06 of 09

Live evidence — calibrated, currently true

~5 min read

What the substrate currently claims, with provenance.

Every figure on this page comes from substrate-state.json — the substrate's machine-readable self-attestation document, regenerated every nightly harvest, served at /state.json. Every claim names the ADR that governs it. The substrate is in pre-Phase-2 stealth: the public surface is this page plus the JSON; full source + audit-trail access is NDA-gated until first-paying-pilot validation. The audit IS the build, just currently within bounded reach.

Generated at

18 May 2026, 04:39 UTC

Schema · Substrate

promethean-state-1.0 · v0.1.0

Observation window

CVE corpus 0
Active corpus 66
Signatures 5,677
ADRs 50
PASSall claims have audit trailPASSevery high-confidence claim has n ≥ 24PASSevery convergence is cross-stream

⚠ Unsigned · this state.json carries no Ed25519 attestation

Trust degraded. Regenerate via npm run substrate:state with the attestation private key in place.

✓ Promotion-log transparency surface · 66 entries

last entry 2026-05-10T16:59:47.237Z
algosha256 · canonical-form v1 · append-onlyhead69480a3b7bff394f(64 hex chars)anchorcommitted in state.json (covered by Ed25519 signature above)

Verify the full chain + anchor: npm run substrate:verify-log -- promotion-log.jsonl --state=state.json. Tampering with any entry breaks the chain. Appending entries past the signed state requires re-signing state.json.

✓ Corpus Merkle tree · 66 entries committed

algosha256 · canonical-form v1 · per-entry leaf hashroot8e7286eee18147ce (covered by L1 signature)

Inclusion proof for any entry: npm run substrate:verify-corpus -- corpus.jsonl --signature="…" --state=state.json

Honeypot canaries · 5 defined · 0 hits (§47.3 gates holding)

manifest65ece45b94341ea1 (silent removal breaks this hash)scanned2026-05-18T04:39:03.596Z

✓ Per-stream signing keys · 7 streams · trust-weighted §47.6

public-harvestw=0.70+WyC/QMnFU1hFHtPW8dopDg4cve-harvestw=0.90QxieWANZzJD9bRoVzZGZe7Z1fix-commitw=0.85Hpg/F9QJGgh53Txj1GylUS/Yghsaw=0.85F95/pMzUqNnFW/WVD+ZewN+wstackoverfloww=0.40/JHPf7Fv1MMGfdvYrCWxayiPproduction-runtimew=1.00H8AYPsveeVp6TBMkbZ3RMH6Slibrary-patchnotesw=0.88GWUlFNsJxJ9aDobxPIfP87Qe

Each stream signs its observations with its own Ed25519 key. Cross-stream poisoning becomes asymmetric: production-runtime weighs 1.0, stackoverflow weighs 0.4.

01 — Statistical claims

0 CWE growth-rate projections at high confidence.

§65.1 ordinary-least-squares linear regression on monthly count series, with 95% prediction intervals and explicit sample-size disclosure. Every claim has n ≥ 24 months of NVD evidence. Every claim is reproducible from npm run failure:landscape against the same corpus.

No high-confidence claims yet — the substrate is refusing to over-claim. As the harvest accumulates more months of CVE evidence, projections graduate from 'insufficient-data' to 'high' at the §65.1 thresholds (24 months × 50 events).

02 — Cross-stream evidence

0 strength-3 jackpots, 2 strength-2 candidates.

A predicate observed independently across multiple structurally-independent evidence streams (public-repo harvest + CVE fix-commits + production-runtime) is the strongest possible signal that the predicate is universal — different communities, different contexts, all arriving at the same correctness property. The §47.6 jackpot is real when this list is non-empty.

Strength 2 · cross-stream candidates

2
null-check|<id>runtime + public
repos=61 · op-var=1[promoted to corpus]
2
equality|<id>.status_coderuntime + public
repos=15 · op-var=11[promoted to corpus]

03 — Substrate self-adapts

Evidence shapes the next product, by construction.

Until May 9 the substrate observed its empirical landscape and produced reports. It did not act on its observations at the gate level. ADR (aa) closed four specific wires that turn the substrate's analytics layer from reporting-mode into decision-shaping. Every promotion verdict now incorporates current empirical evidence; every product the factory ships injects an operator-specific corpus subset.

iWire 1

§47.6 strength → promotion bar

A signature observed across multiple structurally-independent evidence streams (public-repo + fix-commit + production-runtime) carries stronger evidence than three same-stream observations. The substrate now lowers the recurrence threshold by 1 per cross-stream class above 1 — strength-3 jackpots auto-promote at recurrence-1. Verdict reasons carry the [§47.6 strength-N] note for audit.

