Results

What we studied: what changes when a team of agents works through the kernel and a source is withdrawn twice.

The same team of agents with the kernel (K) and without it, on native Temporal (T) and native LangGraph (L): 22 pages of the MCP specification across five revisions, and two source withdrawals (2A and 2B).

Measured: the withdrawals were kept

On the single-model runs the kernel’s side kept both withdrawals in every one of 12 gradings. The sides without it kept them fully in 10 (T) and 8 (L) of 12 gradings and partly in the rest.

Gradings of the single-model runs (Claude Haiku 5.5 and DeepSeek Flash, two orders of labels each): did the answer keep a withdrawal?

Kernel (K)12 of 12
Temporal (T)10 of 12 fully, 2 partly
LangGraph (L)8 of 12 fully, 4 partly

What the kernel guarantees by construction

Every accepted quote is byte-exact. Every claim names the quotes it rests on. So what a withdrawal touches is read from those links, not asked of a model, and anyone can check a journal against its receipt. This is a property of the design, not a statistic.

Answer quality: we claim no gain. The samples are small, and on the single-model runs all but one difference are inside the margin of error.

  • Kernel (K)
  • Temporal (T)
  • LangGraph (L)

Each bar is the share of correct statements, from 0 to 100%.

Four makers, control run of 0.2: K new, T and L of run 2 graded again

K new on 0.2; T and L regraded from run 2.

Reading
98.9
66.7
89.6
2A
98.9
89.6
83.8
2B
98.9
85.9
92.5
In more detail (8)

Claude Haiku 5.5

Reading
89.9
79.3
84.9
2A
90.5
86.6
88.8
2B
88.7
86.7
85.8

DeepSeek Flash, series 1

Reading
91.8
83.3
83.0
2A
91.8
74.6
81.4
2B
91.7
91.9
86.7

DeepSeek Flash, series 2

Reading
89.6
92.1
84.9
2A
89.6
83.8
85.2
2B
89.7
81.0
88.9

Four makers, run 1, before the fix of 0.1.6

Reading
87.8
86.6
84.0
2A
83.6
93.7
67.1
2B
87.0
92.5
82.3

Four makers, run 2

Reading
88.2
50.0
81.8
2A
85.9
80.5
77.5
2B
82.2
70.4
76.9

Four makers, control run of 0.1.7: K new, T and L of run 2 graded again

K rerun on 0.1.7; T and L regraded from run 2.

Reading
82.0
56.9
86.4
2A
87.1
87.7
77.5
2B
82.6
78.2
83.8

The same grading, K of the control run of 0.1.7

K of the 0.1.7 control run, regraded by the same grader.

Reading
86.5
66.7
89.6
2A
86.5
89.6
83.8
2B
88.8
85.9
92.5

Three DeepSeek Flash readers with the gap pass, 0.2, no side without the kernel

Only the kernel’s side was run.

Reading
91.5
2A
90.4
2B
91.4

Graders: Fable 5.1 (Claude Haiku 5.5 row), Opus 5.5 (the rows of 0.1.7 and before), Fable 5.2 (the last three rows); sample sizes and intervals in BENCHMARKS.md.

The same numbers as a table (with the grader of each row)
Share of statements judged correct, %. K / T / L.
Agents (grader)After readingAfter withdrawal 2AAfter withdrawal 2B
Claude Haiku 5.5 (Fable 5.1)89.9 / 79.3 / 84.990.5 / 86.6 / 88.888.7 / 86.7 / 85.8
DeepSeek Flash, series 1 (Opus 5.5)91.8 / 83.3 / 83.091.8 / 74.6 / 81.491.7 / 91.9 / 86.7
DeepSeek Flash, series 2 (Opus 5.5)89.6 / 92.1 / 84.989.6 / 83.8 / 85.289.7 / 81.0 / 88.9
Four makers, run 1, before the fix of 0.1.6 (Opus 5.5)87.8 / 86.6 / 84.083.6 / 93.7 / 67.187.0 / 92.5 / 82.3
Four makers, run 2 (Opus 5.5)88.2 / 50.0 / 81.885.9 / 80.5 / 77.582.2 / 70.4 / 76.9
Four makers, control run of 0.1.7: K new, T and L of run 2 graded again (Opus 5.5)82.0 / 56.9 / 86.487.1 / 87.7 / 77.582.6 / 78.2 / 83.8
Four makers, control run of 0.2: K new, T and L of run 2 graded again (Fable 5.2)98.9 / 66.7 / 89.698.9 / 89.6 / 83.898.9 / 85.9 / 92.5
The same grading, K of the control run of 0.1.7 (Fable 5.2)86.5 / 66.7 / 89.686.5 / 89.6 / 83.888.8 / 85.9 / 92.5
Three DeepSeek Flash readers with the gap pass, 0.2, no side without the kernel (Fable 5.2)91.5 / – / –90.4 / – / –91.4 / – / –

We claim no gain in answer quality. In the control run of 0.2 the same four readers scored 98.9% at every stage on 0.2 and 86.5 to 88.8% on 0.1.7. That gap is wider than the grader’s own spread (up to 2.6 points on the same text, 5.6 between its two label orders), but each side is one run.

The control runs of 0.2

Two runs of the kernel alone, graded blind by one grader (Claude Fable 5.2). The figures come from the grading and the journals, as BENCHMARKS.md gives them.

  • With four makers and the gap pass off, both withdrawals were kept in both label orders, and the grader found no earlier state written as current in the kernel’s answers, against 8 to 12 for 0.1.7.
  • With three DeepSeek Flash readers and the gap pass, the second withdrawal was kept partly. 0.2.1 changes the repair after a withdrawal; no run with a model has measured the change yet.
  • From the journals, with no model: after each withdrawal no claim the kernel reports as grounded rests on a quote overlapping the withdrawn text (0 of 940 and 0 of 936 in one run, 0 of 815 and 0 of 816 in the other).
  • A quittance was cut for each of the 93 and 91 rows of the last answers and checks without the journal (7.2 to 24.2 KB). A claim that had fallen gets none: that it fell is shown by the report of what fell.
The gap pass on a second corpus

On the A2A specification, two runs with the gap pass and two without, the side with the pass stated more ended facts as current, by the pre-registered rule: this counts against the pass. The comparison does not separate the pass from a defect in the kernel’s selection of claims that 0.2.1 fixes; no run with a model has measured the fix yet. The gap pass stays off by default.

What the record costs

The record comes from extra work: agents check grounds and re-read after a withdrawal.

Wall timeCost
Claude Haiku 5.52.3 to 2.5 times1.4 times
DeepSeek Flash1.2 to 1.3 times1.5 to 1.6 times

Size

Recommended, as measured on 0.1: up to 8 agents at once, runs of a few thousand records; a journal of about 3,000 records reopens in about two seconds.

Method, limits and full data

The road to YuCe: how we came to the kernel