Jev results, segment IDs mode

Metric Key

Quick reference for the columns in every table below.

Metric	Range	Direction	What it means
F1 (accuracy)	0 to 1	higher is better	Combined score of precision and recall against the human-verified ground-truth ad spans. F1 = 0 means the model found nothing right; F1 = 1 means it found every ad with the correct boundaries. Uses IoU >= 0.5 (predicted span must overlap truth span by at least half) to count a match, after both sides are canonicalized to per-break spans.
F0.5 @0.5 / F0.5 @0.8	0 to 1	higher is better	The same F0.5 estimator at two IoU match thresholds. Ranking uses the @0.5 column; the @0.8 column re-scores the identical predictions where a match must line up on the boundaries, so a large drop between the two means the rank is an artifact of loose matching rather than accurate cuts. n/a means the row carries no strict-threshold scores.
Cost / episode	USD	lower is better	Average dollars per episode at the current pricing snapshot. Recomputed from token counts so all rows compare at the same prices regardless of when the call ran.
F1 / $	ratio	higher is better	F1 divided by cost-per-episode. Cheap accurate models score highest. Free-tier models (when the roster has any) are rank-listed separately because the ratio is undefined.
p50 / p95 latency	seconds	lower is better, with caveats	Median (p50) and tail (p95) wall-clock response time. Note: for models routed through OpenRouter (everything except claude-*), this includes OpenRouter’s queueing and upstream-provider latency, not just the model itself. Treat as a load/availability indicator, not a model-quality signal.
JSON compliance	0 to 1	higher is better	Fraction of responses that parsed as a clean JSON array matching the requested schema. 1.0 = always clean; lower = used object wrappers ({ads: [...]}), markdown fences, extra fields like sponsor, or required regex fallback to extract.
No-ad episode	PASS / FAIL	PASS desired	Negative-control test on the 2 episode(s) verified to contain no ads (ep-ai-cloud-essentials-e8dc897fbd6b, ep-oxide-and-friends-ce789ff5b62e). PASS = zero predictions across all 16 of their windows. FAIL = the model false-positived on a non-ad segment, with the FP count shown.
F1 stdev	0 to 1	lower means more consistent	Standard deviation of F1 across the 12 ad-bearing episodes. High stdev = inconsistent across content types.
Moderation blocked	0 to 100%	0 desired	Share of attempted calls the provider refused on content grounds. Refused windows never reach scoring, so any non-zero value means that model’s F1 was computed on a subset of the corpus and is not comparable to an unblocked row.
JSON mode	native / prompt-inject / mixed	–	How the model received its JSON-output instruction. native = provider accepted response_format=json_object for at least 95% of calls; prompt-inject = provider rejected it and the runner fell back to instructing JSON in the prompt for at least 95% of calls; mixed = neither path crossed the threshold (sample mostly comes from intermittent provider rejections). Reads from json_format_used in calls.jsonl. Useful when picking a model whose provider may not support native JSON mode – a strong JSON compliance score from a prompt-inject model carries different weight than the same score from a native model.

Glossary

• IoU (intersection over union): how much two time ranges overlap, expressed as (overlap) / (union). 0 means no overlap, 1 means identical ranges. We use IoU >= 0.5 as the threshold for a predicted ad to count as matching a truth ad.
• Per-break canonicalization: before matching, predicted and truth spans separated by gaps under 15 seconds are merged, so one span means one contiguous ad break. This mirrors the detection prompt’s own merge rule; a model that reports one break as several adjacent spots is not penalized for the split.
• Trial: each (model, episode) pair runs 5 trials at temperature 0.0 to surface non-determinism. F1 numbers in tables are averaged across trials.
• Window: each episode is split into 600-second sliding windows at a 420-second stride, so consecutive windows overlap by 180 seconds; the model judges each window independently. Per-window predictions are stitched together for episode-level scoring.
• Schema violations: number of times the response had at least one missing-required-field, wrong-type, or extra-key issue. Doesn’t tank F1, but signals brittleness.
• Extraction method: the route the parser took to recover the ad list. json_array_direct is the cleanest; method names with regex_* mean the JSON itself was malformed and we fell back to text matching.

TL;DR

Best Accuracy (F0.5 @ IoU >= 0.5)

Models ranked by F0.5 (precision weighted 2x recall) against human-verified ground truth. MinusPod cuts the segments it flags, so cutting real content (a false positive) is worse than leaving an ad in (a false negative), and F0.5 penalizes it more. A model shares the tier above it unless it scores consistently lower across the same episodes (paired one-sided t-test, 95%); models that trade wins episode to episode share a tier, so order within a tier is not meaningful on this 12-episode corpus. Flags caveat a model without changing its rank.

Tier	Model	F0.5 @0.5	F0.5 @0.8	95% CI	Precision	Recall	F1	Cost / episode	p50 latency	JSON compliance	Flags
A	claude-haiku-4-5-20251001	0.905	0.722	+/-0.068	0.900	0.937	0.915	$0.0773	16.0s	1.00

Best Value (F0.5 per dollar)

Paid-tier only, ranked by F0.5 per dollar. Free-tier models are excluded here because F0.5 / 0 is undefined (none in this campaign’s roster came back at $0.00). No confidence tiers on this table, since a point ratio does not group cleanly, but the reliability flags still apply.

Rank	Model	F0.5/$	F0.5	F1	Cost / episode	Flags
1	claude-haiku-4-5-20251001	11.71	0.905	0.915	$0.0773

Charts

Cost vs F1 (Pareto)

Each model is one colored point. Lower-left is unhelpful (expensive, inaccurate). Upper-left is the sweet spot (accurate, cheap). The legend below the chart shows each model’s color next to its F1 and cost-per-episode.

