Datasets:

LucioLiu
/

track-record

This is the discipline's manual operation layer: the five-loop calibration cycle compressed into a card a human can run by hand — with pen and paper, a spreadsheet, or any Markdown file; equally fit for personal decision reviews and team prediction ledgers. The rule source remains SKILL.md; this card compresses, it does not add rules. All numeric thresholds live in the SKILL's single-source threshold table — this card never re-copies numbers.

Log (record the bet when you make it): only material, genuinely uncertain judgments that reality will answer. One row, seven fields: date | judgment/prediction | basis | settlement trigger | confidence & probability | actual outcome | bias distilled. Hard requirements: write the basis as an explicit bet ("betting the hot data concentrates and cache hit-rate clears 80%"); write the settlement trigger as an observable event ("check the P95 weekly report two weeks after launch", not "see later"); numeric probability, no bare adjectives ("likely" is a fig leaf; "75%" can be proven wrong); tag confidence HIGH / MED / LOW.
Triage (log only the Goldilocks zone): near-certain calls — don't log (settling them yields no information); judgments with no observable trigger — don't log (they'd rot unsettled). Multiple judgments in one task → one parent record + numbered sub-judgments, settled independently.
Settle (a hard mechanism, not a virtue): keep the ledger's live section visible at the top; scan it every time you open the file. For entries past their trigger window, force a three-way choice — there is no fourth option called "wait a bit more": ① settle now on available evidence (write "insufficient evidence" honestly if so) ② mark 【unsettleable】 with the reason ③ explicit extension — must state a new observable trigger; extension cap per the SKILL threshold table. Independent settler: ideally a different person (or a later you) settles purely by comparing the entry against real-world evidence, without hearing the bettor's explanations. Settling reads reality; it never pings people to hurry reality up.
Backfill (how to write outcomes): settle what can be settled, tag 【partial】; overturned judgments get recorded too — the entries that slapped you are the most valuable in the book; "was wrong" alone is not a backfill — name the failing step and whether you over- or under-estimated; single-sample conclusions carry an explicit n=1 warning.
Distill (bias → rule → mechanism): an isolated case stays a case note; the same bias fingerprint (direction + object type + failure mechanism all matching) recurring past threshold → write a rule that names the scenario; a rule violated again after landing → hand it to process / checklist / tooling. Confidence routing: HIGH may change rules directly; MED goes to a watch list pending one more case; LOW is record-only, rules untouched. Every promotion states "fed back into §
" — otherwise it's just a feeling. Settled entries move to the archive section; the live section keeps only open and partial ones.
Three anti-self-deception lines: "circumstances were special" → specialness goes into the distillation column, never a reason to cancel settlement; "results haven't fully landed" → the three-way choice has no "wait a bit more"; "I remember being roughly right" → remembering ≠ verifying.

Maintained in sync with the SKILL; recheck this card whenever the SKILL version bumps. If they diverge, SKILL.zh-CN.md is authoritative (Chinese is this repo's rule source).