Measure whether the model actually loops, and shorten the README

[punnerud]

The gate the plan put in front of the loop detector. It changed two things.

The gate passes — but the threshold I proposed was wrong by 4×

Over 153 steps in 24 questions (simple, multi-step, malformed, ambiguous):

repeat — says something an earlier step already said	4.6 %
drift — new words, no new subject matter	3.9 %
neither	91.5 %

Drift does not dominate, which is the condition that would have stopped the work — drift is the failure a repeat detector cannot see by construction. A near-identical study elsewhere found the opposite (2 repeats in 41 steps, stateless drift at 49 %), so this had to be checked rather than inherited.

But the plan's 0.90 threshold flags 21 % of steps. The median step already sits at 0.846 cosine to some earlier step and the 75th percentile at 0.900 — so that bar catches the more-similar quarter of ordinary steps. Inspection settled it:

cos 0.9013
  step1: To find 17% of 250, we need to calculate a percentage of a given number.
  step2: We can use the formula: (percentage / 100) * number. In this case (17 / 100) * 250.

That is progress, not repetition. At 0.96 the flags are restatements — two of them character-identical to an earlier step:

cos 1.0000
  step1: The formula for the area of a circle is A = pi*r^2. Substituting r = 3 ... 28.27.
  step3: The formula for the area of a circle is A = pi*r^2. Substituting r = 3 ... 28.27.

The title bar needs no such care: exactly the same three steps fire anywhere from 0.95 to 0.999.

Per step it looks ignorable; per question it is not

Repeats cluster. A quarter of questions produce at least one, and one run produced four. That is the number the work is justified by, not the 4.6 %.

The script audits itself first

Elsewhere a first run reported a 58 % loop rate that was entirely the step extractor mistaking a code fence for a step and seeing it again — a measurement of the harness that nearly became a finding about the model. So the normalisation is checked for manufacturing duplicates before its numbers are used, and that check is pinned.

Also measured, because it decides whether a similarity threshold is a coherent idea at all: the mutual-kNN step graph does not percolate at k≤4 (largest component 5.2 %). Where that collapses, "the same" cannot be a threshold and has to be an exact canonical form instead.

Every raw step is committed and --reanalyse recomputes from them, so a threshold can be revisited without model time. Six new claims, written as guards and comparisons rather than point values — one battery run has no error bar, and pinning a rate to three decimals from it would be precision this has not earned.

Also: the README is a fifth of its former length

248 → 72 lines. The program that knows what is installed is better placed than a README to say what is missing, so startup now reports the models it found — or names the one command that fixes what it did not:

  Local Knowledge Graph

  Ollama did not answer at http://localhost:11434.

  This app runs a language model on your own machine through Ollama.
  1. Install it from https://ollama.com
  2. Pull a model:   ollama pull llama3.2:3b
  3. Start this again.

The detail moved to docs/: models and configuration, internal layers, design, development.