# Decision Journal > Use when sharpening Stage 4 (inversion / strict FOREVER gate) of the v2 onboarding machinery, or when attaching a decision-journal artifact to any conviction call. Provides process-vs-outcome (resulting) discipline, base-rate and skill/luck awareness, calibrated probabilistic framing, pre-mortem/inversion protocol, and a per-stock DECISION-JOURNAL template (what we believed, bet size, kill-criteria, what would change our mind). Enriches — but does NOT replace — the binding Munger/Buffett CIO (vault) and Munger's Psychology of Misjudgment layer already embedded in the pipeline. Triggers: "decision journal", "pre-mortem", "kill criteria", "what would change my mind", "am I falling in love with this stock", "bias check on FOREVER", "was I right for the right reasons", "resulting", "process vs outcome", Stage 4 inversion on any FOREVER candidate. - Skill: `build-with-dhiraj/decision-journal` (Agent Skill) - Install (CLI): `npx skillmds@latest add build-with-dhiraj/decision-journal` - Raw SKILL.md: https://api.skillmd.com/api/skills/build-with-dhiraj/decision-journal/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: DevOps & Infra - Author: build-with-dhiraj (https://skillmd.com/u/build-with-dhiraj) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/build-with-dhiraj/decision-journal --- # Decision Journal (the behavioral / decision-quality keeper) **What this skill adds:** the binding Munger/Buffett CIO already owns the *substance* here — process-over-outcome, crowd psychology, and the hold-forever-vs-cardinal-sin-of-delay tension are all dense in the binding layer. The gap is not doctrine; it is **instrumentation**: the CIO has no persistent artifact, no scheduled re-decision cadence, and no structured way to memorialize a call so it can be re-read post-outcome. This skill supplies exactly that — a contemporaneous, re-readable artifact per conviction call (to distinguish *good process that went wrong* from *bad process that got lucky*), the missing mechanism inputs the canon method requires, and a re-decision loop so the journal cannot quietly become a rationalization engine. Every addition below is *instrumentation of a CIO-native principle*, not a new first-principle. ## STEP 0 — Pull the canon (binding-safe) ```bash set -a && source /Users/Dhiraj/dev/invest/.env && set +a && /Users/Dhiraj/dev/invest/.venv/bin/python \ /Users/Dhiraj/dev/invest/data/scripts/32_consult_brain.py \ --corpus blended \ --company " decision journal process outcome resulting bias premortem inversion calibrated probability base rate fear and greed market cycle sentiment crowd psychology hold-forever inconsistency-avoidance status-quo re-decision second-order and-then-what psychological distance" \ --model general \ --step inversion \ --k 12 \ --json-out extracted/grilling/_journal.json ``` `--corpus blended` surfaces the desk-synthesis behavioral atomics (`vault/desk/behavioral/atomic/`, weight 0.9) alongside the binding CIO. The extended query terms (fear/greed cycle, inconsistency- avoidance, status-quo re-decision, and-then-what, psychological distance) are deliberate — the grilling proved those crowd-psychology / temperament atomics are in the corpus but the old query string under-retrieved them. `cites_principles` MUST be a strict subset of returned slugs. If `principles` is empty, ABSTAIN — do not fabricate behavioral claims. Also run the binding consult (`--corpus binding`, Munger/Buffett CIO) for the moat/verdict judgment; the canon supplies the *behavioral method*, the CIO is still the *arbiter*. **Reliably-surfaced behavioral slugs** — this is a *reference list of what the consult tends to return, NOT a citation allowlist.* Re-read the rule above: cite a slug ONLY when THIS run's `principles[]` actually returns it; if the consult comes back thin or empty, ABSTAIN. Never cite from this list because it appears here. - `[[decision-journal-method]]` — maintain a contemporaneous journal documenting context, alternatives, reasoning, expected outcomes, probabilities, and emotional/physical state to counteract hindsight bias - `[[decision-journaling-pre-outcome]]` — deconstruct and memorialize decisions *before* results occur to enable unbiased post-mortem analysis - `[[resulting-outcome-bias-test]]` — evaluate decision quality independently of outcome; never let a good result validate a flawed process or a bad result indict a sound one - `[[motivated-reasoning-diagnosis]]` — identify and control for motivated reasoning (processing information to confirm existing beliefs); the FOREVER confirmation trap is motivated reasoning in disguise - `[[pre-mortem-group-analysis]]` — solicit independent disconfirming analysis *before* the outcome - `[[probabilistic-decision-discipline]]` — frame the buy as a probability-weighted bet; state the odds, not a binary call; accept that well-calibrated bets can lose - `[[calibrated-probabilistic-thinking]]` — force probabilistic quantification of beliefs; refine with new information; resist collapsing to "it will work" certainty - `[[bayesian-updating-priors]]` — continuously update probability estimates by integrating new information with relevant base rates; resist anchoring to the original thesis - `[[accuracy-vs-rightness-reasoning]]` — reason for accuracy (objective truth), not to be right (affirming priors), especially in post-hoc review - `[[review-for-pattern-recognition]]` — periodically review past journal entries to identify recurring errors, cognitive biases, and judgment patterns - `[[skill-luck-continuum-placement]]` — place the investment activity on the skill-luck continuum; stock-picking is HIGH-luck, requiring larger samples to evaluate analyst performance - `[[mean-reversion-analysis]]` — analyze mean reversion rates to infer skill vs luck; a hot streak in a high-luck domain is not a signal of repeatable skill - `[[can-you-lose-on-purpose-test]]` — if you can trivially construct the losing case, luck dominates; use this to calibrate how much weight to place on recent track-record evidence - `[[single-outcome-deep-dive]]` — focus analysis on ONE specific failure scenario in full concrete detail rather than generic contingency planning - `[[sample-size-tailoring]]` — require large samples before inferring skill from outcomes; resist concluding "our process works" from one or two good outcomes ## WHERE THIS SKILL PLUGS IN **Primary hook: Stage 4 — Inversion / strict FOREVER gate** (runs alongside the adversarial disconfirmation agent). Before the Opus/Fable verdict, run through the DECISION-JOURNAL TEMPLATE below and flag any bias-pattern findings to the inversion agent. A FOREVER candidate that cannot pass the resulting-test and pre-mortem is demoted by default. **Secondary hook: any conviction call (Stage 3)**. Attach the filled template to the `vault/research/dossiers/.md` as a "Decision Journal" section at the bottom. It becomes the re-readable artifact for post-outcome calibration review. ## METHOD — process-vs-outcome discipline ### 1. Resulting test (run FIRST) Before writing the journal entry, ask: "Would I grade this process identically if the stock had declined 30% in the past year?" (`[[resulting-outcome-bias-test]]`). If recent performance is influencing the conviction grade, correct for it — process quality and outcome are independent. A FOREVER belief built on a strong recent chart is motivated reasoning, not analysis. ### 2. Base-rate / skill-luck calibration Stock-picking sits on the HIGH-LUCK end of the continuum (`[[skill-luck-continuum-placement]]`). Implications: - A two-year track record of good picks means almost nothing statistically (`[[sample-size-tailoring]]`). - Any analyst's "consistent outperformance" claim needs the `[[can-you-lose-on-purpose-test]]` and `[[year-to-year-correlation-test]]` checks before the Substack analyst's thesis is up-weighted. - ROIC persistence (e.g., IGI/CAMS high ROIC sustained 5+ years) is a harder signal than analyst conviction, because ROIC mean-reverts in competitively weak businesses (`[[mean-reversion-analysis]]`). ### 3. Pre-mortem (mandatory for FOREVER candidates) Run the pre-mortem as an independent step (`[[pre-mortem-group-analysis]]`): *assume the thesis failed 3 years from now* and write the most plausible autopsy. Solicit the inversion agent's version independently before comparing. Red flags: if the failure scenario is easy to write, the moat is not as durable as claimed. Hard rule: a pre-mortem that produces the same 3-4 failure modes as the bull thesis re-stated in reverse is not a real pre-mortem — push harder. ### 4. Motivated-reasoning check (the FOREVER confirmation trap) The FOREVER tier is the highest-stakes judgment — and therefore the highest-risk locus of motivated reasoning (`[[motivated-reasoning-diagnosis]]`). Before confirming, explicitly list *three pieces of evidence that would have