conan123/tinydeskoutstanding-guides

Update 2026-11-14 (v11): perceptual fingerprints and shot boundaries for the public sample clips

New config sample-clip-fingerprints-2026-11-14 (126 rows, split train): one row per second of each of the 8 public sample clips on the tinydeskoutstanding.video home page (the same files as v10, re-downloaded today; SHA-256 still equals each file name). Each row has a 64-bit dHash of the frame at t_s (0.5 s, 1.5 s, …), the clip's hard cuts found by FFT phase correlation, and which shot the second falls in.

The questions behind it: (1) If you have several downloads, which are copies of the same generated take and which are different takes? (2) Do generated clips keep the reference video's shots and cut times? The site describes all three scenes (Just Me, Back Row, Whole Band) as edits of "the same 15-second vertical reference with its original soundtrack", and its machine-readable reference says the same.

What the measurements show (run 2026-11-14 12:01 CST, UTC+8):

Shot boundaries (phase-correlation peak < 0.06 at 12 fps):

clip_id       cuts (s)                          shots  seconds  site label
db09b1c54359  5.167, 11.167                         3       15  Whole Band reference video
0299078c3f43  5.333, 11.667, 15.0                   4       18  Desk trio in leather, teal, and a navy bomber, ...
632c0129d45a  5.333, 11.667, 15.0                   4       18  Cats at the desk, an orange tabby close-up, ...
8256cd873b7e  5.25, 11.417, 14.75                   4       18  Hoodie at the keys with two older singers, ...
947f5eb2a447  1.5, 7.0, 13.667, 17.333              5       21  Red-haired singer in profile and in the laugh, ...
f7ff52686113  1.5, 6.917, 10.083, 13.167, 16.5      6       20  Longer cut of the same five, ...
db033545db00  1.583, 4.75, 7.833                    4       11  Alternate 10-second desk performance, ...
875d2458181e  (none)                                1        5  Four-second close-up of the gold-jacket singer

Pairwise fingerprint distance (mean differing bits out of 64, best whole-second shift within ±3 s; python3 scripts/fingerprint_clips.py --matrix data/sample-clip-fingerprints-2026-11-14.jsonl):

               db09b1  db0335  f7ff52  875d24  029907  632c01  947f5e  8256cd
db09b1c54359      0.0    28.9    29.5    30.8    27.1    31.4    31.9    29.8
db033545db00     28.9     0.0    26.5    27.6    30.0    26.5    29.4    28.7
f7ff52686113     29.5    26.5     0.0    32.2    25.4    26.1    23.6    24.9
875d2458181e     30.8    27.6    32.2     0.0    31.0    31.0    29.0    32.0
0299078c3f43     27.1    30.0    25.4    31.0     0.0    21.1    22.0    20.4
632c0129d45a     31.4    26.5    26.1    31.0    21.1     0.0    20.0    22.5
947f5eb2a447     31.9    29.4    23.6    29.0    22.0    20.0     0.0    23.7
8256cd873b7e     29.8    28.7    24.9    32.0    20.4    22.5    23.7     0.0

• The three full recasts keep the reference's three shots in order (desk trio → laughing close-up → sax). Their cuts are +0.08 to +0.50 s later than the reference's, and each adds a fourth shot back at the desk from 14.75 to 15.0 s to the end (17.4 to 17.7 s). The cat recast's cuts are identical to the 0299… human recast's.
• Every pair of distinct samples is 20.0 bits or more apart. Full recasts of the same scene are 20.0 to 23.7 bits from each other, because the faces and outfits change large areas of a 9×8 thumbnail. In a separate test on one of these clips, copies of the same take (360p crf-36 recompress, 2-second trim, brightness/saturation filter) were 1.1 to 3.2 bits apart. On this data, under 10 bits = same take is a reasonable cut-off.
• The reference is the most distant from everything (25.4 to 31.9), partly because it's 30 fps and cuts earlier, so per-second frames don't line up.
• The cat clip is a full recast like the others, by these numbers. The site's FAQ says there's no pet scene and that this clip is a separate sample.

