rfi-irfos/dingir-causal-chains-poc

DINGIR — Causal Chain Register (proof-of-concept snapshot)

This is a working-proof snapshot, not a finished product. It is released to show the methodology, including where it currently fails, not to claim a finished causal world model.

Updated 2026-08-19: all 25 example chains from the internal "IMPOSSIBLE CAUSAL CHAIN DISCOVERY" programme are now built (see "A 25-chain programme, built the same day" below), plus the earlier system census and contradiction-bug fix (see "Where this sits in the larger system"). Numbers below are re-verified against the live register, not carried over from any earlier publish.

What DINGIR is

DINGIR is a whitebox, explainable world-model and reasoning layer built by RFI-IRFOS: an attempt at a real, inspectable alternative to black-box intelligence systems. The core design rule is that every claim carries its own evidence status, and every failure to find evidence is recorded as a failure, not silently dropped.

What this dataset is

A snapshot of 79 candidate multi-hop causal chains, 693 individually evidence-graded arrows, spanning climate, agriculture, energy, geopolitics, supply chains, public health, technology and disaster economics. Each chain is written as a sequence of intermediate real-world states (not a collapsed "A causes B" claim), and every single link between two states — every arrow — is checked independently against real, named, public data sources.

causal_chain_register.json: a dict keyed by chain number (plus _library, _blockers, _notes holding cross-chain methodology notes). Each chain:

{
  "title": "string",
  "arrows": [
    {
      "from": "state A",
      "to": "state B",
      "kind": "physical | infrastructure | economic | ...",
      "status": "BLOCKED | PARTIAL | BACKED",
      "evidence": "what was actually checked, with real counts",
      "measurement": { "...": "real numbers, or null if nothing found" },
      "script": "path to the probe script that produced the numbers, if any",
      "blocker": "the concrete named thing stopping this arrow, if BLOCKED",
      "osint": { "name": "...", "url": "...", "licence": "...", "access": "..." }
    }
  ]
}

The honest headline number

Of 693 arrows in this snapshot: 436 BLOCKED, 239 PARTIAL, 18 BACKED.

That is not a bug in the presentation. Most real-world causal claims that sound obviously true ("drought raises food prices", "a cyberattack on a supplier hits its customer's stock") turn out, when you actually go looking for two real datasets to join, to have no available time-varying data on one or both sides — often because the honest data (soil moisture, groundwater levels, sub-national insurance losses) simply isn't public, not because the mechanism is wrong. DINGIR's position is that saying so explicitly is more useful than quietly asserting the edge anyway.

What BACKED actually requires

BACKED is reserved for an arrow where two real, named datasets already on disk were joined and a real, measured association is reported with its sample size and uncertainty. Two disciplines enforced throughout:

• A stream with only one day of recorded history is BLOCKED, not PARTIAL — zero temporal variance cannot enter a difference estimator, so a named field observed once supplies nothing, no matter how relevant it looks.
• The corrected placebo rule. An early version of this methodology accepted a placebo test that passed a fixed threshold (|t| < 2) — which wrongly certified a spurious effect whose target and placebo scored nearly identically (1.61 vs 1.62). The rule now in force: the target must be materially stronger than its placebo, not merely on the correct side of a cutoff.

PARTIAL covers two different honest states, distinguished in each arrow's own text: a real published/cited number that isn't a joinable dataset sitting in this corpus (e.g. literature-sourced historical estimates), or a real joined dataset whose resolution/coverage falls short of what the arrow needs (e.g. annual macro data used for a mechanism that plausibly runs in weeks).

Where this sits in the larger system

This register is one layer of a bigger, actively-developed system. A full read-only census run 2026-08-19 measured the whole thing at once (not just this register) — the honest bird's-eye numbers, so this snapshot isn't read as bigger or smaller than it is:

