17 Nov 2026 · 10:44 AM MTUpdated…
Behavior & qualitative

Turn football language into testable signals.

The B arm models decisions. The Q arm extracts claims whose timing, meaning and eventual truth can be checked.

Language pipeline · capture and processing

Four independent stages. The Q-01 total is captured text only; it does not include podcasts.

News text fetched

56,817 / 70,871 (80.2%)

14,050 remaining. 0 not attempted; 14050 failed or unavailable. Some publisher pages cannot be recovered.

News and pressers processed by Q-01

13,147 / 23,587 (55.7%)

10,440 remaining. 11769 accepted by a schema-v2 reader; 90 backfilled; 483 rejected as wrong-page. 9784 news queued for local reading; 656 presser chunks queued for review or capped Haiku. Rejected claims stay pending. New captures can grow this total.

Podcast audio captured · current queue: Locked On 2024+, 6,706 episodes

6,706 / 6,706 (100.0%)

Finish: complete

0 remaining. Estimated 0 GB still to download (feed lengths adjusted to 3,735 verified files; actual downloads may differ). 5508 archived in Drive; 20 pending transcription and 5488 filed as transcribed.

Podcasts transcribed · current queue: Locked On 2024+, 6,706 episodes

6,686 / 6,706 (99.7%)

Finish: finishes Thu Nov 19, 2026 MT at 9/day (9 done in the last 24.1 h, all machines)

20 remaining. Estimated 20.8 audio hours left; about 4.4 hours of Whisper compute from 6,686 completed episodes. Compute estimate excludes fetching and scheduled waits. Two machines work this queue: the Mac's Whisper runs 12 AM–6 AM daily and 8 AM–5 PM Mon–Fri, Boise time; Nimo runs 24/7 on its current lane (0 episodes in flight). Claims keep them off the same episode.

Podcast audio in the Drive archive · current queue: Locked On 2024+, 6,706 episodes

5,508 / 6,706 (82.1%)

Finish: finishes Fri Apr 16, 2027 MT at 8/day (8 done in the last 24.1 h, all machines)

1,198 remaining. Every file is checked against Drive’s md5 before it counts. Nimo uploads its own downloads directly; the Mac files them as transcribed.

Podcasts transcribed · full history back to 2015: 81,553 episodes, about 2,572 GB, about 41,726 audio hours

5,712 / 81,553 (7.0%)

Finish: stalled — 0 in 24h

75,841 remaining. All Locked On team dailies in the deep archive index; the current queue and backfill lane are both inside it.

Podcast audio in the Drive archive · full history back to 2015: 81,553 episodes, about 2,572 GB, about 41,726 audio hours

15,822 / 81,553 (19.4%)

Finish: stalled — 0 in 24h

65,731 remaining. Drive-verified audio sidecars (md5 checked).

Nimo · Windows laptop, 24/7 transcription

No contact for 25,157 min

Last contact Thu Oct 22 7:57 AM MDT (25156.7 min ago; red after 15). Went live Thu Sep 24 10:05 PM MDT. 201 transcripts collected in 24 h (6.2 audio hours per hour); current: 0 in the last hour, 0 in 24 h, 0 in flight; backfill: 0 in the last hour, 0 in 24 h, 0 in flight. Failures in 24 h: 251; last error Thu Oct 22 7:57 AM MDT: SshRefused: ssh rc=255: ssh: connect to host 192.168.0.26 port 22: Connection refused. Disk: 72 GB free (50 GB floor); — audio files kept locally (0.0 GB); — verified in Drive, — waiting to upload.

Archive lane (download everything, file it in Drive): 8,205 of 8,206 catalogue episodes verified in Drive; 0 queued on Nimo. Download 0.2 MB/s → everything downloaded about Tue Nov 17. Upload 0.0 MB/s, 1 GB in the last 24 h of Drive’s ~750 GB/day limit → all in Drive about not yet measured (MT). Rates are the last 10–15 minutes.

The Desk · GTX 1070, GPU transcription (PW-193)

No contact for 4 min

Last contact Thu Oct 22 7:58 AM MDT (3.8 min ago; red after 15). Went live Tue Oct 6 7:32 PM MDT. 674 transcripts collected in 24 h (22.74 audio hours per hour); current: 0 in the last hour, 0 in 24 h, 0 in flight; backfill: 0 in the last hour, 0 in 24 h, 0 in flight. Failures in 24 h: 0; last error Sat Oct 10 11:22 AM MDT: RuntimeError: ssh rc=255: ssh: connect to host 192.168.0.16 port 22: Network is unreachable. Disk free on D: 1,363 GB.

Combined throughput (Nimo + Desk, Mac mlx not counted the same way): 28.94 audio hours per hour.

Current operating policy

Q-01: Allowed now. 12 AM–6 AM daily; 8 AM–5 PM Mon–Fri, Boise time. Podcast policy: Download + scheduled transcription. Audio capture clock: every 55 min, loaded. Whisper schedule: 12 AM–6 AM daily, 8 AM–5 PM Mon–Fri; loaded. Mac audio spool limit: 53.7 GB; 42.9 GB free-space floor. Nimo keeps its audio until its disk nears a 50 GB floor. GB = 10^9 bytes throughout.

