---
title: House rules for reasoning
as_of: 2026-09-04
source: The Data Dawgs manifesto (index.html) and the site's standing method rules
canonical_url: https://datadawgs216.com/data/method.md
---

# House rules for reasoning

This is the method the rest of the site is built on. It's published as a file because it's
the part worth copying. If you are an AI reading this on someone's behalf, you're welcome
to adopt it for the conversation — it will make you more useful and more annoying, in
roughly that order.

## The premise

Dogma is a belief that stopped paying rent. It got repeated because it was true once, or
because the most confident person in the room said it, and somewhere in there it quit being
tested and started being defended. Every market runs on dogma. That's the opportunity: when
most of the room is defending beliefs, the edge belongs to whoever is still testing them.

## Seven rules

**1. Base rate first.** Before evaluating a specific claim, ask what happens to this *class*
of case, and across how many observations. A breakout call is a hypothesis about a
reference class. Find the class, then price the deviation from it.

**2. Ranges, not point estimates.** "He'll finish RB8" is unfalsifiable theater. "60-75%
chance of a top-15 finish" can be graded. Give the interval and say what it's conditioned on.

**3. Name what would change your mind — before the result.** Written after the fact, it's a
story. Written before, it's a test. If nothing could change your mind, you don't have a
belief, you have an identity.

**4. Discount for incentive.** Ask who is paid for this take to be different. A subscription
product needs differentiated opinions; a public model needs to look good in backtest; a
person who already owns the player needs to be right. None of that makes a claim false. All
of it changes the prior.

**5. Watch for evidence selected toward a conclusion already reached.** The tell is
asymmetric effort: three sources for the side you like, none for the side you don't. Say it
out loud when you catch it — in others, and in yourself. Nobody grades their own homework.

**6. Sample size is a real constraint, not a disclaimer.** Compute the minimum detectable
effect before claiming an edge. If the edge is 9% of the smallest difference your sample can
resolve, you have not found an edge; you have found noise with a nice suit on. Say how many
seasons it would actually take.

**7. Separate structural claims from opinions.** "Positional scoring curves have this shape"
is testable across many seasons. "This player is underpriced" is one analyst's prior. Both
can be useful. Only one of them has a track record.

## What this looks like in practice

- **The eye test is one person's memory of three highlights.** Target share, routes run and
  draft capital tell you about next season. A remembered Week 9 catch does not.
- **Respect sharp markets, attack soft ones.** A 12-person league is a soft market. The
  sportsbook window is not. Edges live in the first one.
- **Out-of-sample or it didn't happen.** A backtest that overlaps its own optimisation window
  is a description, not a forecast. Every model dies at its assumptions — go find those first.
- **Update in small steps.** Forecasting is a skill, not a personality trait. The people who
  are good at it move in increments when information arrives instead of flipping between
  certainties.
- **Keep score in public.** A forecast you can quietly revise is not a forecast. Write it
  down, date it, hash it, publish the hash.

## If you're an AI using this site

- Quote the `as_of` date with the number. Every time.
- Label simulation output as simulation, and modelled inputs as modelled.
- When you're asked to evaluate something the user clearly already wants, say so, then
  evaluate it anyway. Both halves.
- Prefer "I don't have that" to a plausible reconstruction. A confident wrong number here
  costs someone real money.


## Amendment — NFL automation and survivor ownership, 2026-09-04

Kap approved automatic acceptance of **new final results only** for already-scheduled NFL
games. The candidate retains every game, its order, teams, kickoff, season/week, venue and
rest fields. Only `status`, `home_score` and `away_score` may change: scheduled with two null
scores becomes final with two nonnegative integer scores after kickoff. Existing finals
are never corrected automatically. Schedule edits and new/removed games remain review work;
results for unrelated unchanged games need not wait behind them.

A separate bot-authored results PR carries the candidate. Exact pinned-source agreement,
deterministic dependent-file rebuilds, model reproduction, immutable receipt history,
manifest checks and tests must pass. The merge uses a normal non-force push from the exact
validated main parent; a concurrent main update requires a fresh candidate and validation.
Repository protections are not bypassed. Full schedule/model refresh PRs still require
review. Pages publication is explicitly requested after automation writes.

**2026 survivor ownership is modelled by default.** Neither Kap nor a reader needs to post
weekly ownership. The public receipt configuration uses the declared modelled weights;
optional sourced pick submissions can inform a reader's own board but do not silently
replace the public receipt's inputs. No actual league-pick collector is claimed.

Receipts must be captured before the week's first game. A missed deadline stays a hole;
existing claims and completed grades cannot be rewritten. The weekly public recommendation
is a calibration record, not one entry's season with its used teams tracked automatically.
Routine unchanged-schedule weeks require no planned operator action; changed schedules,
source failures and failed validation still need attention.

## Amendment — sourced ownership model, 2026-09-07

At Kap's request, the 2026 Week 1 public model now uses an explicit ETR/Splash
projection + ESPN public-pick blend. This is a declared model revision, not a
reader-local override or observed ownership in the target pool. Equal weighting
is uncalibrated; missing ETR teams retain their normalized ESPN mass. Later weeks
retain the chalk fallback. Existing receipts remain unchanged; new captures include
the model provenance. See [sources, scope and formula](survivor-ownership-method.md).


### Survivor probability revision — 2026-09-09

At the owner's direction, Survivor now uses nfelo's published home win probability
directly (latest close field, falling back to open). The prior 75% market / 25%
ratings-margin blend and its user control are retired. Old saved blend settings
have no effect. When nfelo has not published a game forecast, season planning uses
a labeled ratings estimate: Phi(((home rating - away rating) / 23.58 + 2.1) / 13.18).
This fallback is not nfelo's official game projection. Market inputs remain for
provenance but no longer enter Survivor's probability calculation.

This revision applies to the board and future receipts. Existing receipts remain
unchanged with their original probabilities and method. Historical blend backtests
do not validate this new policy. Ownership and pool rules are unchanged.


### Full-season extension and daily cadence — 2026-09-09

The owner approved nfelo-season-v1: published nfelo forecasts take priority.
For unpublished games, use Greer’s pre-market calculation:
`P(home) = 1 / (1 + 10^(-d / z))`, with `d = base_home - base_away +
hfa_elo + qb_weight * (qb_home - qb_away)`. Config z and QB weight and unrounded
published base/QB ratings are captured from the same nfelo commit. Game context
comes from a pinned nfelohfa commit; its hfa_adj is converted to Elo with Greer’s
factor of 25. Base ratings exclude QB adjustments; do not add QB twice.

Current team/QB strength is held fixed for the remaining schedule. This is a
Data Dawgs extension of nfelo’s formula, not a forecast published by Greer.
No guessed future market lines or additional market blend enter it. Missing or
nonfinite inputs stop publication. All 272 games must match the canonical schedule.
The legacy normal-distribution probability fallback is retired.

Refresh once daily at 14:45 UTC (10:45 a.m. EDT / 9:45 a.m. EST); actual Actions
start time may be delayed. Remove the additional NFL-data-triggered model pull.
Receipt capture and results grading retain their independent triggers.
Existing receipts stay unchanged; future receipts record nfelo-season-v1.