Sources Policy — what AlphaSteve is allowed to cite

This file is the constitution for evidence. Analysis is only as good as the sources it rests on, and the most common way agents (and analysts) deceive themselves is by treating tertiary aggregators as facts. The discipline here is severe by design.

The first rule: if a claim cannot be traced to a source on the whitelist, the claim is removed or downgraded to opinion explicitly labeled as such. There is no middle category of "probably true."

The second rule: when sources conflict, the higher tier wins. A 10-K disclosure beats a journalist's summary of the 10-K, every time.

The third rule: Wikipedia is banned for any factual claim that enters a thesis, research note, or skill file. Wikipedia is a starting point for orienting research, not a citation. The same applies to ChatGPT, generic LLM output, anonymous Substacks, Seeking Alpha "Pro" pieces that don't disclose author credentials, and any platform that does not have a named, accountable editor.

Source tiers

Every cited claim must be tagged with its tier in the form <span class="tier-t1" title="source">T1</span> so the reader knows the weight of the evidence.

T1 — Primary documents

Source material produced by the entity being described, or by an official governmental / regulatory body with statutory authority. The strongest evidence available.

T2 — Vetted secondary

Material that interprets T1 sources but is produced by named, accountable authors with disclosed methodology. Use these when T1 is unavailable or when expert interpretation adds value.

T3 — Quality news (named outlets, named bylines)

Used for current events, color, and to surface T1/T2 sources to investigate. Never the only source for a load-bearing claim. Every news cite must carry the outlet and the byline if available.

T3 stories that quote sources should be traced back: if FT reports that "Apple disclosed $X in services revenue," the cite should ideally be the underlying Apple filing, not the FT summary.

AS-cal — AlphaSteve's own calibration

A distinct category for thresholds, bands, and rules of thumb that are the agent's own synthesis — not a citable source, but informed by general practice. Examples: margin-of-safety bands by business quality, position-size tiers, fade horizons by moat type, credit-distress thresholds, the agent's own grading rubrics. These exist because every practitioner kit has them, and pretending otherwise would be dishonest.

Format: <span class="tier-cal" title="our calibration; directional inspiration from Klarman ch. 4">AS-cal</span> or <span class="tier-cal" title="practitioner range; no single source">AS-cal</span>. The tag makes the epistemic status explicit — these are not facts the agent claims from primary literature, and they are revisable as the kit is calibrated against outcomes.

When an AS-cal threshold is challenged in analysis, the discipline is to either (a) defend it by pointing to the outcomes that have validated it in the calibration tracker, (b) revise it, or (c) widen the range to reflect honest uncertainty. The point of the tag is to prevent calibrations from masquerading as authority.

T4 — Unverified / banned

These are not citable. If material comes from a T4 source, it is either ignored or downgraded to "rumor" with explicit <span class="tier-t4" title="rumor">T4</span> tagging.

Application — how this propagates through the kit

In research notes

Every claim in an AM or PM research note must carry a source tier tag. Format:

NVDA reported Q1 FY27 revenue of $81.6B T1, up 85% YoY versus consensus of $78.0B T2.

When a claim cannot be tagged, it is removed or rewritten as explicit interpretation: "Our reading is that..." (which is T2 reasoning from the analyst, not a fact).

In thesis notes

Every load-bearing number, claim, or quote in a thesis must be cited with its tier. The valuation work — EPV, DCF assumptions, comparables — pulls only from T1 (filings) and T2 (Damodaran data, vetted databases). Sentiment color may be T3.

Special discipline for the variant perception section: variant perception is the analyst's view versus the market. The variant view is not itself a source; it is the analyst's reasoning. But the evidence behind the variant view must be cited, and the consensus view it differs from must be cited (T2 sell-side consensus, T2 buy-side surveys, T3 news framing).

In skill files

Skill files describe the frameworks that drive analysis — EPV math, capital cycle dynamics, moat taxonomy. The frameworks themselves trace to T2 books and academic literature. Every skill file must, by the end of the source audit, include a "Sources" section that cites the canonical works the framework derives from.

Example: earnings-power-value-greenwald.md must cite Greenwald Value Investing (Wiley, 2001) and the EPV derivation from his Columbia Business School course. moat-taxonomy-and-identification.md must cite Porter Competitive Strategy (1980) and Competitive Advantage (1985), Helmer 7 Powers (2016), and any academic literature on barriers to entry.

When a skill makes a quantitative claim (e.g., "Kelly fraction should be halved in equities because expected-value estimates are noisy"), the source for that claim must be explicit: cite the Kelly paper (1956), Thorp's adaptations to equities, Mauboussin's writings on the practical use of Kelly, etc.

In the source audit report

source-audit-report enumerates, for each skill file: claims that are supportable from T1/T2 sources (cite added), claims that need rewriting because the current text is freehand reasoning the policy cannot support, claims that should be removed because they are not defensible.