caused you not to reach this conclusion*. If you cannot list three, you have a confirmation problem, not a conviction. Run the same lens *outward* on management, not just on yourself: score the company's capital-allocation narrative for self-serving bias (`[[self-serving-bias-detection]]`) and incentive distortion (`[[institutional-incentive-analysis]]`) — a buyback/empire-building story told by the people whose comp depends on it is a flagged input, the bridge to the capital-allocation and forensic desks. The FOREVER label *itself* is also a behavioral hazard (`[[forever-label-as-commitment-trap]]`): it manufactures commitment-and-consistency and endowment bias, turning Munger's cardinal sin of inconsistency-avoidance into the gate's default — which is why the holding must be periodically *re-decided* (see Hard rule 8). ### 5. Probabilistic framing (not binary) Translate the verdict into a probability statement (`[[probabilistic-decision-discipline]]`, `[[calibrated-probabilistic-thinking]]`): *"I believe there is X% probability this business sustains its moat and compounds owner-earnings at ≥Y% for ≥10 years."* Then state what the probability would need to be to justify buy-below at the MoS threshold. If the required probability seems heroic, widen the MoS or demote. Express uncertainty honestly — "70% confident" is more useful than "high conviction." ### 6. Bayesian update discipline (ongoing) Each quarter-result, management change, or analyst revision is a Bayesian update opportunity (`[[bayesian-updating-priors]]`). The kill-criteria in the journal (see template) define the evidence that should *update the probability down materially*. If a trigger fires but the thesis is unchanged, that is a motivated-reasoning warning sign. ## DECISION-JOURNAL TEMPLATE Attach as a `## Decision Journal` section at the bottom of `vault/research/dossiers/.md`. Fill BEFORE outcome is known (pre-outcome memorization per `[[decision-journaling-pre-outcome]]`). ```markdown ## Decision Journal **Date:** YYYY-MM-DD **Verdict:** FOREVER / COMPOUND / WATCH / AVOID / TOO_HARD **Conviction:** N/5 **Buy-below (MoS gate):** ₹X **Probability statement:** "I believe there is ___% probability this business sustains its moat and grows owner-earnings at ≥___% for ≥10 years." **Decision-time state:** [one line — emotional / physical / time-pressure state right now, per [[decision-journal-method]]; "rushed and excited after a great quarter" is a real input, not noise.] ### Sentiment-cycle locus Where in the fear–greed cycle is this buy being made — **panic / normal / euphoria**? Run the `[[be-fearful-when-others-are-greedy]]` contra-check: am I buying because the business is cheap relative to IV, or because everyone is? (`[[sentiment-cycle-locus-of-a-decision]]`, `[[fear-and-greed-drive-markets]]`, `[[market-cycles-are-driven-by-human-nature]]`.) A FOREVER formed at self-identified **euphoria** is demoted by default unless the MoS is widened. This is a *qualitative temperament check*, never a market-timing call — it adjusts MoS demand, it does not gate entry on a macro forecast. ### What we believed (the thesis, in one paragraph) [Write the bull case *without* using the stock's recent price action as evidence.] **And-then-what chain:** [for the core bull driver, write 2–3 second/third-order steps — "prices rise → and then what? (competitor adds supply / regulator caps fees / customer substitutes)" per [[and-then-what-as-second-level-thinking]] / [[and-then-what-sequence]]. A thesis whose first-order story is bullish but whose second-order chain turns bearish has not been thought through.] ### Alternatives considered and rejected [Name ≥2 other verdicts you considered and why you rejected them. Per [[explicit-alternatives-evaluation]] / [[probabilistic-decision-discipline]].] ### Key variables and assumed odds | Variable | Assumed value | Why | If wrong → | |---|---|---|---| | Moat durability | [e.g., "stable 10yr"] | [reason] | [verdict change] | | ROIC 5yr avg | [e.g., "≥28%"] | [reason] | [conviction drops to N-1] | | Mgmt alignment | [e.g., "owner-operator"] | [reason] | [re-evaluate] | ### Kill-criteria (what would change our mind) List the 3–5 concrete, observable triggers that would cause a verdict downgrade or sell decision: 1. [e.g., "ROIC drops below 18% for 2 consecutive years"] 2. [e.g., "Regulatory cap on registry fees imposed with no pass-through"] 3. [e.g., "Promoter stake falls below 40% without strategic rationale"] 4. [e.g., "New entrant clears SEBI approval for competing registry"] 5. [e.g., "Price breaches IV-high by >30% with no earnings upgrade"] ### Pre-mortem (failure autopsy) *Assume this bet failed 3 years from now. What is the most plausible explanation?