DecisionGraph

Ask what your team decided. Get the decision that still stands.

Teams write a decision down, then change their minds and write another. Search finds the old one as easily as the new one, so agents and new joiners act on decisions that were replaced years ago. DecisionGraph links every decision to the ones that replaced it, walks those links to the current answer, and learns from every correction.

Answers come from 128 real architecture decision records published by the UK Ministry of Justice Cloud Platform, its Modernisation Platform, and GOV.UK. This live graph has already learned from the 50 corrections in the test below. The example questions are worded differently from that test. Ask something new, correct it, then ask again in other words.

Does it learn?

We ask 50 questions in five rounds of ten. Every system answers each question before it is told the right answer, then gets the correction, so every score is on a question it had not been told about. Mean of five question orders. On 24 held-out questions written by an independent AI reviewer, the trained graph is right 96% of the time, against 42% for vector search.

Runs the whole test, question by question, on a fresh copy of the graph in a second HydraDB cell. The live graph above is not touched. Takes a minute or two.

Inside the graph

Only a few decision records say what replaced them, so four simple rules propose the missing links. Corrections confirm or reject individual links, and adjust how far each rule is trusted. So far, only links with an explicit status line are trusted enough to follow without a correction; the bars show where each rule stands.

What the corrections changed

Each row is a “replaced by” link that a correction changed, read live from HydraDB. The thin line is what the rules guessed before any correction. The thick line is what the graph does now. A link that is not followed is never walked, however close the text.

    The evidence, in three questions

    Each answer is measured on the same 50-question test, before the system is told the right answer. Brackets are 95% intervals from resampling the questions.

    How it works

    1. Find the decision

      Vector similarity, terms that link to decisions in the graph, and hints learned from past corrections each vote for where the question starts.

    2. Walk to what replaced it

      One algo.MSpaths call in HydraDB follows trusted replaced by links, with trust stored as path cost, until it reaches a decision nothing has replaced. The agent gets that decision and its history, about 130 tokens instead of 590. If nothing is close enough, it says there is no decision on record.

    3. Learn from the correction

      Each correction is stored as a question in the graph. It confirms or rejects the links it walked, proposes a missing link, adjusts rule trust, and points the question's words at the topic. When the topic comes back in new words, the graph is right 84% of the time, against 47% for vector search with memory.