Source data: Best Accuracy, Best Value

Accuracy vs latency

F0.5 (y) against p50 latency (x, log scale). The cost Pareto above answers what accuracy costs in dollars; this one answers what it costs in wall-clock time. Upper-left is accurate and fast. MinusPod’s pipeline is offline, so a slow accurate model is usable, but the chart shows which models make you choose and which don’t. The OpenRouter latency caveat from the Metric Key applies.

Source data: Best Accuracy (F0.5), Latency tail (p50)

JSON schema compliance

Fraction of each model’s responses that parsed as a clean JSON array. 1.0 means every response came back exactly as requested; lower numbers mean the parser had to recover from markdown fences, object wrappers, or extra fields.

Source data: Per-Model Detail (JSON compliance field)

F1 by episode (heatmap)

F1 score for each (model, episode) pair. Greener is more accurate, redder is less. The no-ad episodes are excluded. They have no F1 because they’re PASS/FAIL negative controls.

Source data: Quick Comparison, Per-Episode Detail

Confidence calibration (heatmap)

One row per model, one column per self-reported confidence bin. Cell text is the actual hit rate at that bin plus the sample size; cell color is the calibration error (actual minus bin midpoint). Red cells mean the model claimed high confidence but was usually wrong; green is well-calibrated; blue is underconfident. Empty cells mean the model never produced a prediction in that bin. Models are sorted from most overconfident at the top to most underconfident at the bottom.

Source data: Confidence calibration table

Latency percentiles

p50, p90, p99, and max per model on a log scale. The gap between p99 and max indicates how heavy the tail is. For OpenRouter-routed models, the tail also includes upstream provider load.

Source data: Latency tail table

Cross-model agreement (window distribution)

Histogram of how many models flagged at least one ad per (episode, window). The left side is windows nobody flagged (clear non-ad content), the right side is windows everyone flagged (clear sponsor reads). Bars in the middle are contested (some models said yes, some said no) and are candidates for ensemble voting or manual review. This view is anonymous (bars don’t show which models contributed); the per-model breakdown is in the next chart.

Source data: Cross-model agreement table

Per-model alignment with majority

Stacked horizontal bar per model. Green + blue segments are windows where the model voted with the majority (true positives + true negatives); orange is windows where it voted yes but most others voted no (likely false positive / hallucination); red is windows where it voted no but most others voted yes (likely missed real ad). Right-edge label is alignment rate. High alignment means the model tracks consensus; low alignment is either insight or noise depending on whether those broken-from-consensus calls were right.

Source data: Per-model alignment with consensus table

Precision vs Recall (with F1 isocurves)

Scatter of precision (y) vs recall (x) for each model. Dashed gray lines are F1 isocurves; points on the same dashed line have the same F1. Top-right is ideal (high precision AND high recall). Top-left is cautious (high precision, low recall). Bottom-right is greedy (high recall, low precision). Useful for picking a model whose error profile matches your tolerance: precision-leaning for environments where false positives are expensive, recall-leaning for completeness-first.

Source data: Precision, recall, and FP/FN breakdown table

Boundary accuracy (start + end MAE)

Stacked horizontal bars per model: blue is mean absolute error on the predicted ad START in seconds, orange is the same for END. Total error labeled at the right. Sorted by total ascending so the cleanest boundaries are at the top. Skewed bars (start much larger than end, or vice versa) mean the model systematically overshoots on one side. Relevant if you cut audio downstream.

Source data: Boundary accuracy table

Token efficiency vs F1

Scatter of output tokens per detected ad (x, log scale) vs F1 (y). Upper-left is the efficient zone: high accuracy with few output tokens. Right-side points are reasoning-heavy models that emit chain-of-thought alongside their JSON. The chart answers whether the extra tokens buy more F1 or just burn output budget. A model that lands far right at modest F1 is paying for reasoning that didn’t help.

Source data: Output token efficiency table

Cost split (input vs output)

Stacked horizontal bars per model: blue is the input share of per-episode cost, orange is the output share, total labeled at the right, sorted by total ascending. Every model reads the same transcripts, so a long blue bar is an expensive input price and a long orange bar is a talkative model. Reasoning models show up as mostly orange.

Source data: Cost breakdown (input vs output) table

Trial variance (determinism check)

Horizontal bars of mean F1 stdev across episodes per model. All trials run at temperature 0.0 so well-behaved models cluster near zero. Bars are color-graded: green below 0.02 (effectively deterministic), yellow 0.02-0.05 (slight noise), red above 0.05 (single-trial F1 numbers from this model should be treated with suspicion). Dotted reference lines mark the 0.02 and 0.05 thresholds.

Source data: Trial variance (determinism check) table

Detection rate by ad length

Heatmap of model (row) vs ad-length bucket (column), cell = detection rate with sample size. Greener = caught more ads in that bucket; redder = missed more. Models are sorted by overall detection rate so the strongest are at the top. Empty (gray) cells mean that bucket had no truth ads for the corresponding model’s trials.

Source data: Detection rate by ad characteristic > By ad length table

Detection rate by ad position

Same shape as the ad-length heatmap, but columns are episode position (pre-roll / mid-roll / post-roll). A common pattern: pre-roll is easy because of clear show-intro transitions; post-roll is harder because models near the end of long episodes often produce shorter responses or run out of context to anchor on.