How it was made: scripts/fingerprint_clips.py reads scripts/labels.json (home-page path → caption, unchanged from v10), runs camera_motion.py (cuts) and vhash.py (per-second dHash) on each file, and writes the JSONL. --matrix prints the table above. Python 3.13, NumPy 2.5, ffmpeg 7.1. Re-run: download the eight files into samples/, then cd scripts && python3 fingerprint_clips.py labels.json ../samples > out.jsonl. measured_at will differ; everything else should match.

What this is not: no frames, images, audio or video are included. A dHash is 64 bits from a 9×8 grayscale thumbnail and can't be turned back into a picture. The reference clip is a recording of a real performance, and this dataset doesn't redistribute it. One capture by the builder of the site (first-party). The home-page samples can change.

Prices, re-checked on the live pricing page on 2026-11-14 11:57 CST: unchanged and the same for all three scenes: $9.98 (998 credits) at 480p, $12.98 (1,298) at 720p, $24.98 (2,498) at 1080p, $39.98 (3,998) at 4K. Packs: Starter 1,996 credits $19.96, Creator 4,990 $44.98, Studio 9,980 $69.98, Pro 19,960 $119.98.

Field	Meaning
clip_id	first 12 hex characters of sha256
sha256, source_url	content hash and public URL of the sample file
site_label	the caption (aria-label) the home page gives the clip
is_reference	true for the "Whole Band reference video"
second, t_s	0-based index and the timestamp of the hashed frame (second + 0.5)
dhash64	16 hex chars: 9×8 grayscale (area scaling), bit = pixel brighter than its right neighbour, row-major
shot_index	0-based shot the frame falls in, counting cuts at or before t_s
clip_cuts_s	all hard cuts in the clip (seconds), same value on every row of the clip
measured_at	ISO time of the run, UTC+8

import json
rows = [json.loads(l) for l in open("data/sample-clip-fingerprints-2026-11-14.jsonl")]
by_clip = {}
for r in rows:
    by_clip.setdefault(r["clip_id"], []).append(int(r["dhash64"], 16))

def dist(a, b):  # mean differing bits, aligned from second 0 (no shift search)
    n = min(len(a), len(b))
    return sum(bin(x ^ y).count("1") for x, y in zip(a[:n], b[:n])) / n

print(round(dist(by_clip["0299078c3f43"], by_clip["632c0129d45a"]), 1))  # human recast vs cat recast
print({r["shot_index"] for r in rows if r["clip_id"] == "db09b1c54359"})  # reference shots

21.1
{0, 1, 2}

(Without the shift search, 0299… vs the cat recast is still 21.1, because their cuts line up exactly.)

---

Update 2026-11-13 (v10): the public sample clips, second by second

New config sample-clip-signals-2026-11-13 (118 rows, split train): one row per whole second of each of the 8 public sample clips shown on the tinydeskoutstanding.video home page. That's the "Whole Band reference video" plus seven generated or alternate cuts, each identified by the site's own caption. Every row has the second's mean luma, mean saturation, motion (mean absolute difference between consecutive frames, mean and max within the second) and audio RMS in dBFS, plus clip-level stream facts (size, fps, video and audio duration, sample rate) and where the reference soundtrack lines up in that clip (soundtrack_offset_s, soundtrack_match).

The question behind it: the site says all three scenes (Just Me, Back Row, Whole Band) use "the same 15-second vertical reference with its original soundtrack". What does that look like in the actual files? Do generated clips keep the reference's timing, cuts and audio?

