Step 6 · Operations evidence · Operations evidence · K3 near-cap proof: 2 MiB and 99.76% PT-BREN
CLIProxyAPI 7.2.70 · kimi-k3 · Visual course

K3 near-cap proof: 2 MiB and 99.76%

Strong proof keeps both the first failure and the second success: the failure was the byte envelope, not the token policy.

Read first (primary source)
CONFIG-MODEL-POLICY-EVIDENCE.md

A redacted local record for the running gateway and the kimi-k3 example.

Read the simple version or open technical detail in any section.
1

Big idea

Strong proof keeps both the first failure and the second success: the failure was the byte envelope, not the token policy.

Think of it as… It is like passing a door with a backpack: making the backpack slightly smaller solves the door width without proving the corridor became shorter. The analogy breaks because serialized bytes and tokens are different units..

Under the hood

With slack 2500, both protocols received HTTP 400 above 2097152 bytes. With slack 4000, chat accepted 1046096 tokens and Responses 1046050: about 99.76%, with 3/3 needles, marker, and checksum.

2

In one picture

01 · policy target 1048576policy target 1048576governsfirst slack 2500Slack 2500 received HTTP 400 at 2 MiB.
policy target 1048576 — governs — Slack 2500 received HTTP 400 at 2 MiB.
3

In the code

This is the real local path and a safe read-only command.

kimi-k3-full-context/EVIDENCE.md 
# trecho real, não editado
CLIPROXYAPI_SKIP_OUTPUT=1 CLIPROXYAPI_MODEL_FILTER=kimi-k3 node validation-harness/validate-gateway-full-window.mjs

Open it yourself

CLIPROXYAPI_SKIP_OUTPUT=1 CLIPROXYAPI_MODEL_FILTER=kimi-k3 node validation-harness/validate-gateway-full-window.mjs

4

Try it

Use the controls to retrieve the evidence before reading the explanation.

02 · first slack 2500first slack 2500matches2 MiB byte envelopeSlack 4000 accepted about 99.76%.
first slack 2500 — matches — Slack 4000 accepted about 99.76%.
03 · 2 MiB byte envelope2 MiB byte enveloperoutes tochat 400 failureAll three needles and checksum returned.
2 MiB byte envelope — routes to — All three needles and checksum returned.
04 · chat 400 failurechat 400 failureexcludesResponses 400 failureSlack 2500 received HTTP 400 at 2 MiB.
chat 400 failure — excludes — Slack 2500 received HTTP 400 at 2 MiB.
05 · Responses 400 failureResponses 400 failureprovessecond slack 4000Slack 4000 accepted about 99.76%.
Responses 400 failure — proves — Slack 4000 accepted about 99.76%.
06 · second slack 4000second slack 4000boundschat 1046096 tokensAll three needles and checksum returned.
second slack 4000 — bounds — All three needles and checksum returned.
07 · chat 1046096 tokenschat 1046096 tokensmaps toResponses 1046050 tokensSlack 2500 received HTTP 400 at 2 MiB.
chat 1046096 tokens — maps to — Slack 2500 received HTTP 400 at 2 MiB.
08 · Responses 1046050 tokensResponses 1046050 tokensrestarts aschat stop statusSlack 4000 accepted about 99.76%.
Responses 1046050 tokens — restarts as — Slack 4000 accepted about 99.76%.
09 · chat stop statuschat stop statuspreservesResponses completed statusAll three needles and checksum returned.
chat stop status — preserves — All three needles and checksum returned.
10 · Responses completed statusResponses completed statuscompares withthree needlesSlack 2500 received HTTP 400 at 2 MiB.
Responses completed status — compares with — Slack 2500 received HTTP 400 at 2 MiB.
11 · three needlesthree needlesreturnschecksum 134Slack 4000 accepted about 99.76%.
three needles — returns — Slack 4000 accepted about 99.76%.
12 · checksum 134checksum 134confirmsoutput intentionally skippedAll three needles and checksum returned.
checksum 134 — confirms — All three needles and checksum returned.
13 · output intentionally skippedoutput intentionally skippedqualifiespolicy target 1048576 verification recordSlack 2500 received HTTP 400 at 2 MiB.
output intentionally skipped — qualifies — Slack 2500 received HTTP 400 at 2 MiB.
14 · policy target 1048576 verification recordpolicy target 1048576 verification recordrecordspolicy target 1048576Slack 4000 accepted about 99.76%.
policy target 1048576 verification record — records — Slack 4000 accepted about 99.76%.

Three-frame summary

1 Retain the 2500-slack failure
2 Fit under the 2 MiB envelope
3 Qualify the 99.76% success
With slack 2500, both protocols received HTTP 400 above 2097152 bytes. With slack 4000, chat accepted 1046096 tokens and Responses 1046050: about 99.76%, with 3/3 needles, marker, and checksum.
Record the observation, preserve the exception, and choose the next narrow proof.
retrieveSlack 2500 failureclick to flip
Slack 2500 received HTTP 400 at 2 MiB.
retrieve2 MiB envelopeclick to flip
Slack 4000 accepted about 99.76%.
retrieveSlack 4000 successclick to flip
All three needles and checksum returned.
Predict before revealing

What does the slack-4000 success establish?

It proves near-cap transport and semantic recall, not literal maximum input or long output.

Worked example
1Preserve both HTTP 400 errors from slack 2500.
2Run the separate slack-4000 attempt.
3Compare 1046096 and 1046050 with the policy target.
Retrieval review
What does the slack-4000 success establish?

It proves near-cap transport and semantic recall, not literal maximum input or long output.

Ask which boundary you would prove next; the next lesson narrows a different operational risk.