菜单
Review a single stated probability or confidence judgment, and its reasoning, for the failure modes that make numbers miscalibrated; use when checking one forecast's number.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/grammy-jiang/intelligence-analysis-agent --skill calibration-and-probability-hygiene命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| name | calibration-and-probability-hygiene |
| description | Review a single stated probability or confidence judgment, and its reasoning, for the failure modes that make numbers miscalibrated; use when checking one forecast's number. |
| kind | skill |
| status | ready |
| provenance | {"principles":["P002","P003","P010","P014","P018","P028","P031","P032","P038","P052","P058","P073","P076","P078"],"claims":["C00700","C00701","C00111","C00112","C00514","C01108","C00393","C00394","C00463","C00464","C00470","C00736"],"evidence":["E00094","E00095","E00009","E00010","E00050","E00225","E00025","E00026","E00035","E00036","E00042","E00109"],"source_anchors":["d88ef7714f29-c0008","3a2b4f822beb-c0000","3a2b4f822beb-c0015","e3c7c0b4e46f-c0010","3a2b4f822beb-c0010","3a2b4f822beb-c0013","d88ef7714f29-c0010","e3c7c0b4e46f-c0009","e3c7c0b4e46f-c0022","d88ef7714f29-c0004","5da0a790f5ae-c0011","d88ef7714f29-c0019"],"authored_from_digest":"24028903e9c7b37d57cf5630e73fc3333a43c16b5e8def0f7d82e597f7d87d92"} |
A single stated probability, forecast, or confidence judgment is easy to state and hard to get right. This skill reviews that one number — and the reasoning behind it — for the specific failure modes that make probability judgments miscalibrated, overconfident, or simply incoherent:
This skill reviews and corrects the probability statement itself. It does not produce the underlying forecast, decide the substantive question the number answers, or make the call the number feeds into — that stays with the forecaster and the decision-maker.
Work through these checks, in order, against the single probability judgment under review. Each check that turns up a problem becomes one finding in Output below; a check that is already satisfied is noted rather than silently skipped.
Set the ceiling on extremeness by predictability, and question where the confidence came from (P002, P003). In a domain with essentially no genuine predictability, the right call is one steady central estimate for every similar case, not a case-by-case extreme guess; only sanction a wider spread of predictions once real predictability is demonstrably higher. A tidy, story-like match between the inputs and the claimed outcome is not proof of predictability, and that illusion holds even once the inputs behind it are already known to be weak (P002). Also check input independence: when two or more inputs are correlated and effectively say the same thing, treat any added confidence from the second one as illusory — accuracy will not rise to match it — and prefer inputs that are independent of one another and roughly comparable in their own validity (P002). Separately, interrogate any claim that a large volume of information itself justifies the stated confidence: beyond a fairly small threshold, more information keeps pushing confidence upward without a matching rise in accuracy, a pattern seen across multiple expert fields, including cases where practitioners were properly calibrated only when working from the barest minimum of information. Treat this as a strong default to test for, while allowing that specific circumstances can genuinely break it (P003).
Kill the default 0.5 (P032). A flat 0.5, or "toss-up," offered purely because the case feels unreadable is a hygiene failure, not honest humility: spread across an unpacked set of sub-cases, a reflexive 0.5 can push the total probability mass above 100%, and it wastes whatever sliver of real information the assessor does hold. When the one fact actually known is simply how long the thing being judged has already persisted, prefer a duration-based estimate (reasoning from how long something has already lasted to how much longer it likely has) over a reflexive half-and-half guess.
Check for anchoring (P038). Treat any number that was already sitting in view before the estimate was formed — a target figure, a prior number, a round figure embedded in how the question was asked — as a live anchor on the final answer by default: people under-adjust away from an anchor reliably, experts are barely more resistant to the effect than novices, and it is not a bias a person can simply will away by trying harder. Where the stakes justify the effort, confirm someone deliberately searched for reasons the anchor itself might be wrong before the final number was accepted.
Cap certainty short of the endpoints (P078). Flag any probability stated or implied at exactly 1.0 in the assessor's own view, or exactly 0.0 in a rival's — at that point the odds ratio involved is undefined, and no future evidence can shift the number again. Ask that the ceiling be drawn back to something short of the extreme (figures like .95 and .05 are a reasonable working limit) so the belief stays open to revision, noting that beliefs genuinely converge toward the truth through the accumulation of evidence — fastest when a strong prior meets evidence that is highly diagnostic — not by declaring the matter closed in advance.
One finding per problem surfaced in the Procedure, ordered highest-impact first, each in this four-part form:
Even where every check is passed, produce at least one finding-shaped note naming which specific checks the judgment already satisfies — never a bare "looks fine" with nothing named.
In advise mode (a single upcoming hygiene question rather than a full review), collapse this to one recommendation in the same four-part shape. In compare mode (weighing two ways of stating or handling the same probability — for example a tight interval against a widened one, or a flat 0.5 against a duration-based bound), lay out what each choice costs against these same checks and close with a calibration-weighted recommendation.
This skill never substitutes its own probability for the one under review, and it never makes the decision the number feeds into — the corrected judgment and the residual odds are handed back to the forecaster and the decision-maker.
