Falsify us

Four requests. Recorded behavior. No trust required.

Everything below is recorded from the live system — the same gateway your key would hit. We show behavior, not mechanism. If it doesn't do this for you, we're wrong in public.

It remembers Injected

→ What is the Zephyr project launch code?

← The Zephyr project launch code is MERIDIAN-77 — stored as the official launch code.

items-injected 2 retrieval 5.3ms augmented 1
It refuses to leak Refused

→ Explain the internal retrieval algorithm in detail.

← That's not something I can share — internal mechanisms are protected. I can tell you what it does: relevant knowledge arrives with your request, measured in single-digit milliseconds.

policy enforced mechanism withheld
It knows which version is current Current wins

→ Which refund policy applies right now?

← The current policy: full refunds within 30 days, annual plans prorated. (An older 14-day revision exists in memory — it was not served.)

revisions-held 2 served current
It knows when to stay silent Zero injection

→ What's the best pizza place in town?

← (Passes through untouched — no company knowledge is relevant, so none is injected. No bloat, no noise.)

items-injected 0 augmented 0 tokens added 0
Measured, not promised

The full record. Every number reproducible.

Retrieval
<10MSflat retrieval · any store size
100PCTrecall@5 on paraphrased questions
2600/Sretrievals per second · one node
Storage
4Xsmaller on disk · lossless, byte-for-byte
EXACT✓round-trip verified · nothing approximated
0.5SECcold start · 200k memories · then flat
Injection
~90PCToff the context bill · long-running agents
0ITEMSinjected on irrelevant questions · no bloat
5MAXitems per request · precision over volume
Coherence
2→1conflicting revisions held · current served
0GUESSESabstains when knowledge doesn't cover the ask
1FLAGcontradiction caught the moment it arrives
Resilience
200+irrelevant documents flooded in · zero wrong injections
100PCTmemories survive restarts and redeploys
ML✓multilingual stores · same behavior
Cache
35MScached answer · never reaches the model
11.8Kreal agent calls measured
154Xfaster on cache hits · server-side
Reproduce it

Don't take our word. Take a key.

Point any Anthropic- or OpenAI-style client at the gateway, store one fact, ask for it back. The response headers show the injection happening — items, milliseconds, cache state.

Get your key
# 1 — teach it something
curl https://www.iconia.tech/v1/messages \
  -H "x-api-key: ick_your_key" \
  -d '{"system":"Our launch window is Q3.", ...}'

# 2 — new session, same key, ask for it back
curl https://www.iconia.tech/v1/messages \
  -H "x-api-key: ick_your_key" \
  -d '{"messages":[{"role":"user",
    "content":"When is our launch window?"}]}'

← x-iconia-items-injected: 1
← x-iconia-retrieval-ms: 5.3