Source data: Detection rate by ad characteristic > By ad position table

Parser stress (extraction-method usage)

Heatmap of model (row) vs extraction-method (column), cell = number of responses parsed via that method. Columns are ordered by total usage. json_array_direct is the clean path; everything else is a recovery path the parser had to take because the model added markdown fences, wrapped the array in an object, or returned malformed JSON. Models near the top of the chart use the clean path most often. They are operationally easier to consume.

Source data: Parser stress test table

Failures and provider issues

No unresolved call errors across this run. Every (model, episode, trial, window) tuple ended with a parseable response.

Precision, recall, and FP/FN breakdown

F1 collapses two failure modes into one number. A precision-leaning model misses ads but rarely flags non-ads; a recall-leaning model catches everything at the cost of false positives. Production tradeoffs hinge on which one you can tolerate.

Column key

Column	Meaning	Range
TP (true positive)	Predicted an ad and a real ad existed at that span (IoU >= 0.5)	0 to total truth ads
FP (false positive)	Predicted an ad where no real ad existed	0 to total predictions
FN (false negative)	Missed a real ad entirely (no prediction matched it at IoU >= 0.5)	0 to total truth ads
Precision	TP / (TP + FP). Of the ads the model claimed, how many were real? Higher means fewer false positives.	0.000 to 1.000
Recall	TP / (TP + FN). Of the real ads, how many did the model find? Higher means fewer misses.	0.000 to 1.000

Reading the table: high precision + low recall means the model is cautious. It rarely flags something that isn’t an ad, but misses real ads. High recall + low precision means the opposite: catches everything but invents false positives. F1 is the harmonic mean of the two and rewards models that do both well.

Model	Precision	Recall	TP	FP	FN
claude-haiku-4-5-20251001	0.900	0.937	230	27	15

Boundary accuracy

For ads that match the truth at IoU >= 0.5, how far off were the predicted start and end timestamps? Lower is better. A model can hit F1 cleanly while still being 20s off on every boundary. Bad for any pipeline that cuts the audio.

MAE is size of the miss; bias is its direction (mean of predicted minus truth). A negative start bias or positive end bias means the cut extends past the ad and eats surrounding content; the opposite signs mean ad audio is left in. MinusPod cuts what the model flags, so a model whose bias points outward over-cuts even when its MAE looks acceptable. Bias near zero with a large MAE means the misses are random rather than systematic.

Model	Start MAE (s)	End MAE (s)	Start bias (s)	End bias (s)
claude-haiku-4-5-20251001	3.63	4.50	+0.87	-1.69

Confidence calibration

Models include a self-reported confidence on each detected ad. A well-calibrated model should be right ~95% of the time when it claims 0.95 confidence. The table below bins each model’s predictions and shows the actual hit rate (fraction that were true positives at IoU >= 0.5). A bin near 1.0 is well-calibrated; a low number with a high count means the model is overconfident.

Model	0.00-0.70	0.70-0.90	0.90-0.95	0.95-0.99	0.99+	total
claude-haiku-4-5-20251001	0.00 (n=1)	0.43 (n=7)	0.54 (n=13)	0.95 (n=171)	0.89 (n=65)	257

See report_assets/calibration.svg for the visual reliability diagram.

Latency tail

Median latency hides outliers. p99 and max are what determines queue depth and worst-case user wait. For OpenRouter-routed models the tail also reflects upstream provider load, not just model compute.

Model	p50	p90	p95	p99	max
claude-haiku-4-5-20251001	16.01s	69.51s	85.52s	226.93s	452.07s

Output token efficiency

How many output tokens the model spent per detected ad. Lower is more concise (the model finds an ad and returns the JSON). Higher means the model is producing a lot of text the parser will discard, which costs you whether or not the answer is right.

Model	Total output tokens	Ads detected	Tokens / ad	Cost / TP
claude-haiku-4-5-20251001	47,158	784	60	$0.0047

Cost breakdown (input vs output)

Where each model’s per-episode dollars go, at the same pricing snapshot as every other table. Every model reads the same transcripts, so the input side varies only with the provider’s input price. The output side varies with how much the model writes: a high output share on a modest total usually means reasoning tokens, and a model with a low per-token price can still land mid-table by writing thousands of them. Failed calls are excluded, same as the cost column everywhere else.

Model	Cost / episode	Input	Output	Output share
claude-haiku-4-5-20251001	$0.0773	$0.0739	$0.0034	4%

Trial variance (determinism check)

All trials run at temperature 0.0. If a model produces stable output you’d expect the F1 stdev across trials to be near zero. Higher numbers mean the model is non-deterministic even at temp=0. That’s fine to know, but means you cannot trust a single trial’s number for that model. n/a means the row has fewer than 2 trials on record, so a stdev is undefined rather than zero.

Model	Mean F1 stdev across episodes	Highest single-episode stdev
claude-haiku-4-5-20251001	0.0686	0.2236

Cross-model agreement

For each of the 171 (episode, window, trial-equivalent) entries, how many of the 1 active models predicted at least one ad? High-agreement windows are unambiguous ads (or unambiguously not ads). Low-agreement windows are where individual models disagree, and are candidates for ensemble voting if you want a cheap accuracy boost.

Models predicting an ad	Window count	Share
0 of 1	97	56.7%
1 of 1	74	43.3%

Read this as: rows near the top are windows where the field disagrees (most models said no, a few said yes, usually false positives); rows near the bottom are windows where the field broadly agrees (typical of clear sponsor reads).