iiWire 2

§65.1 rising-CWE → relaxed threshold

When §65.1 forward-projection identifies a CWE class as 'high' confidence rising, predicates targeting that CWE clear a lower bar. Today CWE-79, CWE-862, CWE-22 and seven others are at high-confidence rising — predicates with those topics now promote at recurrence-2 instead of recurrence-3. The substrate adapts to where failure is concentrating right now.

iiiWire 3

§62 production-runtime → weighted recurrence

Production-runtime evidence is the strongest signal — actual deployed code hit this constraint. Each runtime repo counts as 2 public-harvest repos for the effective-recurrence calculation. A group with 1 public-repo + 1 production-runtime hits effective-3 and promotes — production traffic accelerates pattern-promotion of what your stack actually exercises.

ivWire 4 · the demo unlock

§66 fingerprint → operator-specific products

Two operators with different stacks now get different products from the same factory. The §66 fingerprint of each operator's repo determines which corpus subset gets injected as test stubs. A fintech operator gets SQL-injection / SSRF / path-traversal guards; a healthtech operator gets info-exposure / authz guards; both get xss-guard + universal predicates. Same corpus, same factory, deterministic-but-divergent output.

Verify Wire 4 yourself

# Two-operator side-by-side comparison
npm run factory:fingerprint-demo -- \
  --repo-a=path/to/fintech \
  --repo-b=path/to/healthtech

# Or with explicit CWE overrides (teaching mode)
npm run factory:fingerprint-demo -- \
  --repo-a=. --repo-b=../other \
  --demo-mode \
  --cwes-a=CWE-89,CWE-918,CWE-22,CWE-79 \
  --cwes-b=CWE-200,CWE-862,CWE-863,CWE-79

# Verified output: two operators, same corpus, different selections.
# May 9 demo run: 5 shared, 3 only-A, 3 only-B (out of 16 total).

Wires 1-3 fire continuously inside `cve:nightly` and the live §47 harvester; their effects are visible in the audit trail (`audit.jsonl`). Wire 4 fires at product-build time when an operator runs the factory against their fingerprint. Honest framing: Wire 4's selection differentiates only when the corpus contains CWE-tagged entries, which arrive as fix- commit signatures clear §47.3 promotion (Wires 1-3 accelerate that). The `--demo-mode` flag uses a synthetic CWE-rich corpus to demonstrate the mechanism in the meantime.

04 — Evidence streams

3 of 7 streams flowing tonight.

Six structurally independent dataset sources, each with its own audit-trail ADR. None require external operator participation — the substrate compounds from public evidence today, with the §62 reverse-loop primed for operator-runtime contributions when Phase 2 GTM lands.

§47

public-test-pattern-harvest

cycle 636; cross-cycle signature accumulation in signatures.jsonl

docs/adr (§47 series)

636
records
§54

cve-vulnerability-harvest

0 CVE records mapped; 0 GHSA-tagged

docs/adr/2026-05-08-dataset-harvest-cve-mvp.md

0
records
§54.1

cve-fix-commit-pattern-harvest

GitHub fix-commit patches harvested per CVE; routed into §47.1 signature memory via ADR (r)

docs/adr/2026-05-08-fix-commit-cve-harvest.md

0
records
§56.1

eu-enforcement-decisions

ICO English + CNIL French DPA enforcement decisions ingested as HTML

docs/adr/2026-05-08-eu-enforcement-decisions.md

2
records
§68

stackoverflow-near-miss-corpus

Accepted-answer pairs with expected/actual phrasing — developer near-miss evidence

docs/adr/2026-05-08-stackoverflow-adapter.md

0
records
§69 + §70

bug-bounty-and-postmortem-narratives

adapters shipped; awaiting first ingest

docs/adr/2026-05-08-bug-bounty-postmortem-adapters.md

0
records
§72

library-patchnotes-harvest

4 release entries across 8 libraries; 18 CWE-shape predicates extracted via deterministic token matching

docs/adr/2026-05-11-library-patchnotes.md

18
records

05 — Audit trail

Every claim, traced to a specific ADR.

50 architecture decision records cover every claim on this page. Latest: ADR (oo) on 2026-05-14. Each ADR carries falsifiable predictions — claimed today, verified tomorrow, falsified or confirmed in the open. The ADR paths are listed on each claim card above as canonical references; the documents themselves are NDA-gated during pre-Phase-2 stealth (see § 5).

50
ADRs
audit-trail integrity verified at build time · all claims linked · latest 2026-05-14
7
evidence streams
all carrying audit-trail references to specific ADRs

06 — How to verify

Audit IS the build. External verification, four surfaces.

