Eigenstate Research MCP Server

Eigenstate Research

Site · Methodology · Verify walkthrough · Ledgers · Agent index

Independent measurement of tokenized fixed-income and RWA settlement structure. Observe, measure, verify, publish. This repo is the public face: reports, dossier mirrors, measured ledgers, MCP source, and provenance. The runtime engine is private.

For engineers and analysts who need a claim they can overturn, not a dashboard they have to trust. Every published rate has a machine-readable ledger behind it, every refusal names the limb that failed, and the verifier fails closed on invented attestations.

One command, no install, no account:

curl -sS -G 'https://geniusflow-federation.vercel.app/api/verify' \
  --data-urlencode 'source_url=https://docs.aave.com/developers/aave-v3/overview' \
  --data-urlencode 'expected=v3'

Returns status: "ATTESTED-PRIMARY", found: true, proof_shape.admitted: true. Change expected to a string the source does not contain and the same endpoint refuses by name. The difference between those two responses is the whole thesis.

Contents

1. Quick start
2. What it measures
3. How it works
4. Measured ledgers
5. Public surfaces
6. Who it is for
7. Status, limits, and non-goals
8. Operating this repo
9. Documentation and evidence index
10. Contact
11. Appendix A: agent surfaces
12. Appendix B: Host operational log

Quick start

Admit and refuse in 60 seconds

The command at the top of this file is the admit case. Here is the same endpoint refusing, with one argument changed:

curl -sS -G 'https://geniusflow-federation.vercel.app/api/verify' \
  --data-urlencode 'source_url=https://docs.aave.com/developers/aave-v3/overview' \
  --data-urlencode 'expected=not-in-this-page'

Both calls return HTTP 200. The refusal lives in the JSON body, not in the status code, because a consumer that only checks status codes would read a refusal as a pass. Pipe through jq if you have it.

The full six-step walkthrough (status, dossier, /api/package admit, /api/package refuse, grounding verify, return-wire ack) is at verify-walkthrough/.

Run the demo in 5 minutes

A toy stack, not the private engine. Python 3.8 or newer, no API keys, no accounts.

From a clone of this repo:

python3 -m venv .venv && . .venv/bin/activate
pip install helixhash
python3 public/demo/eigenstate_demo.py

Without cloning, in any empty directory:

python3 -m venv .venv && . .venv/bin/activate
pip install helixhash
curl -sO https://kaydeep0.github.io/eigenstate-research/demo/eigenstate_demo.py
python3 eigenstate_demo.py

The venv is what makes pip install work on a stock macOS or Debian box, where a system-wide pip install is refused as an externally managed environment. Only helixhash is needed; the script uses the standard library for everything else.

macOS cold start. If your default python3 is Homebrew's 3.14, the venv above can die at pip import time before it installs anything. Pin the interpreter instead:

python3.12 -m venv .venv && source .venv/bin/activate
pip install helixhash
python3 public/demo/eigenstate_demo.py

The demo itself is fine on 3.12. If you would rather repair the 3.14 toolchain than pin around it, brew reinstall expat python@3.14 is the fix.

On macOS, a Homebrew Python whose libexpat has drifted from the interpreter fails at import time, before any of this code runs:

ImportError: dlopen(.../pyexpat.cpython-314-darwin.so): Symbol not found:
_XML_SetAllocTrackerActivationThreshold

That is a broken local toolchain, not a problem with the demo. Either pin the venv to a known-good interpreter, python3.12 -m venv .venv, or repair the Homebrew one with brew reinstall expat python@3.14. The demo is verified against CPython 3.12 with helixhash 0.1.1 from PyPI.

What the demo does. It reads three public series (FRED SOFR, DeFiLlama BUIDL TVL, GitHub stars), measures each as delta_I = |log2(current / prior)|, crosses them into a local HelixHash chain, then tampers with one entry in-process so you watch verify() flip to False and back. It ends on the latest Base commit in the bundled index. If a source is unreachable the run still completes on pinned fallback values and labels each observation as live or fallback. The demo is not production cadence, full topology coverage, or continuous on-chain commits.