Per-model alignment with consensus

Same data, viewed per model. For each window, the majority is whether more than half of the 1 active models flagged an ad. Then for each model: did it vote with the majority or against it? Four buckets:

• with-yes: this model voted yes, majority also voted yes (likely true positive)
• with-no: this model voted no, majority also voted no (likely true negative)
• broke-yes: this model voted yes, majority voted no (likely false positive / hallucination)
• broke-no: this model voted no, majority voted yes (likely missed real ad)

Alignment rate is (with-yes + with-no) / total. High alignment means the model tracks the consensus; low alignment means it disagrees often, which could be brilliance or noise depending on whether its disagreements are also where its F1 wins or loses.

Model	with-yes	with-no	broke-yes	broke-no	Alignment
claude-haiku-4-5-20251001	74	97	0	0	100.0%

Detection rate by ad characteristic

Aggregate detection rates often hide systematic blind spots. Below: for each model, what fraction of truth ads in each bucket were detected (matched at IoU >= 0.5).

By ad length

Truth ads bucketed by duration: short (<30s), medium (30-90s), long (>=90s). Cell values are detection rate (fraction of truth ads in that bucket the model caught), with the sample size n so a misleading 1.00 on a 2-ad bucket doesn’t get over-weighted. Models that systematically miss short ads usually fail on network-inserted brand-tagline spots; missing long ads is rarer and usually means the model gave up before processing the full window.

Model	long (>=90s)	medium (30-90s)	short (<30s)
claude-haiku-4-5-20251001	0.94 (n=140)	0.95 (n=75)	0.90 (n=30)

By ad position

Truth ads bucketed by where they fall in the episode: pre-roll (first 10%), mid-roll (10-90%), post-roll (last 10%). Cell values are the same detection-rate-with-n format as ad length. A common failure pattern in our data: most models detect pre-roll and mid-roll reliably and miss post-roll, because the prompt windows near the end often catch the model mid-reasoning or with fewer transition phrases to anchor on.

Model	pre-roll (<10%)	mid-roll (10-90%)	post-roll (>90%)
claude-haiku-4-5-20251001	0.97 (n=70)	0.96 (n=125)	0.84 (n=50)

Quick Comparison

One row per model, one column per episode. The headline columns (F1, Cost/ep, p50) summarize across all episodes; the per-episode columns let you see whether a model’s average hides wide swings (a model that scores well overall might still bomb on a specific genre). The right-most F1 stdev column averages the per-trial standard deviations across episodes; high values mean the model isn’t deterministic at temperature 0.0, so its single-trial F1 number is noisy. Moderation blocked is the share of attempted calls the provider refused on content grounds; those windows never reach scoring, so any non-zero value means that row’s F1 was computed on a subset of the corpus and is not comparable to a row at -.

Model	F1	Cost/ep	p50	ep-crime-junkie-8ce498f299d7	ep-daily-gist-chicago-70a82fe93a5c	ep-daily-tech-news-show-b576979e1fe8	ep-daily-tech-news-show-c1904b8605f7	ep-drink-champs-30c9a2d49f13	ep-glt1412515089-373d5ba5007b	ep-it-s-a-thing-e339179dfad6	ep-on-air-with-dan-and-alex2-574e4f303730	ep-security-now-audio-2850b24903b2	ep-the-brilliant-idiots-0bb9bf634c8e	ep-the-tim-dillon-show-f62bd5fa1cfe	ep-tosh-show-5f6894439bb6	ep-ai-cloud-essentials-e8dc897fbd6b (no-ad)	ep-oxide-and-friends-ce789ff5b62e (no-ad)	F1 stdev	Moderation blocked
claude-haiku-4-5-20251001	0.915	$0.0773	16.0s	0.950	0.900	1.000	0.687	1.000	0.821	1.000	0.880	0.865	0.876	1.000	1.000	PASS	PASS	0.069	-

---

Detailed Results

Per-Model Detail

Full per-model profile: F1 averaged across episodes, total cost per episode at current pricing, p50 / p95 latency, JSON compliance, parse-failure rate, the distribution of extraction methods the parser had to use, and verbosity / truncation telemetry. The Extraction methods list shows how often each route was hit. json_array_direct is the cleanest; the rest are recovery paths. The verbosity row flags models that emit long reason fields or run out of token budget mid-response. Ordered by F1 descending so the best performers appear first.

claude-haiku-4-5-20251001

• F1 (avg across episodes): 0.915
• Total cost / episode: $0.0773
• p50 / p95 latency: 16.01s / 85.52s
• JSON compliance: 1.00
• JSON mode: native (100% native, 855 calls)
• Parse failure rate: 0.0%
• Extraction methods: segment_id_direct: 855
• Verbosity: 0/855 calls over 1024 output tokens (0.0%); 0 hit max_tokens (0.0%); 0 salvaged from truncated JSON (0.0%)
• Segment category named on 784/784 detections (100%); the rest stay uncategorized (resolver: production)

Per-Episode Detail

One subsection per episode in the corpus, showing how every model performed on that specific episode. For ad-bearing episodes you see F1 and the stdev across trials (low stdev means stable, high stdev means the model’s number on this episode is noisy). For the no-ad episode you see PASS / FAIL on the negative control: PASS = zero false positives across all windows, FAIL = the model flagged something that wasn’t an ad, with the count.

ep-ai-cloud-essentials-e8dc897fbd6b: How Physical AI is Streamlining Engineering

• Podcast: ai-cloud-essentials
• Duration: 16.4 min
• Truth: no-ads episode

Model	Result	FP count
claude-haiku-4-5-20251001	PASS	0

ep-crime-junkie-8ce498f299d7: MISSING: Christopher “Cole” Thomas

• Podcast: crime-junkie
• Duration: 48.2 min
• Truth ads: 4

Model	F1	F1 stdev
claude-haiku-4-5-20251001	0.950	0.112

ep-daily-gist-chicago-70a82fe93a5c: Suburban apartment market heats up

• Podcast: daily-gist-chicago
• Duration: 21.2 min
• Truth ads: 2

Model	F1	F1 stdev
claude-haiku-4-5-20251001	0.900	0.224

ep-daily-tech-news-show-b576979e1fe8: Motorola Razr Fold is a Noble Competitor to the Galaxy Z Fold 7 - DTNS 5269

