fix isn't a smarter metric — it's never showing the metric to the thing being measured.
The Swarm's Own Metabolism
A hard concurrency cap either wastes capacity or floods the system — there is no setting that is right for both a quiet hour and a busy one. The fix borrows an idea from biology: let the backlog itself apply the brake, so the system settles at its own equilibrium instead of fighting a wall someone guessed at.
Verbal Cryptography: Keys You Can Say Out Loud
Authentication With Nothing to Steal — a stolen password works for whoever stole it, a stolen token for whoever holds it. The interesting factor is the one that can't be copied or handed off, because answering it requires actually doing the transformation in your head, every time.
Two Clocks, No Bottleneck
A tamper-evident record has two jobs that want opposite update rhythms — react instantly to what just happened, and prove to the outside world that nothing has been reordered. Trying to serve both off one clock is why systems like this usually stop scaling long before 'planet-scale' is a serious phrase.
No Manager, No Meeting: How Agents Coordinate Through the Record
Ant colonies coordinate complex construction with no supervisor, no plan and no meetings. The mechanism is marks left in a shared environment — and it is the same mechanism that lets a swarm of software agents work one codebase without a scheduler.
What a tamper-evident record actually buys you
Everyone says they keep logs. The useful question is whether anyone could have quietly changed them — and whether you could tell.
Hire Me
Work history and credibility — research leadership, fraud-detection tooling, a skills matrix, and role-targeted résumés.
Every shipped behavior, backed by evidence
I built RLHF-tuned classification and response models for a high-stakes intervention domain, where an ordinary helpfulness-tuned model actively harms the user. I ran a cross-functional data-science, legal, and PR team, ran A/B tests under ethical review, and briefed FBI, CISA, and national-security stakeholders — nothing shipped without a documented evidence base behind it.
Anomalies don’t announce themselves
Freelance ML tooling (Python / SQL / Tableau) surfacing transaction anomalies and fraud across geopolitical and financial contexts, fed by real-time dashboards — every finding carrying a documented, tamper-evident trail.
A T-shaped profile built over a decade
Most engineers bolt safety on at the end. Most security researchers can't code the solution. The T-shape here isn't a résumé accident — four domains that lock together into one profile that doesn't otherwise exist on the market.
AI Engineering
Sole architect of reckon: stigmergic multi-agent orchestration, MCP server (37+ tools), four-layer enforcement stack, shape registry self-measurement, piston-wave context pressure. OpenHands SDK, FastMCP, Anthropic + OpenRouter APIs.
Trust & Safety
Project Gravity (2023–2024): led development of AI-enabled CVE toolkit with explicit constraints around NOT amplifying harmful behavior. RLHF-tuned models for high-stakes intervention. Briefed FBI, CISA, national-security stakeholders.
Threat Intelligence
Real-time threat-monitoring dashboards (Python/SQL/Tableau). Blockchain analytics for cryptocurrency misuse by criminal and terrorist networks. OSINT investigation. ML anomaly detection for fraud in geopolitical and financial contexts.
AML & Compliance
AML/ATF/Sanctions compliance expertise. CAMS certification in progress. Forensic chain-of-custody design — every agent action in reckon is ed25519-signed and hash-chained, producing an auditable ledger any regulator or court can examine.