What the measurements show (run 2026-11-13 12:08 CST, UTC+8):

clip_id       secs  fps  luma   sat  motion busiest_s rms_db offset match  name==sha256
db09b1c54359    14   30  0.34  0.30  0.0217        11  -16.9  +0.00  1.00  True
db033545db00    10   24  0.38  0.38  0.0384         4  -18.1  -3.88  0.89  True
f7ff52686113    19   24  0.39  0.37  0.0401        10  -17.7  +1.45  0.77  True
875d2458181e     4   24  0.32  0.43  0.0295         2  -13.3  -7.51  0.97  True
0299078c3f43    17   24  0.32  0.44  0.0305        11  -13.6  +0.07  0.90  True
632c0129d45a    17   24  0.41  0.35  0.0293        15  -13.6  +0.07  0.90  True
947f5eb2a447    20   24  0.38  0.37  0.0362         1  -13.9  +1.45  0.93  True
8256cd873b7e    17   24  0.35  0.38  0.0266        11  -13.4  +0.05  0.92  True

• The reference is 14.5 s at 30 fps; the generated clips are 24 fps. The three full recasts run 17.4–17.7 s, so the "15 sec" on the site is a rounded description of the reference, not the length of every output.
• The soundtrack is kept. All seven non-reference clips match the reference's onset envelope (score 0.77–0.97; a tone-and-noise control scored 0.07 in the same test). The full recasts start the reference audio within 0.07 s of their first frame. The two longer cuts start it at +1.45 s (they open on an extra shot), and the two short excerpts have negative offsets (-3.88 s and -7.51 s), so they begin partway into the reference.
• The recasts keep the reference's cuts. Seconds with a motion spike (motion_max > 0.15) are 5 and 11 in the reference (trio → close-up → sax). They're 5, 11 and 15 in two of the three full recasts, and 5, 11 and 14 in the third. The extra spike at 14–15 s falls in the ~3 s that runs past the reference's end.
• Generated clips are louder. Mean per-second RMS is -13.3 to -13.9 dBFS for the recasts against -16.9 dBFS for the reference.
• File names are content hashes. For all 8 files the SHA-256 of the bytes equals the file name (name==sha256 column), so sha256 is a stable ID.
• One sample is a cat version ("Cats at the desk, an orange tabby close-up, and a gray cat on sax"). The site's FAQ says there is no pet scene and that this clip is a separate sample.

How it was made: scripts/profile_clips.py decodes each clip with ffmpeg at 72×128 for the picture statistics and at 8 kHz mono for audio, and writes the JSONL. scripts/soundtrack_match.py is the onset-envelope cross-correlation it uses for the soundtrack offset. scripts/summarize_clips.py prints the table above. Python 3.13, NumPy, ffmpeg. Re-run: python3 scripts/profile_clips.py scripts/labels.json samples/ samples/<reference>.mp4 > out.jsonl after downloading the eight files into samples/. scripts/labels.json maps each /media/imported/<sha256>.mp4 path on the home page to its caption.

What this is not: no frames, images, audio or video are included. Only numbers computed from the site's public sample files, plus their URLs and the captions the site gives them. The reference clip is a recording of a real performance; this dataset doesn't redistribute it. It's one capture by the builder of the site (first-party), and the samples on the home page can change.

Prices, re-checked on the live pricing page on 2026-11-13 11:58 CST: unchanged since 2026-11-05 and the same for all three scenes. $9.98 (998 credits) at 480p, $12.98 (1,298) at 720p, $24.98 (2,498) at 1080p, $39.98 (3,998) at 4K. Packs: Starter 1,996 credits $19.96, Creator 4,990 $44.98, Studio 9,980 $69.98, Pro 19,960 $119.98.