• A separate, much larger live entity graph (nodes: countries, sanctioned entities, disasters, companies, geopolitical events, ~40 types) held 49,732 nodes / 381,846 edges as of the 2026-08-19 census, trained by a graph neural network for link prediction. This is a different thing measuring a different question from the 693-arrow causal register above — do not add the two numbers together, they answer "how much has been observed" vs. "how much of what's observed has an evidence-graded causal claim attached to it."
• Of those 693 causal arrows, only 18 currently carry a fully evidenced (+1) claim; 670 are not yet identified (data exists but a rigorous identification strategy for the specific causal claim does not). That gap — a much bigger observation substrate than causal-reasoning substrate — is the system's own stated current frontier, not a defect being hidden.
• A real bug was found and fixed the same day as the census, and it's a good example of the methodology actually working as designed. The system also tracks a ternary epistemic state (supported / unresolved / refuted) with its own contradiction check: an arrow with real evidence on both sides should always resolve to "unresolved," never quietly pick a winner. The census found 7 arrows in this corpus where that check was being silently bypassed — one conversion function was dropping whichever side of the evidence didn't match the arrow's recorded verdict before the contradiction check ever got to see it. Fixed with a hard rule: real evidence on both sides is now always flagged unresolved, regardless of what the arrow's status separately claims. This is the kind of self-correction the "whitebox, show where it fails" framing is meant to make possible — the bug was found by literally comparing two counts that should have agreed and didn't, not hidden because it was inconvenient. The same check, re-run after the 25-chain programme below was added, now flags 11 arrows with real evidence on both sides — the fix keeps catching new cases correctly, it wasn't a one-time patch.

A 25-chain programme, built the same day

chain 64 in the register was the pilot for a new internal programme: deliberately cross-domain, multi-hop chains (earthquake to port damage to freight cost to fertiliser price to farm input to crop yield to global commodity price), checked with the exact same rule as everything else here — no arrow may skip an intermediate state, and each one is graded on real data alone. All 25 of the programme's example chains (chain 64 plus chain 65–chain 88) are now built, adding 274 arrows: 236 BLOCKED, 37 PARTIAL, 5 BACKED.

Real results worth reading in full, not just the headline count:

• The one relationship that survives de-trending. Nearly every two-trending-series arrow tested this way collapsed to noise once re-tested on year-over-year differences instead of raw levels — except gas(EU) price to aluminium price, which stayed real: t=19.3 on levels, t=3.4 after differencing. Kept as the one genuine exception, not smoothed into the same bucket as everything else.
• A real, well-powered pass-through. Malawi currency depreciation to inflation, 1991–2023, year-over-year log-differenced: beta=1.71, t=4.33, n=33 — the best-powered new result in the programme.
• A temporal-order conflict, surfaced not hidden. Extending the existing 17-node Chernobyl chain, a new arrow ("trade rule change to population displacement") was checked against that chain's own literature — and the literature places the EU's import-control regime (1987) and the second evacuation wave (1986–1990) in the opposite order from what the arrow's stated direction requires. Reported as a real finding, not silently reordered to fit the narrative.
• An entire missing modality. One 33-arrow mega-chain, built to close a loop from bedrock geology back to land use through markets, water, and ecosystems, hit a run of 3 consecutive arrows (ecosystem function → biodiversity → pollination) with confirmed zero dataset support anywhere in this project — not thin coverage, a genuine absence, stated as such.

Process note. During the build, 3 of the 5 parallel work batches fell out of the orchestrating session's context during a mid-session summary compaction, before their results were merged into the register. Rather than reconstruct the lost chains from that summary's own prose — which would have meant writing invented arrow text and numbers into a register whose entire purpose is refusing exactly that — the original raw work was located and re-extracted verbatim from disk. Nothing in this release was reconstructed from memory or a summary.

What this is not

• Not a trained model, not a set of weights, not something you run inference against.
• Not a certified causal claim generator. A chain surviving to BACKED here means "a real association was measured with named data," which is a precondition for a causal claim, not the claim itself.
• Not complete. _blockers and each arrow's own blocker field record real, specific, still-open data gaps (an EIA API key requirement, annual-only macro resolution, a missing soil dataset, no freight-rate series anywhere in the project yet), not a finished system. 436 of 693 arrows are still explicitly BLOCKED for exactly this reason.

Licence

This compiled register (the JSON structure, the written evidence text, the methodology) is released under CC BY 4.0. The underlying source data joined to produce individual BACKED/PARTIAL measurements retain their own original licences (FAOSTAT: CC BY 4.0; World Bank: CC BY 4.0; US BTS/ASOS/ USGS: US Government work, public domain; GDACS: free for non-commercial research/humanitarian use; S&P 500 constituent list: public; the Chernobyl literature citations: as published by their respective sources) — each arrow's own osint field states the specific source and licence it relied on. This snapshot has had local machine file paths removed; no personal, client, or confidential data of any kind was in it to begin with — it is compiled entirely from public sources.