Dynamic Capability Routing
Reads each request and selects the tier, the tools and the data it needs, before any model is paid.
✓ At work: 34 requests routed: 32 to the fast tier first, 2 directly to the top tierNEO AI Evaluations · Official LILI Capability Tests
34 analysis tasks · 5 systems · 170 completed answers
LILI and two leading single models answered the same 34 analysis tasks through real calls at real prices. Every answer was timed and priced by the platform, scored blind by two independent AI reviewers, sealed in the Proof Ledger and time-stamped by DigiCert in the second the test ended.
Evidence pack edition 2 of 20 November 2026. The Public Edition was sealed under the programme's former name, NEOLABS. Its file and fingerprint are unchanged.
Results
NEO AUTO, LILI's balanced profile, against Claude Opus 5.5, the strongest reference on this test.
| Measure | NEO AUTO | Claude Opus 5.5 | Reading |
|---|---|---|---|
| Cost per answer | 0.018 USD | 0.090 USD | 4.9× lower cost |
| Time to a complete answer (median) | 19.0 s | 32.9 s | 1.7× shorter time |
| Judged quality (out of 10) | 8.70 | 8.91 | No statistically significant difference detected in judged quality on this test. |
Judged quality: No statistically significant difference detected in judged quality on this test. Against GPT-5.5, NEO AUTO was judged better by +0.95 points out of 10 (95% interval 0.28 to 1.63; n = 32 paired tasks; significant after Holm's correction). Recompute these figures from the per-task results.
| System | Judged quality /10 | Cost per answer | Median time | Best suited to |
|---|---|---|---|---|
| NEO AUTO | 8.70 [8.30-9.03] n=32 | 0.0182 USD | 19.0 s | Demanding work where quality must hold |
| NEO PULSE | 6.26 [5.64-6.86] n=33 | 0.0116 USD | 1.6 s | Instant replies, simple requests |
| NEO PRIME | 5.32 [4.73-5.94] n=33 | 0.0018 USD | 2.8 s | Very high volume at minimal cost |
| Claude Opus 5.5 | 8.91 [8.56-9.20] n=32 | 0.0896 USD | 32.9 s | Reference |
| GPT-5.5 | 7.82 [7.26-8.35] n=33 | 0.0841 USD | 26.2 s | Reference |
Comparator
Switch the measure. Bars show the value; the white line, its 95% interval.
| LILI profile | Cost vs Claude Opus 5.5 | Cost vs GPT-5.5 | Time vs Claude Opus 5.5 | Time vs GPT-5.5 |
|---|---|---|---|---|
| NEO AUTO | 4.9× | 4.6× | 1.7× | 1.4× |
| NEO PULSE | 7.8× | 7.3× | 20× | 16× |
| NEO PRIME | 50× | 47× | 12× | 9.5× |
Statistics
Every comparison is paired on the same task and corrected for the 24 comparisons made at once. 17 of 24 are significant, 13 of them in LILI's favour.
Dot: mean paired difference over 34 tasks. Bar: 95% interval. A bar that does not touch zero is significant.
| Comparison | Measure | Difference | 95% interval | n |
|---|---|---|---|---|
| NEO AUTO vs Claude Opus 5.5 | Total time | 10.73 s faster | 7.63 to 14.01 | 34 |
| NEO AUTO vs Claude Opus 5.5 | Cost per answer | 0.0714 USD cheaper | 0.0506 to 0.0967 | 34 |
| NEO AUTO vs GPT-5.5 | Total time | 8.77 s faster | 3.58 to 14.00 | 34 |
| NEO AUTO vs GPT-5.5 | Cost per answer | 0.0659 USD cheaper | 0.0488 to 0.0839 | 34 |
| NEO PRIME vs Claude Opus 5.5 | Total time | 26.12 s faster | 22.60 to 29.68 | 34 |
| NEO PRIME vs Claude Opus 5.5 | Cost per answer | 0.0878 USD cheaper | 0.0716 to 0.1102 | 34 |
| NEO PRIME vs GPT-5.5 | Total time | 24.17 s faster | 17.95 to 31.02 | 34 |
| NEO PRIME vs GPT-5.5 | Cost per answer | 0.0823 USD cheaper | 0.0664 to 0.1001 | 34 |
| NEO PULSE vs Claude Opus 5.5 | Total time | 29.14 s faster | 25.48 to 32.93 | 34 |
| NEO PULSE vs Claude Opus 5.5 | Cost per answer | 0.0781 USD cheaper | 0.0624 to 0.0972 | 34 |
| NEO PULSE vs GPT-5.5 | Total time | 27.18 s faster | 21.99 to 33.19 | 34 |
| NEO PULSE vs GPT-5.5 | Cost per answer | 0.0726 USD cheaper | 0.0576 to 0.0892 | 34 |
| Comparison | Difference /10 | 95% interval | n |
|---|---|---|---|
| NEO AUTO vs GPT-5.5 | +0.95 points (LILI judged better) | 0.28 to 1.63 | 32 |