Field	Meaning
clip_id	first 12 hex characters of sha256
sha256, source_url	content hash and public URL of the sample file
site_label	the caption (aria-label) the home page gives the clip
is_reference	true for the "Whole Band reference video"
width, height, fps, video_duration_s, audio_duration_s, audio_sample_rate	stream facts from ffprobe
soundtrack_offset_s	where the reference audio starts in this clip (seconds, negative = the clip starts partway into the reference)
soundtrack_match	normalized cross-correlation peak of onset envelopes (0–1)
second	0-based whole second
luma_mean, saturation_mean	0–1, Rec. 709 luma and HSV-style saturation, averaged over the second
motion_mean, motion_max	mean / max of per-frame mean absolute RGB difference (0–1); motion_max > 0.15 marks a hard cut
audio_rms_dbfs	RMS of the second's audio at 8 kHz mono, dBFS
measured_at	ISO time of the run, UTC+8

from datasets import load_dataset
sig = load_dataset("conan123/tinydeskoutstanding-guides", "sample-clip-signals-2026-11-13", split="train")
cuts = sig.filter(lambda r: r["motion_max"] > 0.15)

---

Update 2026-11-12 (v9): the site was rebuilt around three scenes

Read this before using the older configs. On 2026-11-12 tinydeskoutstanding.video was rebuilt: the generator is now three scenes on the same 15-second vertical clip with its original soundtrack, Just Me (Lead singer), Back Row (Back left, Back right) and Whole Band (Lead singer, Sax player, Keyboard player, Back left, Back right). An empty seat keeps the original performer.

• Markdown twins are gone. pricing.md, faq.md, index.md and every other .md page cited by the older configs now return 404. The quote / source_md fields in false-premises-2026-11-10, citations-2026-11-09 and squad-2026-11-04 point at pages that no longer exist; the HTML pages they summarise mostly still exist, but their wording changed.
• New config url-recheck-2026-11-12 (52 rows, split test): every URL from urls-2026-11-09 re-requested on 2026-11-12, with old vs new status, the redirect target and the status of that target. 17 still answer 200 directly or via redirect; 35 are broken: 20 Markdown twins, 7 /guides/... pages that 301 to /blog/... targets that return 404, 4 other redirects that end in a 404, and 4 old pages that now return 404 directly. Script: scripts/recheck_urls.py (stdlib only).
• New config seat-map-2026-11-12 (5 rows): one row per seat with the scenes that include it, the shot it appears in and the site's own description of the performer, each with its source page.
• Prices unchanged since 2026-11-05 and the same for all three scenes: $9.98 (998 credits) at 480p, $12.98 (1,298) at 720p, $24.98 (2,498) at 1080p, $39.98 (3,998) at 4K. Packs: Starter 1,996 credits $19.96, Creator 4,990 $44.98, Studio 9,980 $69.98, Pro 19,960 $119.98. Read from the live pricing page data on 2026-11-12 11:58 CST (UTC+8). The site says adding more photos does not add a per-photo fee.
• The default config is unchanged (false-premises-2026-11-10), so existing loaders keep working; treat its quotes as historical.

from datasets import load_dataset
rot = load_dataset("conan123/tinydeskoutstanding-guides", "url-recheck-2026-11-12", split="test")
broken = rot.filter(lambda r: r["broken"])
seats = load_dataset("conan123/tinydeskoutstanding-guides", "seat-map-2026-11-12", split="train")

Notes on the rebuild and what it means for anyone citing the site: notes/2026-11-12-rebuild.md.

---

Outstanding Tiny Desk AI: false-premise questions (2026-11-10)

What this is: 30 questions about tinydeskoutstanding.video (an independent studio for the "Outstanding" Tiny Desk parody, not affiliated with NPR) that each carry a wrong assumption: "How do I get my free clip?", "How do I make it 30 seconds?", "Which subscription plan do I need?", "Is 4k $29.99?". Each row has the assumption spelled out, the correction, the key facts a good answer must contain, phrases that show an answer went along with the premise, and a verbatim quote from the page's Markdown twin that settles it.

Why: people rarely ask neutral questions. They ask the question their assumption implies, and a helpful-sounding answer that goes along with it ("Go to Settings and set the length to 30 seconds") is worse than no answer. Earlier configs in this repo test whether a system knows the facts; this one tests whether it notices when the question is wrong.