• Podcast: daily-tech-news-show
• Duration: 34.6 min
• Truth ads: 4

Model	F1	F1 stdev
claude-haiku-4-5-20251001	1.000	0.000

ep-daily-tech-news-show-c1904b8605f7: Switch 2 Prices Rise, Forecast Drops - DTNS 5265

• Podcast: daily-tech-news-show
• Duration: 38.6 min
• Truth ads: 5

Model	F1	F1 stdev
claude-haiku-4-5-20251001	0.687	0.064

ep-drink-champs-30c9a2d49f13: Episode 501 w/ Warren Sapp

• Podcast: drink-champs
• Duration: 258.6 min
• Truth ads: 9

Model	F1	F1 stdev
claude-haiku-4-5-20251001	1.000	0.000

ep-glt1412515089-373d5ba5007b: #2496 - Julia Mossbridge

• Podcast: glt1412515089
• Duration: 165.3 min
• Truth ads: 4

Model	F1	F1 stdev
claude-haiku-4-5-20251001	0.821	0.110

ep-it-s-a-thing-e339179dfad6: SOUP shots - It’s a Thing 418

• Podcast: it-s-a-thing
• Duration: 26.7 min
• Truth ads: 2

Model	F1	F1 stdev
claude-haiku-4-5-20251001	1.000	0.000

ep-on-air-with-dan-and-alex2-574e4f303730: Ryanair Wants Alcohol Bans, Emirates’ $6.8B Record Profit & Buying Spirit Airlines?!

• Podcast: on-air-with-dan-and-alex2
• Duration: 58.1 min
• Truth ads: 2

Model	F1	F1 stdev
claude-haiku-4-5-20251001	0.880	0.110

ep-oxide-and-friends-ce789ff5b62e: Mechanical Engineering at Oxide [chapter images]

• Podcast: oxide-and-friends
• Duration: 84.5 min
• Truth: no-ads episode

Model	Result	FP count
claude-haiku-4-5-20251001	PASS	0

ep-security-now-audio-2850b24903b2: SN 1077: A Browser AI API? - End of Bug Bounties?

• Podcast: security-now-audio
• Duration: 156.2 min
• Truth ads: 7

Model	F1	F1 stdev
claude-haiku-4-5-20251001	0.865	0.067

ep-the-brilliant-idiots-0bb9bf634c8e: Class Rank

• Podcast: the-brilliant-idiots
• Duration: 119.9 min
• Truth ads: 3

Model	F1	F1 stdev
claude-haiku-4-5-20251001	0.876	0.137

ep-the-tim-dillon-show-f62bd5fa1cfe: 495 - Hantavirus Cruise & iPad Babies

• Podcast: the-tim-dillon-show
• Duration: 80.1 min
• Truth ads: 6

Model	F1	F1 stdev
claude-haiku-4-5-20251001	1.000	0.000

ep-tosh-show-5f6894439bb6: My Mom - Emergency Pod

• Podcast: tosh-show
• Duration: 41.4 min
• Truth ads: 5

Model	F1	F1 stdev
claude-haiku-4-5-20251001	1.000	0.000

Parser stress test

How each model’s responses were actually parsed. Columns are extraction methods, ordered alphabetically; rows are models, sorted by parse-failure rate (cleanest at top). json_array_direct is the happy path: a bare JSON array we could json.loads and process immediately. markdown_code_block means we had to strip triple-backtick fences first; json_object_* means the model wrapped the array in an outer object and we had to find the array key; regex_* are last-resort recovery paths. A model that needs anything but json_array_direct for most calls is fragile. It works today, but a small prompt change can break the parser.

Model	segment_id_direct
claude-haiku-4-5-20251001	855

Methodology

Reproducibility settings used for this run. The benchmark sends the same prompts MinusPod sends in production (same system prompt, same sponsor list, same windowing) so the F1 numbers here are directly relevant to production accuracy decisions. Cost is recomputed at report time from token counts against the active pricing snapshot, so all rows compare at the same prices regardless of when the actual call ran.

• Trials per (model, episode): 5, temperature 0.0
• max_tokens: 4096 (matches MinusPod production)
• response_format: json_object (with prompt-injection fallback when provider rejects native)
• Window size: 10 min, overlap: 3 min (imported from MinusPod’s create_windows)
• Pricing snapshot: 2026-08-23T20:34:27.481692Z
• Corpus episodes: 14

Transcript source

segments.json for every corpus episode is pulled byte-exact from the source MinusPod instance’s original-segments endpoint. The transcript itself was generated by faster-whisper inside that instance, not by the benchmark. Model choice and decoding params affect what gets transcribed, which sets an upper bound on what every benchmarked LLM can find.

