Cosolvent Design Roadmap

Cosolvent is designed to be help diverse and remote users to identify one another and, ultimately, to connect and do exchanges with one another. The exchanges can vary greatly to accomodate the wide variety of thin markets that need help. For example, the exchages may involve sales contracts, knowledge sharing, personal collaborations, as well as other constructive interactions. The conceptual model for all of these exchanges involves a series of common-sense steps that must be completed before meaningful exchange can take place. The model for this process is shown below:

The model assumes that the core of any exchange between unrelated parties will need to progress through at least 3 levels of trust building. No one will waste time, let alone money or effort, if they don’t trust the counterparty. Hence, the primary objective of any attempt to build a thin market exchange must focus on building that trust:

Trust to Explore – People are busy. They don’t have time to waste on frivolous exchanges or questionable counterparties. This is especially true in thin markets, where the participants will generally be unfamiliar with their opposite number. Under these conditions, participants may be unwilling to even entertain an advance or to reach out and make an approach of their own.
Trust to Engage – People may become interested in some of the other participants or their offerings. That is still not enough to spur someone to reach out and make contact. Every market has lurkers and looky-loos and it takes a targeted incentive to draw lurking users to reach out and make contact.
Trust to Accept – Assume that two participants discover each other and enter into a dialog. What then? When do they conclude that they know enough to be willing to sign a contract or remit funds? In familiar thick markets like Amazon, the decision may be spontaneous. In thin markets, however, participants will have more questions and need more concrete answers.

Deeperpoint believes that conventional software systems can only seldom achieve these levels of trust due to cost, practical, and technical reasons. However, what if AI technology were added to the design? What if they could be harnessed to make exploration easier, first contact more targeted, and quickly assemble all of the information needed to reassure the participants that their acceptance was justified? That is the design question and the design roadmap that DeeperPoint is trying to follow.

Discovery

The Discover Phase is where Cosolvent should help potential participants to join, participate and explore. DeeperPoint is betting that there are ways to use emerging AI technologies to make this task easier. In particular, Cosolvent wants to make it simpler for a participant to safely share detailed information about itself that could make a counter party less anxious. It then wants to use AI tech to ask and answer deeper and more context-specific search questions. We are hoping that, even if a participant is not an imminent buyer or seller, they will see the Cosolvent-based system as a useful tool for brainstorming and planning their thin market role.

The core idea is to use AI to do two things:

Help individual users to safely upload unstructured materials (text, pdfs, images, even videos) that tell potential counterparties who they are and what they do. With AI, we can look at miscellaneous documents and ask the LLM to craft a summary profile according to guidelines. There is a structured approval system where users can label specific assets as public, visible to registered users, available with approveal, or private. Additional criteria can fit specific thin market situations. The goal is to make easy for users to “fill out” their profile, while giving them control over which users see what material.
If a user wants to view other users’ published profiles, AI prompts can make sophisticated searches. For example, if a grain buyer wanted to know which farmers use John Deere equipment, an AI directed search might give them a set of prospects that match that oddball criterion.

Cosolvent can also integrate contextual information such as the questioner’s own profile and industry context information about a product or service. That would generate the kind of multi-factor answers that would be a big challenge to program with conventional code.

Communication

When participants think they have found an interesting prospect, we must convince them that it is safe and productive to make “first contact”. This requires a safe messaging system that will avoid things like spamming and marketing message blasts. We are hoping that we can tap into AI capabilities to filter and suggest constructive and fruitful first contacts. The messaging mechanism needs to be secure, but it probably doesn’t need to be sophisticated. Once the contact is made, subsequent communications will probably move outside the Cosolvent system to more familiar channels such as email, phone, WhatsApp, etc

Questions that the communication system might answer:

Simple Messaging System
Smart (AI) Message Filtering
Who do I want to contact?
Do I want to be contacted by?
What sort of contacts do I want?
When do I want to be contacted?
How do I tell people (politely) that I am not open to contact now?
How to I tell a specific set of contacts that they are encouraged to contact me?
How do I know if they read my message?
etc.

The main assumption is that “first contact” will involve a big element of psychology and human behavior. How do I ask? How do I say yes? How do I say no. Beyond that, different thin markets may evolve different etiquettes, attractions and fears. The communication system must be flexible enough that it can be configured to handle approaches in multiple ways. We are not ruling out the possibility that this will be complex enough that it will eventually benefit from AI support.

