AffinityQA

Qloo-powered diagnosis and repair for personalization failures.

Comment

Inspiration

A recommendation can look plausible while using somebody else's taste profile. A missing cache-key field or misrouted tool response may survive ordinary output checks. We built AffinityQA to expose that failure, make a supported correction and verify that the right context reaches the agent again.

What it does

AffinityQA is a Qloo-powered personalization diagnosis and repair workbench. The main journey is DETECT → DIAGNOSE → REPAIR → VERIFY. It traces the requested interest through Qloo's response, the consumed context, model input, decision and cache. Three controlled incidents have matching repairs:

  • A cache key omits the requested profile: include the profile in the key.
  • The agent receives a stale profile: restore the current requested profile.
  • The Qloo context belongs to another profile: bind it to the requested profile.

A companion discovery journey takes a film or artist and returns five films. Qloo resolves identity and retrieves up to 100 candidates for that case. Groq Free / GPT-OSS 20B selects eligible indices; the server reconstructs canonical titles and years. Optional preferences reach the model. Film descriptions are tied to the saved source or honestly marked unavailable. Original decorative illustrations are retained; they are not film posters. Artist/song output, playback and exhaustive worldwide coverage are not claimed.

Judge quickstart

1. Inspect a diagnosis and repair

Open DIAGNOSIS & EVIDENCE. No installation, Docker, personal API key or running creator PC is required.

  1. Select Alice Coltrane / Lil Nas X, Qloo context bound to wrong profile, Repeat 1, then Replay this incident.
  2. Follow the four stages. Inspect the requested Lil Nas X profile, the incorrect Alice Coltrane context and the corrected binding.
  3. Open WHAT DID QLOO CHANGE? In this controlled saved comparison only the consumed Qloo context changes; interest, task, ordered catalog, model setup and output format stay fixed. Four of five films change, and the repaired list matches the healthy reference.
  4. Inspect the eight evidence checks, thirteen comparison checks and six recovery controls. The other incidents demonstrate cache/profile repairs.

This path uses captured Qloo responses and saved local-model decisions, with zero new provider calls. It demonstrates controlled integration recovery, not fresh inference, a no-Qloo benchmark or improved human taste.

2. Try your own discovery

Open FILM DISCOVERY, then OPEN OR RESTORE MY SESSION.

  1. Start with a film or artist. Enter the interest and optional direction; artist mode can also include a known favorite film.
  2. Review the plan, use FIND once, and select the intended identity. Confirm a film's title and year before proceeding.
  3. Use CONFIRM & DISCOVER FIVE once. Inspect the five cards, source-linked descriptions and HOW THESE WERE CHOSEN.
  4. Download the verification. Open SAVED, reload in the same browser profile and reopen the case. Reading saved results does not generate new work.

Repeatable mode has no lifetime plan/execution cap per visitor session. Shared Qloo/Groq Free capacity remains finite: follow waiting/failure messages, preserve the browser cookie and do not repeat uncertain sends. The recorded diagnostic route remains available without fresh provider calls. The existing ownership policy extends through 17 November 2026, 04:45 UTC, covering the official 2–5 November 2026 judging window. This is an application policy, not a reservation of provider quota or a guarantee of future key validity.

Why Qloo matters

Qloo supplies live identity resolution and cross-domain affinity data, as well as the context under test in diagnosis. The verifier binds the saved tool response, consumed context, actual model input and observed decision to the requested profile. Repair must restore that binding while retaining legitimate cache reuse. A plausible-looking list alone does not pass those checks.

How we built it

Python/FastAPI handle diagnosis, supported adapter/session/cache corrections, capture and verification. Next.js/React/TypeScript provide the new interface. Session ownership, CSRF, durable scheduling, persistent accounting and versioned readers preserve old cases and receipts. Diagnostic experiments used a local model; current hosted discovery uses Groq Free. The model selects indices; it does not supply invented plot details or explanation prose to the cards.

Verified technical review — 22 October 2026

Deployed revision: 03926c9. Its Windows/Linux checks passed 801 Windows tests and 797 Ubuntu tests, with zero failures or errors. Four native Windows contracts were skipped on Ubuntu and passed on Windows. Types, build and public-source packaging passed on both. Ten Linux browser workflow blocks produced 26 PASS reports across 16 Node invocations using synthetic providers; Windows browser CI steps were intentionally skipped. Synthetic browser tests and real-provider acceptance are reported separately.