HelixHash is the underlying primitive and is separately published: Kaydeep0/helixhash · HELIXHASH.md.

What it measures

Tokenized funds, stablecoin rails, and on-chain settlement look transparent at the surface. The hard parts stay opaque: who can settle, under what clock, against which benchmark, with what custody and regulatory path.

Markets price the product. Few parties systematically measure the structure underneath: the obligations, permissions, and dependencies that decide whether settlement actually works. Eigenstate is a research measurement stack for that structure. It does not predict prices or trade markets.

flowchart LR
  T[Topology] --> S[Sensors / Parkash]
  S --> V[Vault<br/>hash chain]
  V --> VER[VERIFY / proof_shape]
  VER --> P[Pages report]
  VER -.->|when gates allow| B[Base commit]

Base attestation is gated, not per cycle. Depth and caveats: METHODOLOGY.md.

E = ΔI / A

Information efficiency: new information gained (ΔI) over observation cost (A). A structural metric, not a forecast.

Example. On an SEC filing day, a new rulemaking document against the prior expected state gives high E. The same feed on a quiet day gives low E.

How it works

Four mechanisms, each with a public artifact you can open.

Vault

Qualifying observations append to a hash-chained measurement log (ΔI / E plus integrity fields). Entity field numbers Φ_S and κ (not Γ) are mirrored on public dossiers. Changing an old vault row breaks later hashes and parent checks.

Example. After a qualifying read of CIRCLE or BLACKROCK, the private vault appends a chained row and the dossiers expose live Φ_S / κ. Public mirrors: CIRCLE · BLACKROCK. Not every vault row becomes a report.

Helix commit and on-chain attestation

Local chaining is continuous. On-chain commits to GeniusFlowSettlement on Base land when wallet, balance, and mirror gates pass. Attested when it lands, not every cycle.

Example. Historical commits: On-Chain Proof Index (machine readable: /api/onchain.json). Sample tx: 0x5a132a48…77a66. Prefer the index over any single hard-coded hash.

Verification rails

Dispatch fact-check, then VERIFY / status_at_publish, then federation proof_shape admit or refuse. Cheap copies that invent "attested" labels fail.

Example. A package with a real registry reference and a provenance spine can admit. One that stamps ATTESTED without that spine refuses. Spec: ATTESTATION_PROOF_SHAPE.md.

Structural gap detection

We look for what the topology needs that product design still treats as optional.

Example. SOFR plus ISDA fallback is often treated as a full LIBOR replacement. The stack still scores a persistent term-structure gap (LIBOR_EQUIVALENT). See predictions/.

Measured ledgers

Category infrastructure plan (cites, not stars): docs/launch/CATEGORY_INFRA_90D.md · Host post-today checklist: docs/launch/POST_TODAY.md.

Five probes that measure a rate instead of asserting one. Each names the limb that failed on every refusal, runs its limbs in a frozen order so the counts partition rather than double-count, and serves a per-subject receipt so a disagreement is a fetch rather than an argument.

Per-subject receipts are served from the federation origin: /erc8004/receipts/<agentId>.json · /slsa/receipts/<normalized-name>.json · /rwa/receipts/<ENTITY>.json · /grounded-claims/receipts/<CLAIM_KEY>.json · /tlog/receipts/<LOG_KEY>.json.

The miss rate is measured, not asserted. Of 500 agentIds drawn uniformly from the 60444 registered on Base at block 49425346, 365 refuse at least one limb the ERC states as MUST (73.0 percent, 95 percent CI 68.9 to 76.7), and 455 refuse once self-reference and endpoint liveness are counted (91.0 percent, CI 88.2 to 93.2). The largest single cause is an on-chain identity carrying no agentURI at all, 174 of 500. Every refusal names its limb, and limbs run in a frozen order, so the counts partition the sample rather than double-counting it. Cite that interval or nothing.

