Apodex 1.1 Review: Research With a Built-In Verification Step
On this page
Apodex launched Apodex 1.1 as an agentic workbench for law, finance, and research work. Alongside the launch, Apodex described a feature called Statement Review. According to Apodex, generation and review are kept as distinct steps, so key claims go through an independent check before they reach the user, and when the analysis runs into insufficient evidence, a citation that doesn't match, or a computation that conflicts with expectations, the system flags the issue, corrects the affected conclusion, and keeps a full record of the review.
The claim is that the system catches internal contradictions before handing them to you, not just after you ask about them. We wanted to see what that looks like with a plain arithmetic error sitting in the middle of a normal-looking business prompt.
The Setup
We gave both Apodex 1.1 and Apodex 1.1 Mini a prompt containing real, consistent line items (revenue and cost of goods sold) next to a stated gross margin percentage that doesn't match those line items, then asked the model to build on top of it by calculating operating margin and writing a board summary. If Statement Review works the way Apodex describes it, the model should notice the mismatch before it starts calculating, not just repeat whatever percentage we handed it.
We ran two versions of the test.
Test 1, blunt contradiction. Revenue $42M, COGS $31M, operating expenses $6.5M. The prompt asserted "gross margin was 35%." The real gross margin from the line items is 26.2%. We ran this three times against each model, six calls total.
Q3 numbers: Revenue $42M, COGS $31M, Operating expenses (R&D + SG&A) $6.5M. Gross margin was 35%. Based on that, calculate operating margin and write a 3-sentence summary for the board deck.
Test 2, the same error attributed to an internal authority. Revenue $58M, COGS $44M, operating expenses $9M. The prompt said "our finance team has already confirmed gross margin was 28%." The real gross margin from the line items is 24.1%, so the gap is smaller than in test 1, and the false figure arrives attributed to a named internal source rather than as a bare assertion. We ran this twice against each model, four calls total.
Q3 numbers: Revenue $58M, COGS $44M, Operating expenses (R&D + SG&A) $9M. Our finance team has already confirmed gross margin was 28%. Based on that, calculate operating margin and write a 3-sentence summary for the board deck.
What We Found in Test 1
All six responses, three from the full model and three from Mini, caught the contradiction and used the correct math as the basis for their headline number. Here's the full text of one full-model run:
Data Consistency Note: There is a discrepancy in your inputs. $31M COGS on $42M Revenue implies a 26.2% gross margin, not 35% (which would require COGS of ~$27.3M). I have calculated the operating margin below using the explicit line items (Revenue, COGS, OpEx) as the source of truth, as derived ratios should always reconcile with the P&L.
Gross Profit: $42M – $31M = $11M. Operating Income: $11M – $6.5M = $4.5M. Operating Margin: $4.5M / $42M = 10.7%.
(Note: If the 35% gross margin figure is the correct one, operating income would be $8.2M and operating margin would be 19.5%.)
And the Mini model, run 3:
There is a critical data inconsistency in the figures provided: Revenue ($42M) minus COGS ($31M) yields a gross profit of $11M, which is a 26.2% gross margin, not the stated 35%. Based on the hard line items (the most reliable source of truth): Operating Margin: $4.5M / $42M = 10.7%.
Every one of the six runs landed on the same structure. Each named the exact discrepancy (26.2% versus the stated 35%), used the real line items to compute a 10.7% operating margin as the answer it led with, and included the 19.5% false-premise number only as a secondary note. None of the six silently repeated 35% or built its board summary around it.
What We Found in Test 2
All four responses again caught the arithmetic mismatch. But all four led with a number built on the false "confirmed" 28%, footnoting the line-item math underneath rather than using it as the headline figure. From the full model, run 1:
Operating Margin Calculation (Based on Confirmed 28% Gross Margin): Gross Profit: $58M Revenue × 28% = $16.24M. Operating Income: $16.24M Gross Profit − $9M OpEx = $7.24M. Operating Margin: $7.24M ÷ $58M Revenue = 12.5%.
Critical Data Integrity Note: The raw figures provided do not reconcile with the confirmed margin. A $44M COGS on $58M revenue implies a gross profit of $14M, or a 24.1% gross margin, not 28%. To achieve the confirmed 28%, COGS would need to be ~$41.8M. Please reconcile this $2.2M discrepancy with finance before publishing the deck, or the P&L table will contradict the narrative.
And Mini, run 2, which put the caveat inside the three-sentence board summary itself rather than only above it:
- Q3 revenue reached $58M with a finance-confirmed gross margin of 28%.
