The H-A-H Protocol: Human-AI-Human Output Compression

1. Thesis

Most “AI in the loop” systems put the AI on the input side — nudging, coaching, or drafting what a human is about to say, in real time, inside an ongoing conversation.

Vaal’s H-A-H protocol puts the AI on the output side — interviewing a human after they’ve given a raw, messy answer, and compressing that answer into a structured, reusable artifact for a specific person who wasn’t part of the original exchange.

That’s a narrow, specific difference. This paper lays out why it matters, where it sits relative to existing research, and what’s genuinely still unproven about it.

2. The problem

Practical knowledge about navigating real-world friction — which counter to use at a government office, what a hospital bill line-item actually means, how to appeal a pension rejection — lives almost entirely in people who’ve already been through it. Two things are true about that knowledge:

• It’s valuable and current in a way generic AI answers aren’t. An AI can tell you the official process; it can’t tell you which clerk actually enforces it, or what the form rejects you for that nobody documents.
• It’s locked in a format nobody else can use — a stranger’s memory, a rambling forum reply, a WhatsApp voice note. Unstructured, one-off, hard to search, hard to trust.

The bet: an AI interview layer between the person who has the answer and the person who needs it can convert the first into something usable by the second — without requiring the helper to be a good writer, and without the asker having to sift through noise.

3. Where this sits in existing research

The closest academic umbrella term is AI-Mediated Communication (AI-MC), defined in a 2020 paper as interpersonal communication in which an intelligent agent operates on behalf of a communicator by modifying, augmenting, or generating messages to accomplish communication goals. That’s the right category. But most published and commercial work in this category is shaped differently from what Vaal does:

The pattern in the literature is real-time, input-side, relationship-based. Vaal’s pattern is async, output-side, episodic. Nobody indexed appears to be testing the second shape specifically for practical/bureaucratic problem-solving between strangers who never continue the relationship.

This doesn’t mean no one else is doing it — stealth and unpublished work wouldn’t show up in this kind of search, so “no one else has built this” can’t be stated with certainty. What can be said: the specific mechanic doesn’t match anything found in current academic literature or funded products.

4. A caution worth taking seriously

Not every result in this space favors “add AI.” A 2026 experimental study on peer support groups for exam anxiety found that human-facilitated sessions outperformed AI-facilitated sessions across every measured outcome. AI-as-facilitator is not automatically better than a human doing the same job.

This is actually supportive of the H-A-H framing rather than a threat to it: Vaal’s AI never facilitates or replaces the human helper’s judgment. It only compresses what the human already said. The failure mode the 2026 study captured — AI substituting for human presence — isn’t the bet Vaal is making.

5. Two separate hypotheses, one MVP

H-A-H bundles two genuinely different unknowns, and it matters to keep them apart when reading results:

1. Compression quality — does AI-guided extraction produce a more useful answer than the helper’s raw, unstructured response would have been? This is a product/UX question, answerable once real threads complete.
2. Supply — will a stranger give real time to help another stranger, for an uncertain, small reward? This is the harder, older problem every peer-help platform has faced, and it’s mostly independent of whether the AI compression is any good.

A null result on one doesn’t tell you anything about the other. A thread with zero helper responses says “supply problem,” not “compression problem” — the extraction step never got tested at all.

6. What’s actually known vs. assumed, right now

Known / grounded in research:

• The general category (AI-MC) is real, named, and has a five-year-old research agenda behind it.
• Output-side, episodic, post-hoc compression is a distinct enough shape that it isn’t a re-tread of HAILEY, the Intersubjective Model, or Wisdo Health.
• Human facilitation can beat AI facilitation in some peer-support contexts — reinforcing that Vaal’s AI should stay in a compression role, not a facilitation or replacement role.

Not yet known — open questions this MVP can actually answer:

• Does the Agent Q&A output measurably outperform the helper’s raw unstructured answer, by the asker’s own judgment?
• Does a visible, upfront reward change stranger conversion into the helper flow, and by how much?
• Does an episodic, one-time H-A-H exchange retain any of the trust/quality benefits seen in relationship-based peer support, or does the lack of an ongoing relationship change the dynamic entirely?

None of these have data behind them yet. This paper is a positioning thesis, not a results paper — the honest next version of this document is the one written after the first real threads complete.

7. Publishing plan

Suggested sequence, each one a standalone piece:

1. This paper — the category-naming, prior-art positioning piece. Publish first to stake out the “H-A-H” term before writing about results.
2. Compression quality results — once you have completed threads: raw helper answer vs. Agent Q&A output, judged by the asker.
3. Supply results — conversion funnel data (link opened → started → completed) across however many threads you push out, with honest numbers even if they’re small or zero.
4. Reward design findings — what changed when the reward became visible/UPI-linked vs. vague.

Each piece stays grounded in what actually happened — no projected numbers, no assumed outcomes ahead of the data.

8. References

1. Hancock, J.T., Naaman, M., & Levy, K. (2020). AI-Mediated Communication: Definition, Research Agenda, and Ethical Considerations. Journal of Computer-Mediated Communication, 25(1), 89–100. https://academic.oup.com/jcmc/article/25/1/89/5714020
2. Sharma, A. et al. (2023). Human–AI collaboration enables more empathic conversations in text-based peer-to-peer mental health support. Nature Machine Intelligence. https://www.nature.com/articles/s42256-022-00593-2
3. Intersubjective Model of AI-mediated Communication: Augmenting Human-Human Text Chat through LLM-based Adaptive Agent Pair (2025). arXiv:2502.18201. https://arxiv.org/pdf/2502.18201
4. AI-mediated social support: the prospect of human–AI collaboration (2025). Journal of Computer-Mediated Communication. https://academic.oup.com/jcmc/article/30/4/zmaf013/8200809
5. Human vs. AI Facilitation in Peer Support Groups: an Experimental Study on Exam Anxiety (2026). Journal of Technology in Behavioral Science. https://link.springer.com/article/10.1007/s41347-026-00617-3
6. Publicis Health & Talkspace / Wisdo Health partnership announcement (2026). https://www.prnewswire.com/news-releases/publicis-health-and-talkspace-partner-to-improve-treatment-adherence-by-addressing-gaps-in-social-health-and-peer-support-302720177.html

Status: Working draft. To be revised once the first real (non-simulated) helper thread completes.