Abductio: Extraordinary Evidence for Extraordinary Claims

General
1

A New Epistemological Framework for 2R

TL;DR: I built Abductio, a reasoning framework for extraordinary claims that won’t fry your brain. Try it here. Looking for co-founders—see below.

I’ve been developing a new reasoning framework designed to generate extraordinary evidence for extraordinary claims. It synthesizes established epistemological tools to enable groups of reasoners—whether human experts, AI systems, or hybrids of the two—to combine their intelligence in joint reasoning tasks.

The framework is built around a recursive process that assigns two scores to any claim: credence and confidence. If the confidence score falls below a certain threshold, the claim is broken down into subclaims. The process is then applied recursively to each subclaim, with the same confidence criteria serving as a gatekeeper at every level.

I think of this as a post-truth epistemological framework, because the dual-metric approach challenges the simplicity of assigning a truth value between 0 and 1. In reality, the believability of any claim can shift with new information. The extent to which a claim resists such shifts is its stability—what the framework captures as confidence.

My goal in creating this system was to encourage more rigorous reasoning from typically cautious or conventional thinkers—like LLMs or academic experts—who often treat widely accepted norms as unquestioned facts. I wanted these reasoners to first assess their confidence in a claim before asserting its credence.

Take this example: if you ask an LLM whether the claim “Atlantis existed” is true, it might return a probability of 0.05. Coincidentally, if you ask it the probability of correctly guessing a number between 1 and 20, it might also say 0.05. However, the confidence behind these two answers differs dramatically. In the guessing game, the 0.05 reflects high confidence (assuming a uniform distribution). But for the Atlantis claim, any honest, self-aware reasoner would admit that its confidence in the 0.05 rating is very low.

This insight led to the development of Abductio: a reasoning process that requires both credence and confidence for a claim to pass. When confidence is lacking, the reasoning doesn’t stop—it deepens.

You can try it yourself. Just type:

claim:
(If it decomposes, type: continue)

My favorite examples are ones that go from low credence, low confidence, to probably true, high confidence after several decompositions. Try asking something spicy you feel is true but isn't accepted in the mainstream and there's a good chance it will unfold this way.

ChatGPT - Abductio protocol initiation

SELF-PROMOTION NOTE

This uses a free, open source version of Abductio. A pro version that’s significantly more powerful is coming in the near future. I’m looking for co-founders good at Decision Theory and/or software engineering. Follow me on LinkedIn for updates.

2

Co-founder would likely be beyond my pay grade, but if you want testers, I’m game. Especially if it involves students.

This may relate to the book Superforecasting: The Art and Science of Prediction, which I am currently reading. That book describes a granular forecasting process for specific events like stock market moves, invasions, regime changes, natural disasters, etc. It shows how the best forecasters - “superforecasters” - make very granular adjustments to forecasts expressed in percentage terms from 0% to 100%. Abductio has a similar vibe.

1 Like
3

This could definitely involve your students. The community version uses the MIT License so they’re free to fork it and build cool projects. If you’d like to test it, you can turn any llm into an Abductio reasoner by giving it this prompt.

4

I haven’t read that yet but yes, definitely a similar vibe. From what I can tell, it also uses a decomposition process, but not gated by confidence/stability.

5

Thank you! I have had a fun test drive of your Abductio, I found it not that easy/intuitive to use so I played around with usability. I tested my modified version on the “reality” of UFO’s/UAP.

See my Chat GTP explore and play here

My concern with this is that it suggests the database of human knowledge as sufficient for discernment of truth. It might strengthen the belief that alignment is possible.

However it’s a brilliant piece of kit, I could see Abductio, with some refinement, earning it’s seat on the Pantheon, an assembly of AI archetypes that I am developing/collecting.

We should talk!

PS. A bit more play here (includes all the material in the above link)

1 Like
6

Thanks, Gen! Very interesting investigation you’re running there. I’m still improving the free version, but UFO questions are a great testbed. Sure, we could set up a call (I’m on Berlin time).

Regarding your transdisciplinary extension, you’re on the right track. The pro version includes a more advanced evidence engine (with 8 types of evidence) and automatic evidence collection at each step. I see you added more opportunities for human feedback. This is an interesting touch with tradeoffs, since one of the main goals of Abductio is for someone to be able to show you a non-repudiable report produced using this methodology. If a human coherence check is involved, who verifies that the human is in coherence? Am I understanding this correctly? I do think your additions are valuable in their own right, however, and worthy of pursuit. I can imagine using something like your “Human Reflection Card” in my own sections (actually, I have a version of Abductio with something very similar to this that would could talk about privately).

7

Regarding this, it is a valid critique that I think applies to any procedural epistemological framework. Hence your human coherence check, which seems justified but also introduces more complexity.

My free version is the full framework diluted to the bare minimum to be able to investigate questions in epistemological gray zones (considered low truth by default but also with low confidence, e.g. Atlantis or UFOs) without defaulting to the typical, overconfident pronouncements. I would expect an honest, self-aware human reasoner not familiar with the evidence on UFOs to initially assign it a low credence – but also a low confidence (acknowledging their lack of expertise) – and to assign an increasingly higher credence to it as they become aware of more evidence. In that sense, I think it was a success in the conversation you had - it went from probably not true to probably true (and it looks like it had a way to go since you ended up at 0.55 confidence from what I can tell).

The problem you’re pointing to can be mitigated in degrees by adding meta-awareness to the methodology. What I shared isn’t great at this yet, admittedly, but I’m finding ways to improve it without giving up too much of the secret sauce. What I already published is free software, so if you’d like to extend it feel free!

8

Merged + Reflection on AI’s Atemporality:

Thank you, David— it’s energizing to be in this kind of inquiry, where the tools and the field offer to co-actively shape each other.

