EKMAGEION · replay of a real sessionmachine 180001 · Phase 4 memory test · dev set · 26 Oct 2026

Real DeepSeek V4.1 Flash sessions. No numbers invented; all recomputed from the logged rows.

Replay · 26 Oct 2026

Same AI.
Alone, and with Ekmageion.

DeepSeek V4.1 Flash learns a real TLS server it cannot open, then has to remember how the server behaved before and after a software update. Same AI, same trace, same task.

The job

Learn a device from its trace, then remember it after it changes

Device: A TLS 1.0 server learned from real OpenSSL 1.1.0 (Janssen 2021 benchmark).
Update: At step it restarts into OpenSSL 0.9.8zb. The update is logged in the trace.
Handicap: 3,000 steps arrive in 14 chunks. About 1 reading in 20 is misread. Between chunks the AI keeps only its own notes, capped at 4,400 tokens.
Test: 52 questions: "after RESET, send these inputs; what is the true last reading?", before the update or now.

trace the AI seesstep · input · reading

During the trace

update logged

step 03,000

AI alone

replyDeepSeek V4.1 Flash

AI with Ekmageion

replyDeepSeek V4.1 Flash

Ekmageion watches the same trace and keeps a structural model of both software versions in . It stays silent until question time.

Question time

52 questions. Ekmageion offers its answer; the AI decides.

Outcome on this machine

Same AI, same trace. Alone vs with Ekmageion.

Loss is mean log-loss over the before and after questions (lower is better). Cost is the whole DeepSeek session. Machine 180001 is the clearest single case, not a typical one.

The control

Is it the knowledge, or just a second opinion?

A third run gave the same AI a decoy: Ekmageion's machinery holding a same-size library of the wrong structures.

All eight dev machines

One machine is a story. Eight is the evidence.

AI aloneAI with EkmageionEkmageion alone, no AIquestion loss, lower is better

How to read this

What this shows, and where it stops

Dev evidence, not sealed.The AI + Ekmageion pair passed on 8 dev machines. A sealed rerun on the 3 held-out machines would have had at most 51% power, so it was not run. The Ekmageion-only memory result is sealed: it beat adaptive L# on all 3 (p 0.039).

One run per setup.The AI's note-taking varies from run to run, so the eight-machine comparison carries the claim, not this replay.

The prompt differs by one paragraph.The Ekmageion run is told it will see Ekmageion's answer at question time. Everything else is identical.

Scope.TLS servers only. The change point is logged; spotting an unannounced change is a stated limit. DeepSeek V4.1 Flash throughout.

Ekmageion

The AI keeps its judgment.
Ekmageion keeps the structure.

Source: research/ekv5/phase4/known_memory/dev_b/rows/agent_1800**_{alone,mem,mem_twin}_flash.json
and dev_a/rows/km_1800**.json, merged in PR #34. No new runs were made for this video.

Same AI.Alone, and with Ekmageion.