Why this is short
The fantasy basketball encyclopedia this build mirrors ran five evidence grades crossed with four confidence labels crossed with four freshness states — thirteen terms to learn before making a single pick. A policy nobody can recall under a draft clock is not a policy; it is a document.
This one is three grades, two axes, four rules. It fits on an index card, which is the only form in which it will actually get used.
Evidence, three grades
Official. League and team transactions, filed injury reports, published depth charts, official statistics. This is the only grade that establishes identity, status, dates, and observed performance. If the question is did this happen, is he on the roster, is he listed out — the answer comes from here and from nowhere else. Official sources are narrow: they tell you what occurred, never what it means.
Reported. Named beat writers, and established projection services that publish a visible date on their work. This grade carries expectations and expert disagreement — who is expected to get the carries, what a role change is likely to be worth, where informed people differ. The two qualifiers are load-bearing. Named, because an unattributed report cannot be weighed or revisited. Visibly dated, because an undated projection is a projection of unknown vintage, and vintage is most of what a projection's reliability consists of.
Community. Identified social and forum analysis — a poster with a track record, a thread that assembles something real. This grade produces leads and hypotheses only. It can tell you a question exists and point at where to look. It is never a factual foundation. The correct use of a good community post is to go and confirm it at the Official or Reported grade; if it cannot be confirmed there, it does not enter a note as a fact.
Anonymous, undated, or unattributable material is not a fourth grade. It is not evidence.
Confidence, two axes
Every claim about a player reduces to two questions:
Role. Is he the guy? How many touches, how many targets? A target is a pass thrown in his direction, counted whether or not he catches it; a touch is a carry or a catch. Role is the volume question. See Fantasy football glossary.
Health. Will he play? This is a binary that resolves weekly, and unlike role it can go from certain to zero in one snap.
That is the whole confidence vocabulary. Nothing about efficiency, matchup quality, or talent gets its own axis.
Football compresses this cleanly because volume dominates rate. A running back's fantasy points are overwhelmingly a function of how often the ball is handed to him, not how good he is per carry. A receiver's are a function of how often it is thrown at him. Efficiency varies less between players than volume does, and it is far less predictable year to year. So if you know the role and you know the health, you know almost everything about the forecast — and the remaining uncertainty is mostly touchdown luck, which no amount of research reduces. See Touchdown variance.
A basketball warning. Fantasy basketball does not compress like this, because it pays across several rate categories and a player's shooting percentages genuinely swing his value. Carrying that instinct across produces hours spent on football efficiency metrics that will not change a single decision. The football version of that effort belongs entirely on the two axes above.
Four rules, and why each earned its place
1. Every time-sensitive claim carries an as_of date.
Nearly everything in fantasy football is time-sensitive, and staleness is invisible without a timestamp. A depth chart assertion from July and one from September look identical on the page and are worth entirely different amounts. The date is what lets a reader — including a later version of you — discount correctly rather than either over-trusting or discarding wholesale.
2. Official evidence outranks commentary, and a sourced opinion is attributed, never silently converted into fact. The failure this prevents is quiet laundering: an analyst's projection becomes "he's projected for X," which becomes "he'll get X," which becomes a number in a table with no source attached. Each step is small and the end state is a fabrication. Attribution is not politeness — it is what makes a claim revisable when the source turns out to be wrong.
3. A missing input propagates as a missing output, never as a silent default. The most important rule here, and the one most likely to be broken by code rather than by prose. If a projection is missing, the row is dropped loudly or the build fails; it never quietly becomes zero, or an average, or last year's figure. A defaulted value is indistinguishable from a real one downstream, so the error survives every check and shows up as a confidently wrong ranking.
This is not theoretical. In the basketball build this policy came from, the engine broke this rule three separate times — and all three defects were found by reading source code, not by running the audit. An audit that consumes the pipeline's own output cannot see a default that the pipeline invented. Data/build_board.py is written to fail closed for this reason: a missing file raises rather than substitutes.
4. An audit finding is verified independently before it is applied, and rejecting a wrong finding counts as much as applying a right one.
An automated or agentic check produces findings at a rate that makes applying them all feel like progress. It is not. A wrong finding applied is a defect introduced into a file that was previously correct, and it arrives wearing the costume of a fix. The basketball build rejected 91 findings as unverified, and that rejection was as much of the work as the corrections were. The corollary for this build: when a check disagrees with the plan, the disagreement gets investigated, not obeyed. One entry in _Progress.md records a case where the plan's own asserted number did not survive recomputation — and the recomputation won, but only after being checked.
The card
Evidence: Official for facts · Reported for expectations · Community for leads. Confidence: role and health, nothing else. Rules: date it · attribute it · never default it · verify before applying it.
Open questions
- The two-axis model assumes volume dominates rate strongly enough that rate can be ignored entirely. That is close to true at running back and receiver, and least true at quarterback, where touchdown rate does real work. Whether the simplification costs anything at quarterback is untested.
- There is no explicit rule here for what to do when two Reported sources of equal standing flatly disagree. Current practice is to record both and lower confidence on role, which is a habit rather than a policy.
- Rule 3 is enforced in code for the board pipeline. It is enforced only by discipline in prose notes, and nothing currently audits for a silently defaulted claim written by hand.