The substrate is in pre-Phase-2 stealth — the source repository is currently private, opening as part of first-paying-pilot validation. Four verification surfaces are public today: this page's state.json for machine-pullable continuous attestation; the runnable audit bundle with its embedded plain-JS verifier (~500 lines, Node 18+ stdlib only); three live L12 runtime-AI receipt chains (paysafe / cliniclens / civicgate — every LLM decision a substrate-emitted product made, signed, prevHash-linked); and the audit trail (50+ ADRs available on request under NDA for serious operator-pilot or regulator-partnership conversations).

Snapshot honesty

The state.json above is a signed historical snapshot — that's the whole point of L7's transparency log. The substrate's architecture has advanced to v1.0 stable (May 12 2026) with three additional commitment chains (L9 self-evaluation, L10 audit-receipt, L11 substrate-component) + per-cycle autonomous-loop receipts + the reflexive bootstrap (substrate emits substrate components, two in runtime). The next state.json regeneration surfaces all of that; this snapshot is unchanged because re-signing a historical record would invalidate the L7 chain. See /system · attestation for the full L1–L11 stack.

Public surface — state.json

The state.json that produced this page is served at /state.json. Pull it programmatically. Diff against last week's snapshot to verify compounding. Cite specific claims with timestamps. Schema is locked at promethean-state-1.0 — future shape changes are explicit ADR-documented migrations.

# Pull current state
curl -s https://promethean.software/state.json | jq

# Verify Ed25519 attestation (substrate-side CLI)
curl -s https://promethean.software/state.json > state.json
npm run substrate:verify -- state.json
# → ✓ VALID · signature verified · trusted public key

# Verify schema + self-check
curl -s https://promethean.software/state.json | \
  jq '.selfCheck, .schemaVersion, .attestation.publicKey'

# Audit a specific claim
curl -s https://promethean.software/state.json | \
  jq '.statisticalClaims[] | select(.cwe == "CWE-79")'

The attestation block proves the document came from this substrate. Tamper with any field — corpus counts, audit-trail rows, statistical claims — and the signature breaks. Trust anchor: the Ed25519 public key published on /system · attestation. Compare embedded vs. trusted pubkey before accepting.

Source-level verification. For operator-pilot or regulator-partnership conversations, full source access (including the 50+ ADR documents that govern every claim on this page) is available under NDA. The reproduction recipe — clone, install, backfill, route, regenerate state.json — is identical to the process that produced the document you're reading. Reach out via the /join page.

Public surface — runnable audit bundle

The substrate now produces a self-executing audit artifact — a directory containing the signed state.json + every log file + every receipt + the audit conclusion + an embedded plain-JS verifier. Node 18+ stdlib only, ~500 lines, zero external dependencies, auditable in 30 minutes by a regulator's IT team.

A regulator's workflow:

# After receiving the bundle (one .tgz from Promethean or
# a partner mirror):
tar xzf substrate-audit-bundle.tgz
cd runnable-audit-bundle/

# Verify (exit 0 = PASS, exit 2 = FAIL)
node verify.mjs

# Or override the trusted pubkey for stricter checks:
node verify.mjs --trusted-pubkey="<your-archived-key>"

Verifications performed: (1) bundle hash recomputation over every embedded byte; (2) L1 state.json Ed25519 signature against the bundle's embedded trusted public key (or one overridden via flag); (3) the substrate's self-audit verdict at bundle creation. No npm install. No Promethean code. No TypeScript toolchain. The verifier is the regulator-facing artifact the substrate has produced precisely so trust doesn't require Promethean's continued operation.

Verifiable disappearance. The bundle works offline, after the original substrate directory is wiped (verified end-to-end in the substrate's test suite). Promethean can disappear; archived runnable bundles still produce green/red verdicts against the published Ed25519 trust anchor. See /regulators for the regulator-handoff package wrapper.

Public surface — L12 runtime-AI receipt chains (Phase R)

Three live signed L12 chains served as static JSONL — one per regulated vertical. Each chain records every LLM decision a substrate-emitted product made, bounded by a closed-enum spec, gated by the spec's reviewer-queue policy, Ed25519-signed per entry, prevHash-linked, anchored under the same L1 trust root that covers all other chains.

# Download any of the three chains
curl -O https://promethean.software/paysafe-runtime-ai-receipts.jsonl

# Verify with the substrate verifier (npm install needed once)
npm run substrate:verify-runtime-ai -- paysafe-runtime-ai-receipts.jsonl
# → 12 / 12 entries fully verified · head 14f4987a3cfbd76c…

# OR: extract the embedded verify.mjs from the runnable bundle —
# the same verifier walks L12 chains in 50 KB of plain JS, no deps.
node verify.mjs

Privacy by construction. Inputs are hashed before recording (hash-only redaction policy enforced by the spec). The chain proves an AI decision happened on input X without recording X. See paysafe, cliniclens, civicgate for each chain rendered against its spec with regulator-grade detail.