| NEO PRIME vs Claude Opus 5.5 | -3.64 points (reference judged better) | -4.41 to -2.91 | 32 |
| NEO PRIME vs GPT-5.5 | -2.50 points (reference judged better) | -3.12 to -1.89 | 33 |
| NEO PULSE vs Claude Opus 5.5 | -2.72 points (reference judged better) | -3.44 to -1.92 | 32 |
| NEO PULSE vs GPT-5.5 | -1.56 points (reference judged better) | -2.19 to -0.89 | 32 |
Seven comparisons are not shown: NEO AUTO against Claude Opus 5.5 on judged quality (no statistically significant difference detected in judged quality on this test) and the six exact-match comparisons: the analysis tasks of this test have no single known answer, so exact-match accuracy has no paired task to compare. Under the rules of NEO AI Evaluations a result that is not significant is never reported as a tie or a win. NEO PULSE and NEO PRIME are built for instant and high-volume work; on these demanding analysis tasks the references keep the lead on judged quality, which is why LILI fits the profile to the request.
of judgements returned: 281 of 340. 32 of 34 tasks have every system scored.
of the 118 answers scored by both reviewers are within one point (mean difference 1.04; correlation 0.718).
within one point on the position check (32 pairs), which guards against order effects.
The eight NEO layers
LILI is a layer of intelligence above the best models of every provider: it decides how much compute each request deserves, verifies the result, remembers, and proves what it did.
Reads each request and selects the tier, the tools and the data it needs, before any model is paid.
✓ At work: 34 requests routed: 32 to the fast tier first, 2 directly to the top tierA fast, low-cost model answers first. A top-tier model is called only when the answer needs it.
✓ At work: 32 first answers by the fast tier; 2 escalated to the top tierAn independent model from another provider scores the answer. Evidence decides whether more compute is spent.
✓ At work: 32 answers verified and acceptedLILI sends the facts that matter from its memory graph and a running summary, instead of the whole history.
✓ At work: 34 requests answered with graph factsFor expert work: up to three models from different providers answer in parallel, then one synthesis and one critic.
Built; not at work on this testTwo approaches are sketched in parallel and scored blind; the stronger one is developed.
Built; not at work on this testA verified answer to an equivalent question is reused for the same person instead of being computed again.
Built; not at work on this testDesigned to place models on API, private or self-hosted capacity. Today every model runs through APIs.
Built; not at work on this testLayers 5 to 8 are built and used by other profiles and request types; they were not at work on this test.
LILI decisions
The saving comes from calls that were not made: on 30 of 34 tasks LILI established by verification that a fast-tier answer was good enough. Every escalation is counted in the averages: the 4.9× advantage is net of them.
| What LILI did | Tasks | Typical cost | Total |
|---|---|---|---|
| First answer by a fast-tier model, verified and accepted; top-tier call avoided | 30 | 0.0027 USD | 0.123 USD |
| First answer judged insufficient by LILI; escalated to the top-tier model | 2 | 0.141 USD | 0.281 USD |
| Request routed directly to the top-tier model, then verified | 2 | 0.107 USD | 0.215 USD |
| NEO AUTO, 34 answers | 34 | 0.018 average | 0.619 USD |
| Claude Opus 5.5, 34 answers | 34 | 0.090 average | 3.047 USD |
A founder wants one assistant that remembers users across five different products. Give the strongest argument for a single shared memory, the strongest argument for keeping each product's memory separate, and the design you would recommend, with the consent and deletion rules it needs.
The question is received with its context. Nothing is sent to a model yet.