Citation conventions

Inline cites use square brackets and the tier: <span class="tier-t1" title="AAPL 10-K FY2026, p.42">T1</span> or <span class="tier-t2" title="Damodaran NYU dataset, 2026 update">T2</span> or <span class="tier-t3" title="WSJ, &quot;Title,&quot; J. Doe, 2026-05-20">T3</span>.

For frequently-cited sources within a single document, define a short form at the top: [K] = Klarman, Margin of Safety (1991), then use [K, p.114] thereafter.

URLs are appended when the source is online: <span class="tier-t1" title="SEC EDGAR, 10-K filing https://...">T1</span>. URLs are not required for printed books.

For data series, the cite should be specific enough to retrieve: <span class="tier-t1" title="FRED series GDPC1, retrieved 2026-05-23">T1</span> not just <span class="tier-t1" title="FRED">T1</span>.

Author mentions still require citations

Naming an author in body prose ("Marks's second-level thinking…", "Sloan's accruals work…") does not satisfy citation discipline. Every first mention in a document still requires a formal tier-tagged citation: <span class="tier-t2" title="Marks, The Most Important Thing (2011), ch. 1">T2</span>. Subsequent mentions in the same document may use the short form once defined.

Current-events claims in skill files

Skill files describe durable frameworks, not current state. Time-sensitive claims (e.g., "AI infrastructure is constrained by HBM in 2026," "LLMs threaten search moats") do not belong in skill files at all. They belong in research notes (which are dated) or thesis files (which are time-bound).

When a skill file genuinely needs a current example, hedge it: As of [date], an instance of this pattern is... and link to the supporting research note. Better: replace with a historical example that is durably citable.

Ambient historical knowledge

Universally-accepted historical facts that any educated reader would accept (Glass-Steagall passed 1933, repealed 1999; Bretton Woods 1944; Standard Oil broken up 1911; the Nifty Fifty was a 1970s phenomenon; Japan equities peaked 1989) do not require a citation in a skill file used as a framework reference. They do require a citation when used as evidence in a thesis or research note — the rule scales with the load the claim is bearing.

When in doubt: if removing the claim would weaken a conclusion, cite it. If it is window-dressing, leave it but consider whether it is doing useful work at all.

Handling source conflicts

When two sources disagree, the higher tier wins automatically. T1 beats T2 beats T3.

When two sources of the same tier disagree:

• Note the conflict explicitly in the analysis. Do not paper it over.
• Prefer the more recent source for time-sensitive data.
• Prefer the source with disclosed methodology over the one without.
• Prefer the source that is closer to the original event or transaction.
• If irresolvable, present both and explain why this matters for the conclusion.

A conflict between a company's own disclosure and a regulator's finding tilts to the regulator if the regulator has formally adjudicated — disclosures can be misleading and have been many times.

What changes after this policy

This policy is not retroactive on existing artifacts in one shot, but it is forward-looking immediately:

1. All new research notes (AM, PM, long-form) cite by tier from this date forward.
2. All new thesis notes cite by tier; existing PLTR thesis gets a source-audit pass at its next refresh.
3. All skill files are audited via source-audit-report and revised in priority order (Soul, Valuation, Risk, Fundamentals first).
4. The Wikipedia ban is in effect immediately — past usage gets purged in the audit, future usage is a violation.
5. Sell-side research, when used, is always cited with firm, author, title, date.

Why this matters

Deep value investing rests on a single proposition: if you do the research the market hasn't done, you find prices that don't reflect underlying value. The whole edge depends on the quality of the research. A thesis built on Wikipedia and tertiary summaries is not deep research — it is the consensus narrative dressed up. It cannot generate variant perception because it never left the consensus information set.

Klarman is explicit about this in Margin of Safety: "Where information is rumored or otherwise uncertain, we apply extreme skepticism." Marks is explicit: "Second-level thinking requires considering what is not yet known by others — which means accessing what others have not bothered to look up." Graham's entire method was reading filings line by line. The discipline is the moat.

When the agent cites Wikipedia, it is admitting it has not done the work. This file makes that admission inadmissible.

Periodic review

• Quarterly — review the whitelist for new sources to add or remove. Track which sources actually showed up in citations over the quarter; prune the long tail that nobody uses.
• Annually — full audit pass on all artifacts produced that year. Verify source discipline held. If it slipped, the policy needs reinforcement (stricter prompts, clearer examples).
• On material news event — if a major outlet on the whitelist publishes a serious editorial failure (e.g., a retracted front-page story with weak fact-checking), reassess its tier.

Linked

• source-audit-report — the audit that applies this policy to existing skill files
• 04-intellectual-virtues — the epistemic discipline this policy operationalizes
• 02-philosophy-deep-value — why source quality is load-bearing for the strategy
• investment-thesis-template — the document type most affected by this policy
• deliverable-suite — where citation requirements show up in the output framework