← Cognitive Mesh Architecture
Evidence · Cognitive Mesh Architecture

The fiftieth interaction is better than the first.

Not because the model improved. Because the reasoning behind every decision before it was kept, scored, and put back in front of the work.

Most systems store the answer. This one stores the argument.

A record here is not a document. It is a decision, and it holds the things you actually need a year later — when the person who made the call is you, and you cannot remember why.

Decision record
The call
Ship the deterministic scoring pipeline before the second deployment.
Rationale
The agent loop could not finish inside the platform's request window, so scoring failed silently on the days it was needed most.
Options weighed
Raise the retry ceiling · shrink the batch · move the work off the conversation entirely
Kill criteria
If the new scores disagree with the old scorer beyond the agreed band, the cutover stops and nothing is written.
Reversibility
Reversible — the previous scorer stays deployed and can resume unchanged.

Every field is enforced by schema, not by habit. A record that cannot say why is not written at all.

“Strategic” with a number attached

Every piece of work carries a Strategic Asset Score — seven weighted dimensions summing to one hundred. Four are universal and travel to any deployment unchanged. Three belong to the use case.

Decision rationale completeness20
Reference network strength20
Goal alignment15
Professional intelligence density15
Preference consistency10
Validation status10
Completeness & readiness10
Universal — portable to any deployment Specific to this use case

The heaviest weight in the deployment running today is whether the why was captured. The thing the system cares about most is the thing it was built for.

The work is getting better, and by how much

Not a sense that things are improving. Four consecutive batches of eight pieces of work, each one scored against the same rubric — and each batch better than the one before it.

4.7
6.5
7.1
7.2
first eight
next eight
next eight
most recent eight

A rise of +2.5, against a measured scoring-noise floor of 1.6 — the improvement is larger than the error in the measurement. Nothing declined at any batch size, and the trend is recomputed from the raw records every time it is checked.

+0.703
How much closer an agent's output tracks the recorded rationale when the reasoning is in reach — against a threshold of +0.15 fixed before the test was run
66
Scored pieces of work behind the trend, zero of them counted twice

Unattended, every morning

06:00 UTC
The scoring pipeline runs on its own schedule, with no one watching it
10–24s
To read, grade and file one asset — down from 150–200 seconds
0
Malformed records ever written to the store, across the entire migration
→Candidates are selected, neighbours fetched, results validated and stamped by code. The model does exactly one thing: the judgment itself.
→A judgment that fails validation produces zero rows — never a guessed one.
→Ratified constraints reach the work before it starts, carrying who ratified them and why.
→The improvement trend is recomputed from the source records by code on every check — never from a summary a model kept.

Your reasoning is an asset.

Cognitive Mesh Architecture keeps it, scores it, and hands it back at the moment it matters — in the deployment running today, and in the one after that.

Read the whitepaper →