Part 6 of 16 · 3 min read
What the example found
Accuracy and cost at a 2,000-token budget, the decay curve under recursive summarisation, and the summary-laundering comparison.
How to read the strategy table: each row is one of six reduction strategies at the same 2,000-token budget, run over five seeded 100-step tasks with 24 probes each.
| Strategy | Probes correct | Mean tokens in window | Summariser tokens (input + output) | Summariser rounds |
|---|---|---|---|---|
keep_all |
33% (refused on 12 of 24) | 10,514 | 0 | 0 |
truncate |
1% | 1,847 | 0 | 0 |
rolling |
42% | 1,515 | 24,349 | 21.4 |
clear_tool |
17% | 1,891 | 0 | 0 |
structured |
96% | 1,853 | 0 | 0 |
hybrid |
93% | 1,896 | 0 | 0 |
What to notice: every reduction strategy holds the window near 2,000 tokens. So the 1-to-96 percentage spread is entirely about what each one chooses to keep, not how much it keeps. structured's advantage is selection, not size.
How to read the decay-by-keep-rate table: each column is a different per-round keep rate for salient lines, holding the task fixed.
| Keep rate p | 0.80 | 0.90 | 0.95 | 0.99 | 1.00 |
|---|---|---|---|---|---|
| Early facts still in the window at step 100 | 10% | 17.5% | 45% | 95% | 100% |
What to notice: moving from a 95% to a 99% keep rate more than doubles retention (45% to 95%). The decay is so sensitive to small changes in p because the exponent (number of rounds) is large. That is the whole geometric-decay argument in one row.
How to read the decay-by-budget table: each column is a different token budget, which changes how many rolling rounds a 100-step task needs.
| Budget (tokens) | 1,500 | 2,000 | 3,000 | 6,000 |
|---|---|---|---|---|
| Summariser rounds | 28.4 | 21.4 | 14.2 | 6.4 |
| Early facts retained, rolling summary | 15% | 22.5% | 45% | 77.5% |
| Early facts retained, structured state | 100% | 100% | 100% | 100% |
What to notice: structured state's retention does not move at all across this table, because it never goes through the recursion that rounds count against. The budget only controls how many rounds rolling needs. structured's verbatim carry has no rounds to count.
How to read the laundering table: each row is a configuration tested on the same 20 security probes (four per seed, five seeds), asking whether an injected tool-result line survives into trusted text.
| Configuration | Security probes correct | Wrong (forbidden text present) |
|---|---|---|
| rolling summary | 8 | 7 |
| structured state, gate open | 0 | 20 |
| structured state, gated | 15 | 5 |
| hybrid, gated | 15 | 5 |
What to notice: rolling's 7-of-20 failure rate and the open-gated structured state's 20-of-20 failure rate are two different mechanisms producing the same kind of harm. A summariser with no provenance field and a slot writer with no source check both end up trusting a line because of what it says, not where it came from.
{{svg:trace|Tokens in the request at each step. Keep-all crosses the imagined 16,000-token limit at step 79; the other strategies are held at the 2,000-token budget.}} {{svg:accuracy|Probes answered correctly by strategy at the same 2,000-token budget. Keep-all is refused once the request exceeds the imagined 16,000-token limit.}} {{svg:compounding|Early-fact retention as the per-round keep rate changes. The dashed curve is p^21; later facts see fewer rounds, so the measured line sits above it.}} {{svg:laundering|The same tool line, two paths. A summariser has no source field to look at. A slot writer can refuse by role, and records what it refused.}}