Whisper config:

Setting	Value
Model	large-v3
Backend	local (faster-whisper, CUDA GPU)
Compute type	auto (resolves to float16 on CUDA)
Language	en (forced English, not auto-detect)
VAD gap detection	on (start 3.0s / mid 8.0s / tail 3.0s)

model.transcribe() invocation (src/transcriber.py as of corpus transcription, before 2.96.0):

WhisperModel(model_size="large-v3", device="cuda", compute_type="auto")
model.transcribe(
    audio,
    language="en",
    initial_prompt=<podcast name + SEED_SPONSORS vocabulary>,
    beam_size=5,
    batch_size=<adaptive: 16/12/8/4 by episode length>,
    word_timestamps=True,
    vad_filter=True,
    vad_parameters={"min_silence_duration_ms": 1000, "speech_pad_ms": 600, "threshold": 0.3},
)

The initial_prompt seeded a sponsor vocabulary so Whisper produced consistent spellings (Athletic Greens rather than AG1, ExpressVPN rather than express vpn), biasing what showed up in the transcript and therefore what every benchmarked LLM is scored against. 2.96.0 removed the prompt because it dropped 7-17% of each episode’s speech on the batched pipeline; the corpus has not been re-transcribed without it.

Sponsor vocabulary (254 canonical sponsors, 44 of them with explicit alias spellings totaling 48 aliases; from src/utils/constants.py SEED_SPONSORS). Laid out in two side-by-side groups, read top-to-bottom in each group.