The original recorded capture passed 54 local replays, including 18 controlled context comparisons, with all 384 original files unchanged and zero new Qloo/Groq calls. The interface was also walked through locally across all three supported incidents and their evidence/recovery controls.

Fresh hosted acceptance on this revision: two complete discoveries. The Daft Punk artist case included TRON: Legacy (2010) as a known favorite and an explicitly synthetic science-fiction/electronic direction. The second started with Scary Movie (2000) and a synthetic parody/horror-comedy direction. Each returned five distinct films from a Qloo pool of 100 candidates, with five original illustrations and source-bound descriptions. The favorite was excluded from the Daft Punk output.

The two cases completed with exactly 7 new Qloo requests and 2 Groq Free selections, without retries. All four physical verification downloads were checked: each original/restored pair was byte-identical after reloading and restoring the same browser session. Restoration did not resend provider work. Native accounting and file/row reconciliation preserved all 29 previous cases, including partial and failed cases, and added only these two cases. No quotas were reset; keys, model, illustrations and the current plan were preserved. A normal same-revision restart was included before the final preservation audit; no out-of-memory kill was observed during this bounded acceptance window.

Separately, the current hosted interface was walked through all three supported recorded incidents, all four stages per incident, eight evidence checks and six recovery controls per incident. The wrong-context comparison passed thirteen controls and showed four of five films changing. These diagnostic replays made zero new Qloo/model calls; they used saved local-model decisions, not fresh Groq inference.

VERIFIED_COMPLETE verifies a saved discovery execution and selection; it does not certify causal repair or cultural quality. Discovery's causal and behavioral gates are NOT_EVALUATED. The historical evaluated-agent release gate remains BLOCKED, separately from application availability.

The gallery contains four actual screenshots of the current hosted interface, captured on 22 October 2026: diagnosis, Qloo-context repair, Daft Punk discovery and Scary Movie discovery. Earlier interface images are archived privately, outside the public gallery. Original card artwork is decorative. No external film posters or raw private provider captures are redistributed. The local Daft Punk trial's failed automated download remains recorded; it is not counted as a complete local download/restoration test. Evidence types and technical review.

English translations of screenshot text

English translations of visible Spanish descriptions (Qloo source, not independently fact-checked):

  • District 9: An agent transforms after alien-biotechnology infection and faces corporate exploitation and xenophobia.
  • Rise of the Planet of the Apes: An intelligence-boosting chimpanzee treatment helps trigger an uprising against human arrogance.
  • I, Robot: A technophobic detective investigates a scientist's death and possible robot involvement.
  • Green Lantern: An alien ring recruits a reckless pilot into an intergalactic police force.
  • Jumper: A teleporting young man is drawn into conflict with hunters targeting his kind.
  • Scary Movie 2: Students visit a haunted mansion for university work and encounter absurd supernatural events.
  • Scary Movie 3: A woman investigates crop circles and a cursed tape amid a suspected alien invasion.
  • Scary Movie 4: A haunted-house investigation crosses horror and science-fiction parodies while confronting a growing threat.
  • Not Another Teen Movie: A broad teen-film parody satirizes school romance and coming-of-age clichés with exaggerated gags.
  • EuroTrip: After a graduation breakup, friends take a chaotic European trip to win back an ex.

“Fuente: Qloo · Español” means “Source: Qloo · Spanish.” The Spanish source-fidelity notice means: “AI selects; Qloo supplies the description. Fidelity to the source does not establish its truth or cultural fit.” These translations explain the screenshots; they do not add model-generated explanations to the app.

What we learned and limits

Correct context and traceable inputs matter as much as plausible outputs. Supported repairs cover these three integration faults, not arbitrary program repair or universal recommendation correctness.

Human testing: the creator reports successful informal testing by three real people: himself and two friends. He tested Eminem; one friend used Fast & Furious and liked the car/tuning recommendations, while the other used Scream and received related horror recommendations. This small convenience sample is qualitative, without complete per-film ratings or exact release records for those trials. It is separate from synthetic role-play and automated operator tests. Systematic independent cultural quality remains NOT_VALIDATED; the retained historical remote causal campaign remains INCONCLUSIVE.

Next: independent preference evaluation and broader controlled fault coverage. Original code is MIT licensed; third-party data and models retain their own terms. Public source includes original assets, synthetic fixtures and setup instructions, while excluding keys, private captures and owner case stores. Provider capacity and event-key validity during judging are not guaranteed.

Live diagnosis · Film discovery · MIT source · Judge guide

Built With

Share this project:

Updates

Submission history