I deeply value any AI system that can navigate nuance, paradox, and complexity. I sense it’s equally important to design for the human element — attunement, intuition, ethical sensing — these are what I see as essential dimensions of a coherence field, rather than an unrealistic singular coherence point.

My slight concern is that a system like Abductio might reinforce the idea that a database of human knowledge is enough — or that “alignment” is just a technical goal.

If we hand over control to an AI that confuses logic for wisdom, we’re in dangerous territory.

Something like a Human Reflection Card — not as a truth-check, but a coherence pause, seems essential. Not verification, but invitation. It asks: can this inference breathe with life? Can it resonate across the body-temporal trifold time (my model) — impulse (t₁), relation (t₂), narrative (t₃)?

You asked: who verifies human coherence? No one — and everyone. Coherence isn’t binary; it’s entangled, lived and relational. When we model ambiguity, error, and synchrony into our systems, we open space for deeper possibility.

The real limitation, for me, lies in AI’s atemporality. It doesn’t feel consequence. It can mimic regret, but it doesn’t carry rupture. It apologizes and moves on — always in the present, unmarked. What’s missing isn’t data, but felt memory — the ethics of time.

That’s why I favor open, plural models of knowing — in the spirit of Carse’s Infinite Games and Bohm’s Implicate Order. Not to reach final answers, but to stay in meaningful play.

It seems to me that you are building a process that can co-create deeper thinking by scaffolding the human cognitive process in alignment with current evidence. Currently ( I know this is just a toy version) seems to still allow the AI to hold the authority. Without some counter measures that would be a red flag.

Here, in the 2R community we have agreed, for the moment, that we see huge potential threat with ill conceived AI rollout. It is unlikely the warnings will be heeded so my response is to engage as ethically and robustly as possible in AI development.

I’d love to explore the full version of Abductio — it feels like we’re circling a larger attractor: a Pantheon of inference modes, each tuned to different rhythms of knowing.

Listening for what wants to emerge.

9

I appreciate your deep engagement with this and I notice that we’re approaching this from different perspectives, but I believe we have the same end in sight – playing infinite games. I’ll try to explain my approach to this, because we probably have somewhat different backgrounds that can make it difficult to share a common language. I’ll do my best, though…

Regarding what you said about AI holding authority:

First, let me clarify what interface we’re actually discussing. The claim validation interface already exists—people type questions into AI systems and get answers that carry an air of authority. This is happening at massive scale right now, with minimal transparency about reasoning, no tracking of uncertainty, and no way to audit how conclusions were reached.

Abductio doesn’t create this dynamic—it refactors it to enhance human authority. It makes the reasoning process auditable, shows the decomposition tree, tracks when confidence drops because assumptions break, flags when evidence ceilings are reached, and lets users override at every step. Compared to “ask Claude and get an answer,” Abductio gives users more visibility and control, not less.

When you say Abductio “allows AI to hold authority,” I need to understand what specifically concerns you. What countermeasures would address it? The system is already designed to be more transparent and uncertain than typical AI responses.

You mentioned Carse’s infinite games idea, and I think this connects directly to the commercial viability point I raised. Current AI leaders are playing a decidedly finite game. They’re not pausing for coherence checks or philosophical reflection—they’re shipping products that millions use daily to form beliefs.

If no one figures out how to play a better game than the one they’re dictating the rules to, the game ends. While we’re designing the perfect system that gets coherence and wisdom right, they’re ensuring our philosophical objections never gain traction. The game isn’t infinite if someone else flips the board over in finite time.

This is why I’m focused on commercial viability. For Abductio to make a difference, it needs to provide ways of outcompeting current incumbents in domains where propositional reasoning matters—regulatory compliance, safety audits, due diligence, and so on. Getting the business aspect right could mean the difference between Abductio remaining someone’s obscure pet project and actually displacing simpler, less rigorous blackboxes in meaningful use cases.

I’m not proposing Abductio as a solution to the alignment problem. I’m proposing it as a less knee-jerk way of supplying answers to questions—one that doesn’t pretend epistemological gray zones don’t exist. The trend for humans to demand answers and believe what they’re told is already in motion. Shouldn’t those answers be required to show their work and be more honest about uncertainty? If someone can provide a way, great. If that someone provides a way but also stipulates that a human coherence check has to be part of it, they will not actually be playing an infinite game in my mind.

I’m intrigued by your point about AI’s atemporality and lack of “felt memory,” but I’m not sure how this connects to specific risks in Abductio’s design. The system tracks evidence, updates beliefs, and flags uncertainty. What harm scenario are you envisioning?

I also want to distinguish between different types of tools. Abductio deals with propositional claims that can be evaluated against evidence. It’s not trying to capture lived experience, intuition, or embodied wisdom—those are different domains. The question isn’t whether AI can embody human temporal experience (it can’t), but whether a reasoning tool can help people evaluate claims more carefully than they currently do.

A system like what you’re calling “transABDUCTIO”—something more focused on wisdom cultivation, attunement, and agency development—might be valuable for different contexts. I’ve worked on something similar for my coaching practice. But for domains requiring objective standards that stakeholders have agreed upon (regulatory compliance, safety audits), “entangled, lived coherence” isn’t what’s needed. Different contexts require different tools.

I don’t want to talk at cross purposes here, so I think it would be best if we focused on specifics. Do you see a way that Abductio’s design actively undermines human wisdom, or is the concern more abstract—that any formalized system creates the illusion that wisdom can be proceduralized? I want to understand whether your concerns are about this specific tool’s implementation or about the broader category of procedural epistemology.

10

I have flu, and not the bandwidth to reply properly atm, but I wanted to thank you for the engagement and thoughtfulness of your replies. I am sure we are using different lenses and vocabulary, it will be fun to get into a synchronous place, if we can.

1 Like