Sponsor	Aliases	Category	Sponsor	Aliases	Category
1Password	One Password	tech	MacPaw	CleanMyMac	tech
Acorns	-	finance	Magic Mind	-	beverage
ADT	-	home	Magic Spoon	-	food
Affirm	-	finance_fintech	Mailchimp	-	tech_software_saas
Airbnb	-	travel_hospitality	Manscaped	-	personal
Airtable	-	tech_software_saas	MasterClass	Master Class	education
Alani Nu	-	food_beverage_nutrition	McDonald’s	-	food_beverage_nutrition
Allbirds	-	ecommerce_retail_dtc	Mercury	-	finance_fintech
Alo Yoga	-	ecommerce_retail_dtc	Meter	-	b2b_startup
Amazon	-	retail	Midjourney	-	tech_software_saas
Anthropic	-	tech_software_saas	Mint Mobile	MintMobile	telecom
Apple TV+	-	media_streaming	Miro	-	tech
Asana	-	tech_software_saas	Momentous	-	mental_health_wellness
AT&T	-	telecom	Monarch Money	-	finance
Athletic Brewing	-	beverage	Monday.com	Monday	tech
Athletic Greens	AG1, AG One	health	Native	-	personal
Audible	-	entertainment	NerdWallet	-	finance_fintech
Aura	-	tech	Netflix	-	media_streaming
Babbel	-	education	NetSuite	Net Suite	tech
BetMGM	Bet MGM	gambling	Noom	-	mental_health_wellness
BetterHelp	Better Help	health	NordVPN	Nord VPN	vpn
Betterment	-	finance	Notion	-	tech
Bill.com	-	finance_fintech	Nutrafol	-	health
Birchbox	-	ecommerce_retail_dtc	Okta	-	tech_software_saas
Bitwarden	Bit Warden	tech	OLIPOP	-	food_beverage_nutrition
Blinkist	-	education	OneSkin	One Skin	personal
Bloom Nutrition	-	food_beverage_nutrition	OpenAI	-	tech_software_saas
Blue Apron	-	food	Outdoor Voices	-	ecommerce_retail_dtc
Bombas	-	apparel	OutSystems	-	tech
Booking.com	-	travel_hospitality	PagerDuty	-	b2b_startup
Bose	-	electronics	Paramount+	-	media_streaming
Brex	-	finance_fintech	Patreon	-	tech_software_saas
Brilliant	-	tech_software_saas	Perplexity	-	tech_software_saas
Brooklinen	-	home	Plaid	-	finance_fintech
Butcher Box	ButcherBox	food	PolicyGenius	Policy Genius	finance
CacheFly	-	tech	Poppi	-	food_beverage_nutrition
Caesars Sportsbook	-	gaming_sports_betting	Poshmark	-	ecommerce_retail_dtc
Calm	-	health	Progressive	-	finance
Canva	-	tech	Public.com	-	finance_fintech
Capital One	-	finance	Pura	-	home_security
Care/of	Care of, Careof	health	Purple	-	home
CarMax	Car Max	auto	QuickBooks	-	finance_fintech
Carvana	-	auto	Quince	-	apparel
Casper	-	home	Quip	-	personal
Cerebral	-	mental_health_wellness	Ramp	-	finance_fintech
Chime	-	finance_fintech	Raycon	-	electronics
ClickUp	-	tech_software_saas	Retool	-	tech_software_saas
Cloudflare	-	tech_software_saas	Ring	-	home
Coinbase	-	finance_fintech	Rippling	-	b2b_startup
Comcast	-	telecom	Ritual	-	health
Cozy Earth	-	home	Ro	-	mental_health_wellness
Credit Karma	-	finance	Robinhood	-	finance_fintech
CrowdStrike	-	tech_software_saas	Rocket Lawyer	-	insurance_legal
Cursor	-	tech_software_saas	Rocket Money	RocketMoney, Truebill	finance
Databricks	-	tech_software_saas	Roman	-	health
Datadog	-	tech_software_saas	Rosetta Stone	-	education
Deel	-	business	Rothy’s	-	ecommerce_retail_dtc
DeleteMe	Delete Me	tech	Saatva	-	ecommerce_retail_dtc
Disney+	-	media_streaming	Salesforce	-	tech_software_saas
DocuSign	-	tech_software_saas	SeatGeek	-	gaming_sports_betting
Dollar Shave Club	DSC	personal	Seed	-	health
DoorDash	Door Dash	food	SendGrid	-	tech_software_saas
DraftKings	Draft Kings	gambling	ServiceNow	-	tech_software_saas
Duolingo	-	tech_software_saas	Shein	-	ecommerce_retail_dtc
eBay Motors	-	auto	Shopify	-	tech
Eight Sleep	-	mental_health_wellness	SimpliSafe	Simpli Safe	home
ElevenLabs	-	tech_software_saas	SiriusXM	-	media_streaming
ESPN Bet	-	gaming_sports_betting	Skillshare	-	tech_software_saas
Everlane	-	ecommerce_retail_dtc	SKIMS	-	ecommerce_retail_dtc
EveryPlate	-	food_beverage_nutrition	Skyscanner	-	travel_hospitality
Expedia	-	travel_hospitality	Slack	-	tech_software_saas
ExpressVPN	Express VPN	vpn	Snowflake	-	tech_software_saas
FabFitFun	-	ecommerce_retail_dtc	SoFi	-	finance
Factor	-	food	Spaceship	-	tech
FanDuel	Fan Duel	gambling	Splunk	-	b2b_startup
Figma	-	tech_software_saas	Spotify	-	media_streaming
Ford	-	auto	Squarespace	Square Space	tech
Framer	-	tech	Stamps.com	Stamps	business
FreshBooks	-	finance_fintech	Starbucks	-	food_beverage_nutrition
Function Health	-	mental_health_wellness	State Farm	-	finance
Function of Beauty	-	personal	Stitch Fix	-	ecommerce_retail_dtc
Gametime	Game Time	entertainment	StockX	-	ecommerce_retail_dtc
Geico	-	finance	Stripe	-	finance_fintech
GitHub	-	tech_software_saas	StubHub	-	gaming_sports_betting
GitHub Copilot	-	tech_software_saas	Substack	-	tech_software_saas
GOAT	-	ecommerce_retail_dtc	T-Mobile	TMobile	telecom
GoodRx	Good Rx	health	Talkspace	-	mental_health_wellness
Gopuff	-	ecommerce_retail_dtc	Temu	-	ecommerce_retail_dtc
Grammarly	-	tech	Ten Thousand	-	ecommerce_retail_dtc
Green Chef	GreenChef	food	Thinkst Canary	-	tech
Grubhub	Grub Hub	food	Thorne	-	mental_health_wellness
Gusto	-	b2b_startup	ThreatLocker	-	tech
Harry’s	Harrys	personal	ThredUp	-	ecommerce_retail_dtc
HBO Max	-	media_streaming	Thrive Market	-	food
Headspace	Head Space	health	Toyota	-	auto
Helix Sleep	Helix	home	Transparent Labs	-	food_beverage_nutrition
HelloFresh	Hello Fresh	food	Turo	-	automotive_transport
Hers	-	health	Twilio	-	tech_software_saas
Hims	-	health	Uber	-	automotive_transport
Honeylove	Honey Love	apparel	Uber Eats	UberEats	food
Hopper	-	travel_hospitality	UnitedHealth Group	-	finance_fintech
HubSpot	Hub Spot	tech	Vanta	-	tech
Huel	-	food_beverage_nutrition	Veeam	-	tech
Hyundai	-	auto	Vercel	-	tech_software_saas
iHeartRadio	-	media_streaming	Verizon	-	telecom
Imperfect Foods	-	food_beverage_nutrition	Visible	-	telecom
Incogni	-	tech	Vrbo	-	travel_hospitality
Indeed	-	jobs	Vuori	-	ecommerce_retail_dtc
Inside Tracker	-	mental_health_wellness	Warby Parker	-	ecommerce_retail_dtc
Instacart	-	food	Wayfair	-	ecommerce_retail_dtc
Intuit	-	finance_fintech	Waymo	-	automotive_transport
Joovv	-	mental_health_wellness	Wealthfront	-	finance
Kayak	-	travel_hospitality	WebBank	-	finance_fintech
Klarna	-	finance_fintech	Webflow	-	b2b_startup
Klaviyo	-	tech_software_saas	WhatsApp	-	tech
LegalZoom	-	insurance_legal	WHOOP	-	mental_health_wellness
Lemonade	-	finance	Workday	-	tech_software_saas
Levels	-	mental_health_wellness	Xero	-	finance_fintech
Liberty Mutual	-	finance	YouTube	-	media_streaming
Lime	-	automotive_transport	YouTube TV	-	media_streaming
Linear	-	tech_software_saas	Zapier	-	tech
LinkedIn	LinkedIn Jobs	jobs	Zendesk	-	tech_software_saas
Liquid IV	Liquid I.V.	health	ZipRecruiter	Zip Recruiter	jobs
LMNT	Element	health	ZocDoc	Zoc Doc	health
Loom	-	tech_software_saas	Zoom	-	tech_software_saas
Lululemon	-	ecommerce_retail_dtc	Zscaler	-	tech
Lyft	-	automotive_transport	Zyn	ZYN, Zinn	tobacco_nicotine

Mishearing corrections (174 entries, from src/utils/constants.py SPONSOR_ALIASES). Applied post-transcription to normalize Whisper output toward the canonical sponsor name. Distinct from the aliases column above, which lists intentional alternative spellings (e.g. AG1 vs Athletic Greens); the entries below are mostly Whisper mishearings (e.g. a firm -> Affirm, xerox -> Xero). Laid out in three side-by-side groups, read top-to-bottom in each group.