Negotiation

If parties make contact and want to proceed to an exchange, there will be a lot of details in most thin markets that must be “worked out”. With thin markets, there are fewer precedents and standard methods. If the items being exchanged are costly or important, everyone will have to arrive at a “meeting of the mind” in order to generate the impetus to execute all of the work and details. Fortunately, there is an extensive literature and history concerning how discussions should proceed. Offer, counter-offer, acceptance, questioning are all routine steps in any negotiation. Hopefully, that consistency will make it easier to build a support system for multiple thin market requirements.

AI to propose issues and wording
AI to compare offer against internal requirements (e.g., standard PO terms and conditions or standard sales contract)
AI to suggest (privately to one party) how to adjust offer to match their internal requirementsAI to bring comparables and industry practice to inform discussion
AI to capture and remember key discussion points and areas that need clarification.
AI to remind parties to make progress

Specification and Documentation

IF the parties can reach a point where they have agreement on all of the relevant issues, they will be ready to craft, sign and execute contracts. At that point Cosolvent can’t or shouldn’t go any farther. The best thing it can do is to summarize the process to date in a formal specification document (or set of documents) that summarizes the “meeting of the mind” to be passed to banks, lawyers, suppliers and the like. The actual execution of the deal must be managed in the real-world. As currently conceived, a Cosolvent system should not assume liability for real-world execution of the deal.

A Deal-making process can be modeled as a repetitive series of single point negotiations

One of the universal negotiating rules is the concept that “nothing is agreed until all is agreed”. That’s something that an AI enhanced software system should be able to assist and enforce.

At DeeperPoint, we are uncertain about this last step. It may be a bridge too far for the thin markets we are targeting. The final acceptance step is very close to the state that must be reached in traditional eCommerce. Depending on the thin market, it may be safer to stop short of this point. Alternatively, it may be doable as long as it is not built into the open source Cosolvent, but designed in the market-specific simulator and ecosystem.

Other Components?

These ideas are still tentative. Nonetheless, they are logical developments that DeeperPoint would like to pursue.

Industry Context – Real world deals typically operate in some real-world context that is separate from the specific participants, but can influence what is desired, allowed or possible. Examples are government laws and regulations, industry norms, technical constraints, trade agreements, geographical realities, time frames, etc. These considerations affect many potential deals and may be fairly stable and long lived considerations. Cosolvent views these “industry context” factors as something that should be identified and documented independently for each thin market. The contents may vary from thin market to thin market, but the AI should account for them in every query and response. DeeperPoint is building Cosolvent with some basic scaffolding to support industry context, but the actual information would be part of a subsequent simulator.
Mobile UI – Users should be able to upload assets to their folder from their phones. A farmer could walk his field or farm and take photos (or videos) of his/her crop or equipment or a handful of seeds. The farmer could take a photo of a government report with his/her phone. OCR is a possibility. The farmer should be able to save an attachment from an email and upload it to the document collection. etc.
Institutional Memory – This is probably an essential feature, but DeeperPoint is not yet sure what it should entail or how best to implement it. Our basic observation is likely impossible to create trust without a strong memory function. Cosolvent and/or its application must remember the outcomes of discoveries, communications and negotiations. Privacy and data protection will be big issues and it remains to be seen how such a system can be implemented in a system where many of the most important functions are performed by AI prompts?
Sample Management – Many thin markets will require the interchange of some sort of sample before the parties can seriously entertain a full contract. The sampling process can be challenging with physical goods: The sampling process for conceptual or knowledge goods may have issues protecting intellectual property. DeeperPoint is asking itself whether there might be a general model of this type of process (even a skeleton) that can be integrated with Cosolvent, or is it mainly a concern for the real-world application like GPSim?

Final Thought

The big question for Cosolvent development will be whether a given function should be implemented in the open source Cosolvent or coded into the market simulator for a specific application. Ideally, the open source skeleton should have functions that are truly generic and those should be implemented in a way that is agnostic about specific thin market characteristics. Figuring that out will be an ongoing challenge for Cosolvent or any initiative like it.