0 USDLILI classifies the request and selects a fast-tier model for the first answer.
0 USD - no model callThe fast-tier model writes the full answer: 2,054 output tokens.
0.00156 USDAn independent model from a different provider scores the answer: 1.0 out of 1.0, accepted.
0.00056 USDThe cascade was armed. Because the answer was accepted, the top-tier model was not called.
call avoided| System | Calls | Time | Cost | Judged /10 |
|---|---|---|---|---|
| NEO AUTO | 2 | 13.6 s | 0.00212 USD | 9.0 |
| Claude Opus 5.5 | 1 | 30.8 s | 0.06315 USD | 9.0 |
| GPT-5.5 | 1 | 34.1 s | 0.08114 USD | 7.0 |
Integrity
The fingerprint of all 170 answers and 396 judgements was sealed in one ledger entry, then certified in time by DigiCert.
Every official test is recorded in the Proof Ledger, the built-in register of NEO AI. Cryptographic hashing makes any change to a sealed record detectable, so the integrity of the recorded results can be checked.
This demonstration recomputes the real SHA-256 of entry no. 1285 in your browser. Altering one figure (170 answers becomes 171) changes the hash, and every later link breaks.
In the second the campaign ended, the hash of its sealed entry was sent to DigiCert, a public certificate authority, under RFC 3161, the international standard for trusted time-stamps. DigiCert signed the hash with its own clock. Only 32 bytes left NEO AI: no question, no answer, no figure, no name. The evidence pack published afterwards is signed through Sigstore and time-stamped again by FreeTSA.
Results timestamped by DigiCert · RFC 3161
Download the entry and recompute its SHA-256 with the published rule.
Download the RFC 3161 token and read what it certifies: the hash, the time and the serial.
The cryptographic signature is checked with free software and DigiCert's public root certificate. One command, without NEO AI:
openssl ts -verify -in test-2-ledger-1285.tsr -digest b6aec2b0fd46d143fac82b8ba9c63e0edf960084c93b2fbb24456fd2baedb7f2 -CAfile digicert-trusted-root-g4.pem
Expected answer: Verification: OK. Root certificate SHA-256 fingerprint: 55:2F:7B:DC:F1:A7:AF:9E:6C:E6:72:01:7F:4F:12:AB:F7:72:40:C7:8E:76:1A:C2:03:D1:D9:D2:0A:C8:99:88.
The results digest (fb70d29b...8c5b) covers the 170 answers and 396 judgements; recomputing it needs those records, which are available on request under confidentiality. The ledger entry and the token can be checked by anyone with the files above.
NEO AI (built-in register; not a third party)
Role. The hash-chained, append-only register of NEO AI. Each entry contains the fingerprint of the one before it.
What it confirms. Proof Ledger entry no. 1285 has not changed since it was sealed, and it is chained to every entry before it. It holds the fingerprint of all 170 answers and 396 judgements of the test.
Applies to. Official Test No. 2: Proof Ledger entry no. 1285 (results of 170 answers and 396 judgements)
Scope. Operated by NEO AI: on its own it proves integrity, not a date. The date comes from DigiCert.
Check it yourself, without NEO AI:
node verify-entry.mjs ledger-entry-1285.json
DigiCert, Inc. - public certificate authority
Role. A public certificate authority that issues RFC 3161 time-stamps.
What it confirms. The fingerprint of Proof Ledger entry no. 1285 existed, unchanged, on 18 November 2026 at 12:10:44 UTC.
Applies to. Official Test No. 2: the hash of Proof Ledger entry no. 1285
Scope. DigiCert certifies a date for a fingerprint. It has not reviewed the report, the tasks or the scores.
Check it yourself, without NEO AI:
openssl ts -verify -in test-2-ledger-1285.tsr -digest b6aec2b0fd46d143fac82b8ba9c63e0edf960084c93b2fbb24456fd2baedb7f2 -CAfile digicert-trusted-root-g4.pem
FreeTSA (freetsa.org) - free RFC 3161 time-stamp authority
Role. A free, community-run time-stamp authority (RFC 3161).
What it confirms. The manifest of the evidence pack existed at the stated time. The manifest lists the PDF report, the per-task results and the ledger entry with their SHA-256 fingerprints.
Applies to. Official Test No. 2, evidence pack edition 2: the manifest d98d2df7d7eadfcf… (PDF report, per-task results, ledger entry)
Scope. Certifies a date for the manifest. It does not review its content.
Check it yourself, without NEO AI:
openssl ts -verify -data proofs/manifest.jcs -in proofs/freetsa.tsr -CAfile proofs/freetsa-cacert.pem -untrusted proofs/freetsa-tsa.crt
Sigstore - open-source project of the OpenSSF, Linux Foundation
Role. Fulcio gives the public GitHub workflow a short-lived signing certificate. Rekor records the signature in a public, append-only transparency log.