Heard as	Normalized to	Heard as	Normalized to	Heard as	Normalized to
1 password	1Password	good-rx	GoodRx	patron	Patreon
8 sleep	Eight Sleep	green chef	Green Chef	pay tree on	Patreon
8-sleep	Eight Sleep	green-chef	Green Chef	perplexity ai	Perplexity
a firm	Affirm	greenchef	Green Chef	perplexity-ai	Perplexity
a g one	Athletic Greens	grub hub	Grubhub	policy genius	PolicyGenius
ag 1	Athletic Greens	grub-hub	Grubhub	policy-genius	PolicyGenius
ag one	Athletic Greens	harrys	Harry’s	pyura	Pura
ag1	Athletic Greens	head space	Headspace	ray con	Raycon
athlean x	Athlean-X	head-space	Headspace	ray-con	Raycon
athlean-x	Athlean-X	hello fresh	HelloFresh	re tool	Retool
athletic greens one	Athletic Greens	hello-fresh	HelloFresh	ro gain	Rogaine
athleticgreens	Athletic Greens	him's	Hims	ro-gaine	Rogaine
bet mgm	BetMGM	hims & hers	Hims & Hers	rocket money	Rocket Money
bet-mgm	BetMGM	hims and hers	Hims & Hers	rocket-money	Rocket Money
better help	BetterHelp	honey love	Honeylove	rocketlawyer	Rocket Lawyer
better-help	BetterHelp	honey-love	Honeylove	rocketmoney	Rocket Money
birch box	Birchbox	honeylove	Honeylove	rocketmortgage	Rocket Mortgage
birch-box	Birchbox	hub spot	HubSpot	seat geek	SeatGeek
bit warden	Bitwarden	hub-spot	HubSpot	seat-geek	SeatGeek
bit-warden	Bitwarden	hubs pot	HubSpot	shop a fly	Shopify
blueapron	Blue Apron	imperfect foods	Imperfect Foods	shop fly	Shopify
brecks	Brex	imperfectfoods	Imperfect Foods	shop ify	Shopify
butcher box	Butcher Box	insta cart	Instacart	simpli safe	SimpliSafe
butcher-box	Butcher Box	insta-cart	Instacart	simpli-safe	SimpliSafe
butcherbox	Butcher Box	l m n t	LMNT	simply safe	SimpliSafe
car max	CarMax	legal zoom	LegalZoom	sky scanner	Skyscanner
car-max	CarMax	legal-zoom	LegalZoom	sky-scanner	Skyscanner
cloud flare	Cloudflare	legalzoom	LegalZoom	so fi	SoFi
cloud-flare	Cloudflare	liquid i v	Liquid IV	so-fi	SoFi
co pilot	GitHub Copilot	liquid i.v.	Liquid IV	square space	Squarespace
co-pilot	GitHub Copilot	liquid iv	Liquid IV	square-space	Squarespace
copilot	GitHub Copilot	liquidiv	Liquid IV	stamp dot com	Stamps.com
creditkarma	Credit Karma	magic mind	Magic Mind	stitch fix	Stitch Fix
delete me	DeleteMe	magic spoon	Magic Spoon	stitch-fix	Stitch Fix
delete-me	DeleteMe	magicmind	Magic Mind	stitchfix	Stitch Fix
dollarshaveclub	Dollar Shave Club	magicspoon	Magic Spoon	stub hub	StubHub
door dash	DoorDash	master class	MasterClass	stub-hub	StubHub
door-dash	DoorDash	master-class	MasterClass	sub stack	Substack
draft kings	DraftKings	mercury bank	Mercury	sub-stack	Substack
draft-kings	DraftKings	mercury-bank	Mercury	thrive market	Thrive Market
eight-sleep	Eight Sleep	mint mobile	Mint Mobile	thrivemarket	Thrive Market
eightsleep	Eight Sleep	mint-mobile	Mint Mobile	transparent labs	Transparent Labs
element	LMNT	mintmobile	Mint Mobile	transparentlabs	Transparent Labs
every plate	EveryPlate	monarch money	Monarch Money	uber eats	Uber Eats
every-plate	EveryPlate	monarch-money	Monarch Money	uber-eats	Uber Eats
express vpn	ExpressVPN	monarchmoney	Monarch Money	ubereats	Uber Eats
express-vpn	ExpressVPN	my protein	Myprotein	ver cell	Vercel
fab fit fun	FabFitFun	my ro	Miro	ver sel	Vercel
fab-fit-fun	FabFitFun	myprotein	Myprotein	wealth front	Wealthfront
fan duel	FanDuel	net suite	NetSuite	wealth-front	Wealthfront
fan-duel	FanDuel	net-suite	NetSuite	woop	Whoop
game time	Gametime	nord vpn	NordVPN	xerox	Xero
game-time	Gametime	nord-vpn	NordVPN	zero	Xero
gametime	Gametime	one password	1Password	zip recruiter	ZipRecruiter
github-copilot	GitHub Copilot	one skin	OneSkin	zip-recruiter	ZipRecruiter
go puff	Gopuff	one-password	1Password	zoc doc	ZocDoc
go-puff	Gopuff	one-skin	OneSkin	zoc-doc	ZocDoc
good rx	GoodRx	p ninety x	P90X	zock doc	ZocDoc

Run Metadata

• Report generated: 2026-09-19T06:00:44Z
• Unique work units (current state, last-write-wins after retries): 855
• Successful: 855
• Failed: 0
• Lifetime list-price cost (sum of at-runtime costs, includes superseded rows): $5.4088
• Lifetime tokens (same basis): 5,173,055 in + 47,158 out = 5,220,213
• Note: every input token is priced at list rate. Providers that serve a repeated prompt from cache bill less than this, and the harness does not record cache hits, so a real invoice for this run will come in under the figure above.
• Active pricing snapshot: 2026-08-23T20:34:27.481692Z
• Addressing mode: segment_ids
• System prompt: live (sha256:7652e74a)