← Documentation home

Canonical Markdown source · Oct 20, 2018

Visual similarity v2 — technical recommendation

technical-recommendations/visual-similarity-v2-2026-08-25.md · 37 lines · SHA-256 c84543ee37f5

Impact summary

The governed evaluation now has enough deterministic headroom to test value:

six balanced AIC series cases, thirty repeated observations, and a strong 25/30

metadata baseline. Twenty-four official public-domain records replay before any

SigLIP request. Runtime visual ranking remains opt-in, advisory, candidate-

confined, and unable to authorize identity, attribution, influence, provenance,

deployment, or publication.

Ranked coupling and failure findings

  1. Critical — impossible lift: the v1 baseline scored 10/10, so no model could

achieve the registered 10-point improvement.

  1. High — declarative rights: literal rights fields were accepted without

replaying the linked institutional records immediately before provider fetch.

  1. High — filename authority and gate drift: readiness trusted a receipt path,

while the trial allowed $0.05 despite the audit's $0.02 ceiling.

Actions

| Owner | Action | Acceptance criteria | Validation |

|---|---|---|---|

| Visual discovery | Run v2 against a versioned, cost-reporting SigLIP service | 30/30 expected rankings, stable top one in all six cases, no false-friend selection, max cost <= $0.02, max latency <= 3.5 s | `pnpm ai:visual-similarity:trial -- --receipt=artifacts/ai-agent-value/trials/visual-similarity-latest.json --private-dir=artifacts/ai-agent-value/private/visual-similarity-latest` |

| Curatorial data | Maintain the AIC rights/source manifest | All 24 exact records remain public domain; any redirect, oversized payload, invalid response, or status drift blocks before model invocation | `node --import tsx --test --test-concurrency=1 tests/services/visual-similarity-trial.test.ts` |

| Evaluation operations | Obtain two independent blind relevance/effort reviews | Two distinct timed responses bind the exact private packet, with no reviewer identity or candidate data in public artifacts | `pnpm audit:ai-agent-value:readiness` |

| AI reliability | Preserve strict v2 receipt replay | Manifest, count, cost, latency, safety, authority, or extra-key drift fails closed | `node --import tsx --test --test-concurrency=1 tests/services/ai-agent-trial-readiness.test.ts tests/scripts/ai-agent-trial-readiness-script.test.ts` |

Next-cycle hypothesis

A versioned SigLIP service reaches 30/30 stable expected rankings and receives at

least a 10-point independent usefulness lift over the 25/30 metadata baseline,

without extra reviewer effort or any rights, safety, cost, latency, variance, or

authority failure. The provider receipt plus two-reviewer aggregate falsifies the

hypothesis if any gate fails.