What it confirms. The manifest was signed by the public workflow of github.com/myneogroup-ai/neolabs-proofs, and that signature is recorded in the public Rekor log.
Applies to. Official Test No. 2, evidence pack edition 2: the manifest d98d2df7d7eadfcf… (PDF report, per-task results, ledger entry)
Scope. Confirms who signed the manifest and that it is unchanged since. It does not judge the figures inside it.
Check it yourself, without NEO AI:
cosign verify-blob --bundle proofs/manifest.sigstore.json --certificate-identity https://github.com/myneogroup-ai/neolabs-proofs/.github/workflows/proofs.yml@refs/heads/main --certificate-oidc-issuer https://token.actions.githubusercontent.com proofs/manifest.jcs
GitHub, Inc. (through Sigstore)
Role. Signed build provenance issued by GitHub Actions and recorded in the Rekor log.
What it confirms. The published verifier and scorer (verify/*.mjs) come from the public workflow of the neolabs-proofs repository, at a given commit.
Applies to. The published verifier and scorer of the public proof repository (verify/*.mjs)
Scope. Confirms where the verification code comes from, not that the tested system performs well.
Check it yourself, without NEO AI:
gh attestation verify verify/score.mjs -R myneogroup-ai/neolabs-proofs
OpenTimestamps - open protocol and public calendars, anchored in Bitcoin
Role. A free, open time-stamping protocol anchored in the Bitcoin blockchain, with no company to trust.
What it confirms. Once confirmed: the manifest existed before a given Bitcoin block.
Applies to. Official Test No. 2, evidence pack edition 2: the manifest d98d2df7d7eadfcf… (PDF report, per-task results, ledger entry)
Scope. Confirms a date only, with no company to trust. Anyone can check it on their own node with ots verify.
Check it yourself, without NEO AI:
ots verify proofs/manifest.jcs.ots
Published by NEO TECH LABS TECHNOLOGY L.L.C. on GitHub
Role. Hosts the evidence packs, the verifier and the workflow that signs and time-stamps them, with their full public history. Code under MIT, data under CC BY 4.0.
What it confirms. Nothing by itself: it is where anyone can download the files and rerun the checks offline (node verify/cli.mjs verify).
Applies to. Every edition of the evidence pack and the code that verifies it
Scope. A host, not a certifier.
Shown so that nothing is implied: these services or reviews are planned or waiting for a decision. None of them has confirmed anything yet.
Report text only (Public Edition PDF): the six comparisons not shown besides judged quality are the exact-match comparisons, not time-to-first-text comparisons. Figures changed: no.
Manifest d98d2df7d7eadfcf0175d4f9e927830d96471b917020476c03fdb857b77ad833 · PDF 108b44330f0d5a81b566d14bb8f067980720c09e053368943ec15c74a6915779
First published edition.
Manifest a58336b0134bd4387051c185113ee187fdbfe2b28edeae29be7276a82a08cea0 · PDF 149736a95c717e5482c5060c1168e5dbceff79dfe2cbf6fcd5bccf1ea5e2362b
When a new edition is published, its manifest names the fingerprint of the one it replaces, and receives its own signature and time-stamps. Earlier editions stay online with their proofs at their permanent addresses: each file always matches the fingerprint of its own edition.
Timeline
34 original analysis tasks version-locked before the test.
The first of 170 live calls leaves the Model Lab.
170 answers and 396 judgements stored; their fingerprint is computed.
Entry no. 1285, chained to the 1,284 entries before it.
RFC 3161 token, serial 5da7bc65955b06f9c5efc6871bbf4a10. Only the 32-byte hash left NEO AI.
Methodology, results and proof files made public (the programme was then named NEOLABS).
Manifest signed by the public workflow (Sigstore, Rekor) and time-stamped by FreeTSA; scorer provenance attested by GitHub.
Report text corrected (comparisons not shown); no figure changed. Signed, logged in Rekor and time-stamped by FreeTSA again; edition 1 stays online.
Method & limits
No figure was typed by a person. Each one is computed from records written by the platform while the test ran: the frozen task set, live production calls, the tokens reported by each provider, the server's timing events, the blind review and the statistical engine.
More task families after 34 analytical tasks: reasoning, code, long documents and more languages.
Two independent AI reviewers score blind today; human experts will be added on a sample.
Larger samples, to measure the gap with the strongest reference more finely.
Date and integrity are certified by a third party; next, the test reproduced by an outside party.