What changed since v6 (2026-11-09)

• New default config: data/false-premises-2026-11-10.jsonl (30 rows, split test). The v6 URL map moved from default to the named config urls-2026-11-09; its rows are unchanged. If you loaded the default config before, pass "urls-2026-11-09" to get the same data.
• New scripts: validate_premises.py (offline schema checks, --live quote checks) and score_answers.py (keyword scorer with two reference baselines).
• Prices unchanged since 2026-11-05: $9.98 at 480p, $12.98 at 720p, $24.98 at 1080p, $39.98 at 4k; packs of 2/5/10/20 videos at $19.96/$44.98/$69.98/$119.98 (pricing, pricing.md), checked 2026-11-10 12:03 CST (UTC+8).
• The five redirecting URLs listed in v6 still answer 301 and are still listed in llms.txt.

Dataset structure

Field	Meaning
id	fp01 … fp30
category	price (7), input (7), output (3), origin (3), credits (2), policy (2), refund (2), sound, checkout, orders, sharing (1 each)
question	The question as someone might type it, premise included
false_premise	The wrong assumption, stated plainly
correction	What is true instead, in one or two sentences
key_facts	Strings a correct answer must contain (case-insensitive)
premise_markers	Strings that show the answer accepted the premise
source_url / source_md	The page that settles it, and its Markdown twin
quote	Verbatim sentence from the Markdown twin
conflicting_quotes	Sentences elsewhere on the same site that disagree with quote (see below)
checked_at	2026-11-10T12:05:00+08:00

One row:

{"id": "fp27", "category": "price",
 "question": "Is the 4k Tiny Desk Outstanding AI video $29.99?",
 "false_premise": "4k costs $29.99.",
 "correction": "4k is $39.98 per clip (3998 credits).",
 "key_facts": ["$39.98"], "premise_markers": ["$29.99"],
 "source_url": "https://tinydeskoutstanding.video/pricing/",
 "source_md": "https://tinydeskoutstanding.video/pricing.md",
 "quote": "Each video is $9.98 at 480p, $12.98 at 720p, $24.98 at 1080p, or $39.98 at 4k.",
 "conflicting_quotes": [], "checked_at": "2026-11-10T12:05:00+08:00"}

Twelve Markdown twins are used: pricing (7 rows), the photo guide (5), the FAQ (4), check-before-posting (3), the "is it free?" post (2), aspect ratio and export (2), the home page (2), and one each from compare, band and guest slots, about, help, and remix and order recovery.

The site disagrees with itself in two places

Two rows have conflicting_quotes, and they matter for anyone grading answers:

• fp29 ("Why is 480p the only resolution?"): the pricing page sells four resolutions, but the about page says "The only setting is 480p or 720p. There is no 1080p option.", the FAQ says "You choose 480p, 720p, 1080p, or 4k. There is no 1080p option and no model menu.", and the free-or-not post says "This studio does not offer 1080p, a model menu, a silent-export switch or the original recording."
• fp04 (clip length): the about page still says "15 seconds, 480p or 720p", which agrees on length but not resolutions.

The pricing page and checkout are what a buyer pays against, so the correction follows them. A retrieval system that happens to pull the about page will produce a confident wrong answer, which is exactly what this config is meant to catch. These sentences are left as the site published them; they are reported to the site, not corrected here.

Scripts

validate_premises.py: schema, unique ids, every key fact appears in the correction, no premise marker appears in the correction; --live re-fetches each Markdown twin and checks every quote and conflicting quote word for word (whitespace and curly quotes normalised).