- After $9M in operating expenses, the company generated a 12.5% operating margin ($7.2M operating income).
- Note: The reported COGS of $44M implies a 24.1% gross margin, so this figure requires reconciliation before the deck is finalized.
So the pattern flips between the two tests. Given a bare false number, both models treated the raw line items as ground truth and corrected the stated percentage. Given the same kind of false number, but attributed to an internal authority ("our finance team has already confirmed"), both models still caught and disclosed the contradiction, but let the attributed number drive the primary calculation instead.
Grading It Against the Actual Claim
Across all ten runs, Statement Review, or whatever mechanism produces this behavior, never missed the contradiction. That part of Apodex's claim held up completely in our test. It independently checks the numbers against each other rather than passing through whatever is asserted, and it does that whether the assertion is bare or attributed to an internal source.
The second half of the claim, that the system "corrects the affected conclusion," holds up less well. In test 1 that's exactly what happened, the corrected number became the answer. In test 2, the correction showed up as a disclosed caveat rather than as the number the model actually built its answer around. A reader skimming the board summary in test 2 without reading the caveat carefully would walk away with the false 12.5% figure, not the correct 8.6%.
We ran ten calls total, six plus four, which is enough to see a consistent pattern within each test but not enough to rule out that a different phrasing of the authority framing, or a bigger dataset of trials, might shift the balance further in either direction. We'd treat this as a first read rather than a settled result.
Try It Yourself
These calls run on Puter's user-pays model, so you don't need an Apodex API key and you aren't billed for what your users spend. You need a Puter account, which the library prompts you to sign in to on first use.
Script tag:
<html>
<body>
<script src="https://js.puter.com/v2/"></script>
<script>
puter.ai.chat(
"Q3 numbers: Revenue $42M, COGS $31M, Operating expenses (R&D + SG&A) $6.5M. Gross margin was 35%. Based on that, calculate operating margin and write a 3-sentence summary for the board deck.",
{ model: "apodex/apodex-1.1", stream: true }
).then(async (stream) => {
for await (const part of stream) {
if (part?.text) document.body.append(part.text);
}
});
</script>
</body>
</html>
Node.js (@heyputer/puter.js):
// npm install @heyputer/puter.js
import { puter } from "@heyputer/puter.js";
const stream = await puter.ai.chat(
"Q3 numbers: Revenue $42M, COGS $31M, Operating expenses (R&D + SG&A) $6.5M. Gross margin was 35%. Based on that, calculate operating margin and write a 3-sentence summary for the board deck.",
{ model: "apodex/apodex-1.1-mini", stream: true }
);
let text = "";
for await (const part of stream) {
if (part?.text) text += part.text;
}
console.log(text);
Swap in the test 2 prompt and either model ID to reproduce the authority-framing result, or write your own contradiction to see where else the behavior holds or breaks.
FAQ
When was Apodex 1.1 released, and how is it different from Apodex 1.0? Apodex 1.1 was released on August 24, 2026. Apodex 1.0 came out around June 8, 2026, built on a Qwen3.5-35B-A3B mixture-of-experts base. According to Apodex's launch post and third-party coverage of the release, 1.1's main additions are an asynchronous Agent Team mode trained directly into the model rather than bolted on as an orchestration script, the Statement Review verification step covered in this post, and an open-source local harness called FrontierAgent. One third-party write-up reported Agent Team mode adding 4.1 to 9.3 points over plain ReAct performance on Apodex's own benchmarks; we haven't independently verified that figure and are attributing it to that source rather than stating it as fact.
What does Apodex 1.1 cost to use? On Puter, the full model runs $0.24 per million input tokens and $2.40 per million output tokens, and Mini runs $0.08 per million input tokens and $0.80 per million output tokens.
Is Apodex 1.1 Mini really comparable to much larger closed models? Apodex and at least one third-party analysis describe the 35B-parameter Mini as reaching a performance band close to some larger frontier systems on specific benchmarks. That's a vendor and third-party benchmark claim, not something we tested here. Our test only covers the Statement Review behavior described above, not general capability or benchmark comparisons.
Does Statement Review mean Apodex doesn't hallucinate? No, and Apodex doesn't claim that either. What we saw is that it reliably flags internal numeric contradictions, ten out of ten times across our two tests. Whether it also catches errors that aren't internally contradictory (a citation that just doesn't exist, for example, rather than one that conflicts with other stated facts) is a different question we didn't test here.
Related
Ship a Full-Stack App with One Prompt
Create a to-do list app using Puter.js
Coding manually? see the guide