../../references/calibration-forecasting-principles-index.md
— the package-wide index of every promoted principle cited above, with its full statement,
confidence level, and source anchor.This skill's every check derives from P002, P003, P010, P014, P018, P028, P031, P032, P038, P052,
P058, P073, P076, and P078 (P031 at medium confidence, the rest at high), grounded in Kahneman's
Thinking, Fast and Slow, Tetlock & Gardner's Superforecasting, Tetlock's Expert Political
Judgment, and Heuer's Psychology of Intelligence Analysis; see the frontmatter provenance block
above for the exact claim, evidence, and source-anchor ids.
Widen intervals anchored too tight, and discount precision theatre (P073, P031). Assume any stated interval is narrower than it should be until shown otherwise: subjective ranges are characteristically overconfident, with the true value landing outside a stated 98%-style interval something like three times in ten, because the range gets built by anchoring on one best guess and then under-adjusting the width around it — and rewarding accuracy alone does not cure this. Examine too exactly how the interval was elicited, since the wording or procedure used to draw a range out of someone can itself change the answer given (P073). Separately, do not let apparent precision substitute for evidence of calibration: a number carried to several decimal places, or delivered by a source who sounds sure of themselves, is not by itself proof that the finer distinctions being drawn correspond to any real difference in frequency, and a source's confidence should never be read as a proxy for that source's actual competence (P031).
Handle uncertain evidence probabilistically, and say the uncertainty out loud (P014, P010). Where the judgment leans on a fact whose own reliability is not settled, flag any place that fact gets silently rounded to a plain yes or no: treating something assessed at roughly seventy-to- eighty-percent likely as though it were simply true manufactures overconfidence on its own, and even applying a discount proportional to the fact's assessed reliability tends to leave the result still overconfident (P014). Require that whatever uncertainty remains be written down as an actual number — a probability range or a set of odds — together with where that uncertainty comes from and the specific milestones that would narrow it, rather than left as an unquantified qualifier like "fairly confident" (P010).
Multiply, don't average, a scenario's components, and check the direction of the bias (P018, P028). For a scenario built from several sequential events, recompute its overall probability as the product of the individual event probabilities rather than an average of them: averaging systematically overstates the true figure, since the whole chain can be no more reliable than its weakest link, and note that piling a vivid, plausible-sounding extra detail onto the scenario tends to make it feel more probable to a reader even while it mathematically drives the true joint probability down (P018). Then check which way a chain-structure bias is likely running: when every one of several steps must succeed for the outcome to occur, people tend to rate that outcome as more likely than it is; when any single one of several components failing is enough to cause the outcome, people tend to rate it as less likely than it is. Both errors come from anchoring on the probability of one representative step and under-adjusting for the rest of the chain, so the direction of the likely error can be read straight off whether the structure in front of you is a "must all succeed" or a "just one has to fail" shape (P028).
Run the coherence and scale checks (P076, P052, P058). Wherever the judgment can be broken into a set of outcomes that are mutually exclusive and jointly cover every possibility, add up the probabilities assigned to the individual pieces and compare that sum to whatever probability was assigned to the whole: if the whole was rated less likely than its own parts sum to, that incoherence is a genuine probability-axiom violation, serious enough to expose the assessor to being exploited across a sequence of bets constructed against them (P076). Then consider how the number will actually be read once it leaves this review: a scoring approach that collapses the whole probability scale down to its two ends treats every middling value as interchangeable, rewards a confident call at either extreme when it turns out right, punishes it when it turns out wrong, and treats a confident call that turns out wrong in the wrong direction as the single worst outcome of all (P052). And consider how a decision-maker downstream will actually weight this number rather than how it is written: people systematically give a low probability more weight than it deserves and a moderate-to-high probability less weight than it deserves relative to outright certainty, so a shift from "impossible" to merely "possible," or from "possible" to "certain," moves a decision more than an equal-sized shift in the middle of the scale would — watch in particular for a genuinely uncertain outcome framed in sequential stages ending up treated as though it were already certain (P058).
Produce the calibrated numeric probability — a draft Judgment — for a question under uncertainty: outside view / base rate first, Fermi-decompose, fold in the bias/deception review findings, adjust with case specifics, then state an explicit number with confidence and dissent. Invoke when a defensible probability is needed (not a hedge word like 'likely') and being over- or under-confident is costly. This is the sole entry point for producing the draft probability — the pipeline commits it only after the orchestrator's Step 9a audit, Step 10 human gate, and Step 10a lock. Runs as Step 9 of the structured-analysis workflow, which owns the independent reviewer audit (Step 9a), the human gate (Step 10), and the commit (Step 10a): this skill produces the number, it does not audit, approve, or lock it.
Runs an open-source (OSINT) investigation to collect and verify evidence for an analytic question — search, fetch through the egress guard, hash/EXIF for archival, archive each item into the evidence-ledger as an UNTRUSTED ingested proposal, then verify geolocation/chronolocation as candidate annotations (confirmed downstream at the human gate), and hand off to grading — never grading the item here. Invoke when a question needs external open-source collection. Not for classified sources, and not a fact-asserting tool — every result is a candidate for human confirmation.
Grade an evidence item before it drives a judgment: rate source reliability (A–F) and information credibility (1–6) on two independent axes, weight by diagnosticity, seek corroboration, check for deception, and record the absence of expected evidence. Invoke when weighing raw reporting into an analysis. The Step-3 grading method of the structured-analysis workflow; the bias-perception-reviewer and analytic-method-reviewer — and, once its security gate clears, the deception-detection-reviewer — check the result. Not for producing the analytic judgment, critiquing an existing grade, or collecting the raw material.
基于 SOC 职业分类