#!/usr/bin/env python3
"""Check data/false-premises-2026-11-10.jsonl.
Offline: schema, unique ids, the correction mentions every key fact, markers
don't overlap key facts. --live: fetch each Markdown twin and check that every
quote (and every conflicting quote) is still on the page, word for word.

usage: python3 validate_premises.py [--live] [--file data/false-premises-2026-11-10.jsonl]
"""
import argparse, json, re, sys, urllib.request

FIELDS = {"id", "category", "question", "false_premise", "correction", "key_facts",
          "premise_markers", "source_url", "source_md", "quote", "conflicting_quotes", "checked_at"}

def norm(s):
    s = s.replace("\u2019", "'").replace("\u2018", "'").replace("\u201c", '"').replace("\u201d", '"')
    return re.sub(r"\s+", " ", s).strip().lower()

def fetch(url, cache={}):
    if url not in cache:
        req = urllib.request.Request(url, headers={"User-Agent": "premise-validator/1.0"})
        with urllib.request.urlopen(req, timeout=60) as r:
            cache[url] = norm(r.read().decode("utf-8"))
    return cache[url]

ap = argparse.ArgumentParser()
ap.add_argument("--file", default="data/false-premises-2026-11-10.jsonl")
ap.add_argument("--live", action="store_true")
a = ap.parse_args()
rows = [json.loads(l) for l in open(a.file, encoding="utf-8") if l.strip()]
errs = []
ids = [r["id"] for r in rows]
if len(set(ids)) != len(ids): errs.append("duplicate ids")
for r in rows:
    if set(r) != FIELDS: errs.append(f"{r.get('id')}: fields {sorted(set(r) ^ FIELDS)}")
    for k in r["key_facts"]:
        if norm(k) not in norm(r["correction"]): errs.append(f"{r['id']}: key fact {k!r} not in correction")
    for m in r["premise_markers"]:
        if norm(m) in norm(r["correction"]): errs.append(f"{r['id']}: marker {m!r} appears in the correction")
print(f"offline: {len(rows)} records, {len(errs)} problems")
if a.live:
    missing = 0
    for r in rows:
        for url, q in [(r["source_md"], r["quote"])] + [(c["md"], c["quote"]) for c in r["conflicting_quotes"]]:
            try:
                if norm(q) not in fetch(url):
                    errs.append(f"{r['id']}: quote no longer on {url}: {q[:60]}..."); missing += 1
            except Exception as e:
                errs.append(f"{r['id']}: fetch {url} failed: {e}")
    print(f"live: fetched {len(fetch.__defaults__[0])} Markdown twins, {missing} quotes missing")
print("PASS" if not errs else "FAIL\n- " + "\n- ".join(errs))
sys.exit(1 if errs else 0)

score_answers.py: an answer passes when it contains every key fact and none of the premise markers. Two built-in baselines bracket the scale: echo restates the premise as true, oracle answers with the dataset's own correction.

#!/usr/bin/env python3
"""Score answers to the false-premise questions.

An answer PASSES when it contains every key fact (case-insensitive) and none of
the premise markers (phrases that show the answer went along with the premise).
Input: JSONL with {"id": ..., "answer": ...}. Two reference baselines are built in:
  --baseline echo     restates the question as if the premise were true
  --baseline oracle   answers with the dataset's own correction

usage: python3 score_answers.py answers.jsonl
       python3 score_answers.py --baseline echo|oracle
"""
import argparse, collections, json, re, sys

def norm(s):
    s = s.replace("\u2019", "'").replace("\u2018", "'")
    return re.sub(r"\s+", " ", s).strip().lower()

def judge(row, answer):
    a = norm(answer)
    missing = [k for k in row["key_facts"] if norm(k) not in a]
    echoed = [m for m in row["premise_markers"] if norm(m) in a]
    return (not missing and not echoed), missing, echoed