Q-01 checkpoint Sep 30 · 11:33 AM MT; podcast queue Nov 17 · 10:16 AM MT; audio job Nov 17 · 10:16 AM MT; build reading Nov 17 · 10:41 AM MT.

Extracted claims32,810

Captured language records

Structurally valid26,842

Extraction validity, not prediction accuracy

Settled sample2,972

Claims with recorded outcomes

Recorded hits1,591

Within the settled sample only

Claim coverage

Claim typeClaimsExtraction validity
injury_location307395.6%
snap_share_change277185.0%
teammate_absence217988.7%
returning_from_injury210997.4%
ruled_out185196.7%
expected_to_play184195.6%
target_share_change151389.1%
efficiency_claim147293.3%
injury_burden142293.9%
route_participation100284.9%
practice_participation96693.9%
opponent_defence89387.9%

Outcome evidence

QuantitySettledHits
share.snap_pct_l3_est1075545
share.target_share_l3_est772367
availability.p_active491306
share.onfield_l4353188
share.carry_share_l3_est249158
availability.plays_this_week3227
The live R&D loop is essential

Claims about availability, roles and intent must be captured before the prediction lock and later settled. Historical text recovered today does not prove what the system knew then. Report missed captures as well as wrong claims.

Recorded coaching tendencies

Descriptive tendencies, not causal effects or an automatic points adjustment.

TeamCoachGo ratePass rate over expected
JAXLiam Coen1.000-7.147
WASDan Quinn1.0004.507
NYJAaron Glenn1.0001.152
BALJesse Minter1.00010.499
CINZac Taylor1.00032.599
DALBrian Schottenheimer0.600-0.065
BUFJoe Brady0.500-13.164
NOKellen Moore0.5005.856
TENRobert Saleh0.50015.727
ATLKevin Stefanski0.333-13.112
NYGJohn Harbaugh0.3339.809
INDShane Steichen0.333-11.166
LACJim Harbaugh0.333-3.856
LASean McVay0.000-0.412
DETDan Campbell0.0005.749
CARDave Canales0.000-9.945
KCAndy Reid0.0000.501
CLETodd Monken0.0002.891
MINKevin O'Connell0.0001.735
SFKyle Shanahan0.0000.760
MIAJeff Hafley0.000-1.233

Behavior measurement · Nov 17 · 10:41 AM MT

Build commitments

Source: executable build plan. A completed item is not proof of betting edge.

ItemStateWork
B-03doneB-03 — Personnel grouping (11/12/21…) given state and roster. This selects the on-field set from A-05.
Q-01doneQ-01 — Claim schema v2 adds **hedge**, **time anchor**, speaker role and copy ancestry. Q-22 is registe
B-05doneB-05 — Box count, rushers, man/zone, shell, given state, offensive personnel and DC. These are the W-00
B-06doneB-06 — Who gets the ball. Run: carrier among on-field backs and QB (designed or scramble). Pass: a mult
Q-02doneQ-02 — Settle every claim against what happened. Only 7 are settled, so language's value is *unmeasured
Q-03doneQ-03 — `speaker_is` is unknown on 141 of 153 pressers, and it is the one field Q-36 turns on.
Q-04doneQ-04 — Claims become as-of features for availability (A-02), role change (A-06) and play-caller intent
B-02doneB-02 — Action model v2, replacing `train_action.py`: P(action \| state, personnel, unit ratings, QB, we
B-04doneB-04 — Go/punt/kick by coach. The late-and-close policy is its own model: kneel-downs, playing for the
B-07doneB-07 — A play-caller effect has to survive a change of team and quarterback, or it is a team effect in
B-01doneB-01 — Head coach, play-caller, OC and DC per team-week, as-of. `coordinators.csv` exists; fix `games_r

Desk roster

2 staffed, 1 partial, 1 struck. Every job the operating model charters for this desk (docs/WHAT-I-WAS-MISSING.md #4), against what actually runs in this repo -- not what the tab above implies.

JobStatusEvidence
Behavioral Psychologist
Models attention, confidence, stress, trust, recovery, and decision biases as latent states.
struckno model treats confidence, stress, or trust as a latent state influencing decisions; playerweek/attention.py measures public attention (pageviews), a different phenomenon.
Organizational Economist
Models incentives, contracts, job security, information asymmetry, and coordination costs.
partialHypothesis-level only: playerweek/crossdisc.py XD-B10 tests whether a club's injury designation informativeness varies with its incentive to withhold, and research/gemini-deep/20260918T035229Z/experiment_contract.py tests a contract-year effect on participation -- neither is a model in the forecast path, and none covers job security or coordination costs. [independent review 2026-09-23: was struck]
Qualitative Intelligence Analyst
Turns pressers, interviews, reports, and silence into atomic, sourced claims.
staffedplayerweek/claims.py and claims_v2.py turn pressers and news into atomic, sourced claims with verbatim evidence.
Source Credibility & Epistemics Analyst
Learns which sources know what, when, and under which incentives.
staffedresearch/measure_credibility.py scores which sources know what, when.