* [Write 2–3 sentences. This should NOT be the bull thesis inverted — find the non-obvious failure.] ### Bias check - [ ] Resulting-test passed: conviction grade is identical regardless of recent price performance - [ ] Three disconfirming evidence items listed: 1. ___ 2. ___ 3. ___ - [ ] Motivated-reasoning diagnostic: no evidence of processing data to confirm a pre-existing FOREVER belief - [ ] Psychological-distance pass: re-read the thesis as if a *stranger* wrote it — does it still hold? (`[[psychological-distance-on-own-thesis]]` / `[[objective-evaluation-guards-against-bias]]`) - [ ] Management self-serving-bias check: capital-allocation narrative scored for incentive distortion (`[[self-serving-bias-detection]]`), not taken at face value - [ ] Analyst source check: Substack bull thesis not driving conviction; moat evidence is independent ### Post-outcome review (fill on next annual review or after a kill-trigger) **Date of review:** YYYY-MM-DD **What actually happened:** [facts] **Was the process sound?** Y / N — [reason, citing [[resulting-outcome-bias-test]]] **Bayesian update:** probability revised to ___% **Calibration note:** [what this outcome teaches about our priors / process] ### Standing FOREVER re-decision (re-grade at each annual review or kill-trigger) Re-grade the holding as a **fresh buy at today's price** (`[[status-quo-decision-equivalence]]`): would I open this position now, this size, at this price? **Not selling is logged here as an active decision with its own probability statement** — silence is not a decision. The bar to SELL stays the binding *compelling reason* (`[[hold-with-forever-mindset-unless-reason-to-sell]]`); this is a re-decision *cadence*, not a license to churn on a noisy chart (`[[resulting-outcome-bias-test]]` still guards against selling on a bad recent quarter). **Re-decision date:** YYYY-MM-DD — **Re-buy at today's price?** Y / N — **why:** ___ ``` ## INTEGRATION WITH THE INVERSION AGENT (Stage 4) The inversion agent (Opus/Fable) receives the filled journal in its prompt. The adversarial skeptic must: 1. Evaluate whether the kill-criteria are *real* (specific + observable) or vague (e.g., "moat deteriorates" without a measurable trigger — reject vague criteria). 2. Check whether the pre-mortem failure modes overlap with the kill-criteria (they should). 3. Flag any bias-check box left unchecked — an unchecked box is a FOREVER demotion signal. 4. The inversion agent's verdict carries more weight than the journal author's — default is DEMOTE unless the skeptic is convinced by the journal + canon evidence together. ## Hard rules 1. Fill the journal BEFORE the outcome is known; a post-hoc journal is rationalization, not a record. 2. Kill-criteria must be specific and observable — "moat weakens" is not a kill-criterion. 3. The pre-mortem must name a non-obvious failure; restating the inversion checklist is not enough. 4. The resulting-test and motivated-reasoning check are mandatory for every FOREVER candidate; no bypass. 5. Cite only canon slugs returned by the Step 0 consult; if the vault is thin, ABSTAIN on behavioral claims. 6. The Munger/Buffett binding CIO (vault) remains the final arbiter of moat/verdict; this skill supplies the *process discipline and artifact*, not the *conviction judgment*. 7. Substack analyst conviction is an INPUT, not the verdict — always run the resulting-test and skill-luck calibration (`[[skill-luck-continuum-placement]]`) on analyst track-record claims. 8. **A FOREVER holding is re-decided, not protected.** At every annual review (or kill-trigger), re-grade the holding as a fresh buy at today's price (`[[status-quo-decision-equivalence]]`) and log *not selling* as an active decision — the FOREVER label is a commitment/endowment trap (`[[forever-label-as-commitment-trap]]`). This is a CADENCE, not a churn license: the SELL bar stays the binding *compelling reason* (`[[hold-with-forever-mindset-unless-reason-to-sell]]`), and the perspective voices that preach "quality compounds forever, never sell" illustrate but never arbitrate — Hard rule 6 (the Munger/Buffett CIO is the final arbiter) sits above them.