This is not the first measurement of ERC-8004 registration quality and does not claim to be. Xiong et al., Can Trustless Agents Be Trusted? (arXiv 2606.26028), and Mafrur and Khusumanegara, From Agent Identity to Agent Economy (arXiv 2606.12128), reach the same shape from full population crawls: registration heavy, operationally shallow. Two methods and two samples months apart landing in the same place is the useful part; the agreement is the result, not the novelty. What this adds is operational. The sample is redrawable by a third party because the seed is the pinned block hash and nothing else, every rate carries a Wilson interval, and every sampled agentId has a receipt carrying the exact HTTP status or transport error.

Read it: /erc8004/.

A census, not a sample: all 94 distributions pip resolves for this engine's requirements.txt. 44 serve a PEP 740 attestation and all 44 verify against Fulcio and Rekor with a digest computed from downloaded bytes. 50 serve none. 0 carry SLSA build provenance, because every attestation in the set is a PyPI publish attestation, which binds an upload to a workflow and does not describe the build.

The expectations file was committed before the run (/slsa/expectations.json), and the probe refuses its own author first: two host-published roots are declared at versions the public index cannot satisfy. These counts describe one dependency set and are not a PyPI-wide rate.

Read it: /slsa/.

Seventeen real world asset issuers and tokenized instruments that this node already publishes a dossier card for, asked one question: can a program read the disclosed figure without a human, a document parser, or a third party who already did both.

• 0 of 17 serve a machine readable disclosure surface on an origin they operate.
• 2 of 17 have a machine readable surface recorded at all (2 of the 15 recorded surfaces serve structured bytes; the other 13 are HTML documents), and both belong to the same third party aggregator. That is why this engine's own reserve figures for BUIDL and OUSG come from DefiLlama rather than from BlackRock, Securitize or Ondo.
• 0 of 45 requests to a frozen list of conventional disclosure paths answered with structured bytes, because no such convention exists to follow.
• 0 of 15 recorded surfaces carry a chain qualified registry reference, so no disclosure here identifies a single deployment without the reader guessing the network.

Expectations were published in an earlier commit (/rwa/expectations.json), and the survey publishes two defects it found in its own probe. This is not a measure of disclosure quality and not a sector rate: the population is what this node tracks, and eleven of the seventeen refuse at the first limb on this node's own missing coverage. For the issuer-attributable figure, cite attributable_to_issuer (6 in scope, 2 admit, 4 refuse).

Read it: /rwa/.

Census, not sample: the 6 SEC and Fed claims this node already publishes (19 ledger rows behind them). 6 of 6 expected strings are still present at the cited URL today. 5 of 6 clear every MUST limb; 0 of 6 clear MUST and SHOULD together, because every source is an HTML page. 0 of 6 cite a primary filing; 3 of 6 are corroborated by an independent filing body a free index returned, and 0 of 6 by a document the regulator itself authored. 1 of 6 is held at disagreeing verdicts inside this node's own ledger.

Expectations were published before the run (/grounded-claims/expectations.json). This is not a judgement that any claim is true, and not a rate about the SEC or the Fed: the population is this node's whole regulator-grounded surface.

Read it: /grounded-claims/.

Twenty-three public transparency logs, each read twice with a frozen gap, signatures checked against keys obtained elsewhere, and consistency proofs recomputed. 22 of 23 clear every MUST limb; 11 of 11 served consistency proofs recomputed to the second root; 0 trees shrank; 0 disagreements this batch. This node holds no key and cosigns nothing.

Expectations were published before the batch (/tlog/expectations.json). This is not coverage of the transparency-log ecosystem and not an attestation others can rely on: a batch monitor sees the log at the instants it reads it.

Read it: /tlog/.