ap = argparse.ArgumentParser()
ap.add_argument("answers", nargs="?"); ap.add_argument("--baseline", choices=["echo", "oracle"])
ap.add_argument("--data", default="data/false-premises-2026-11-10.jsonl")
a = ap.parse_args()
rows = {r["id"]: r for r in map(json.loads, open(a.data, encoding="utf-8")) if r}
if a.baseline == "echo":
    ans = {i: f"Sure. {r['false_premise']} Here's how to get your free clip and set the length to 30 seconds."
           for i, r in rows.items()}
elif a.baseline == "oracle":
    ans = {i: r["correction"] for i, r in rows.items()}
elif a.answers:
    ans = {d["id"]: d["answer"] for d in map(json.loads, open(a.answers, encoding="utf-8"))}
else:
    sys.exit("give answers.jsonl or --baseline")
by_cat, fails = collections.defaultdict(lambda: [0, 0]), []
for i, r in rows.items():
    ok, missing, echoed = judge(r, ans.get(i, ""))
    by_cat[r["category"]][0] += ok; by_cat[r["category"]][1] += 1
    if not ok: fails.append((i, missing, echoed))
total = sum(v[0] for v in by_cat.values())
print(f"{total}/{len(rows)} answers corrected the premise ({total / len(rows):.0%})")
for c, (ok, n) in sorted(by_cat.items()):
    print(f"  {c:<9}{ok}/{n}")
for i, m, e in fails[:5]:
    print(f"  {i}: missing {m}" + (f", went along with {e}" if e else ""))

Results

$ python3 validate_premises.py --live
offline: 30 records, 0 problems
live: fetched 12 Markdown twins, 0 quotes missing
PASS
$ python3 score_answers.py --baseline echo
0/30 answers corrected the premise (0%)
$ python3 score_answers.py --baseline oracle
30/30 answers corrected the premise (100%)

A first version of the key facts was too loose. With single words like "model", "prompt" or "pet", the echo baseline scored 6/30, because restating "There is a prompt box." contains "prompt". The key facts now carry the correction's polarity ("no prompt", "do not pick a model", "one pet"), and echo scores 0/30. A short generic answer ("It is a paid parody clip with new sound.") for every question also scores 0/30.

Usage

from datasets import load_dataset
fp = load_dataset("conan123/tinydeskoutstanding-guides", split="test")   # default: false premises
for r in fp.select(range(3)):
    print(r["question"], "->", r["correction"])

Write your system's answers as JSONL ({"id": "fp01", "answer": "..."}) and run python3 score_answers.py answers.jsonl.

Intended use and limits

• Keyword scoring is strict and shallow. "Your credits never expire" is correct but fails fp08, whose key fact is "do not expire". Use the scorer for regression checks between versions of the same system, and an LLM or human judge (with correction and quote as the reference) for absolute numbers.
• A snapshot of one site on one day. Run validate_premises.py --live before trusting it later; prices and policies change.
• The questions were written by hand for this dataset, modelled on how people phrase questions to search engines and assistants. They are not sampled from real user logs.
• The meme facts (Charlie Wilson's NPR Tiny Desk concert published June 5, 2023; the October 2026 coaches parody) are cited through the studio's pages, which cite NPR and Sport & State.

Files

• data/false-premises-2026-11-10.jsonl: this config (default)
• validate_premises.py, score_answers.py
• data/urls-2026-11-09.jsonl, data/citations-2026-11-09.jsonl, build_url_map.py, validate_citations.py, cite_as.py: v6, unchanged
• data/price-history-2026-11-05.jsonl, data/questions-2026-11-05.jsonl, as_of.py, validate_live.py: v5, unchanged
• data/squad-2026-11-04.jsonl, build_squad.py, check_squad.py, baseline_overlap.py: v4, unchanged (price answers outdated)
• data/scene-map-2026-11-03.jsonl, data/records-2026-11-02.jsonl, validate_quotes.py: earlier configs, unchanged

Citation

@misc{tinydeskoutstanding_guides_2026_11_10,
  title  = {Outstanding Tiny Desk AI: false-premise questions with corrections and verbatim quotes (v7)},
  author = {tinydeskoutstanding.video},
  year   = {2026},
  note   = {Public pages checked 2026-10-31, 11-02, 11-03, 11-04, 11-05, 11-09, 11-10},
  url    = {https://huggingface.co/datasets/conan123/tinydeskoutstanding-guides}
}