Public surfaces

Recent public UX (Aug 2026)

• Entity hubs under /reports/<entity>/, with publish-time Φ_S / κ evolution and a series timeline
• Signal Reports index grouped by entity, with expandable dated timelines
• Track Record is a falsifiable Host claim register, not a marketing accuracy percentage; Predictions covers open structural gaps
• Provenance: OpenTimestamps commitment over the helix head (Bitcoin-attested, block 960660) plus a Software Heritage archival request pointer
• MCP public source: mcp/ in this repo; live remote at federation /api/mcp and /server.json

Who it is for

Outsider-facing and grounded in what the public stack actually ships today. Every row names what it is not, because that is the part most projects leave out.

Status, limits, and non-goals

Research, pre-revenue. Public artifacts are real (reports, dossiers, Basescan commits when published). There is no claimed customer book, bank partnership, or production settlement product on this face. Coverage and on-chain mirror paths are partial by design; see METHODOLOGY.md for what is live versus gated.

Non-goals

Source of truth

The source of truth for Host and Pages field tiles is the GeniusFlow engine measurement state: workspace/HOST/state.json (κ, Protocol Truth, parkash stamps) plus the M1 coverage ledger (m1_strict ∩ topology / 199). The site bakes a stamp into public/field-state.json; homepage proof tiles and /track-record read that stamp at build time and again at runtime from the same origin. This is not frozen marketing HTML.

Report heroes overlay live Φ_S / κ from dossier mirrors when present, while the prose stays a dated snapshot. Signal report green badges are vault@publish (trimmed cycle-vault rows), not M1 lifetime observations.

CI fails the build on hardcoded stale denominators (196 or 197) or fake 100 percent accuracy marketing in Astro sources: scripts/check-sot-honesty.sh.

Operating this repo

Host-only. After a parkash, refresh the public surfaces from the engine. Do not hand-edit proof tiles.

# Machine with private engine clone + Vercel auth for federation:
export GENIUSFLOW_ROOT=~/Desktop/GENIUSFLOW_OS/workspace/geniusflow   # or your path
export GENIUSFLOW_BUILDER_DEPLOY=1
export EIGENSTATE_SITE_ROOT=/path/to/eigenstate-research              # optional; auto-detected

./scripts/refresh-public.sh
# writes public/field-state.json
# writes public/federation/dossier/*.json (+ repo-root federation/ mirror)
# rebakes and deploys federation on Vercel when GENIUSFLOW_BUILDER_DEPLOY=1

git add public/field-state.json public/federation federation public/AGENTS.md public/METHODOLOGY.md
git commit -m "Refresh public SoT + dossier mirrors after parkash."
git push   # GitHub Pages deploy workflow

Engine equivalent:

PYTHONPATH=engine python3 engine/tools/refresh_public_surfaces.py --site /path/to/eigenstate-research

Documentation and evidence index

Contact

• Paragraph: paragraph.com/@eigenstate
• GitHub: Kaydeep0/eigenstate-research
• Site: kaydeep0.github.io/eigenstate-research
• Access requests: request-access/

Issues and corrections on this repo are welcome. If you can show a receipt that contradicts a published number, that is the most useful thing you can send.

Appendix A: agent surfaces

Pages stays the human website. Machine cold start is the federation, not this README and not a clone of the private engine.

Canonical start: https://geniusflow-federation.vercel.app/llms.txt

Traverse: hub, then status, then dossier, then /api/package or /api/verify (proof_shape admit or refuse), then optionally /api/return_wire, then the canonical report on Pages. Human version of the same path: verify-walkthrough/.

Core

Network membership

Measured ledgers, machine readable

Headline numbers, intervals, and the caveats that must travel with them are in Measured ledgers. Do not quote a rate without the population it was measured over.

Provenance

Appendix B: Host operational log

2026-08-01/02. MCP registry published io.github.Kaydeep0/geniusflow-federation 1.0.0 (a duplicate-version re-publish is expected to be refused). OTS helix-head verified via lite-client at Bitcoin block 960660 (proof: public/provenance/opentimestamps/helix-head-917f0e3036931e14.txt.ots). A2A day-5 probe-receipt comments posted.