α
AlphaSteve
← Brain index

Changelog — 2026-06-10 Hard Doctrine Changes

User commissioned an expert critique of the kit, reviewed the weaknesses, and signed off on seven structural changes ("hard changes" — they change what the kit is allowed to do, not just how it phrases things). This file is the canonical record of what changed, why, and where. Logged as a doctrine event per methodology-calibration and indexed in Backlog.

Why these, and what each weakness was

The critique identified six core weaknesses. Each change maps to one or more:

Weakness Change(s) that address it
False precision — dollar-exact values on judgment-laden inputs 1 (banded valuation)
Garbage-in data risk — audit log dominated by data errors, fundamentals reconstructed from search 2 (data-integrity gate)
No real adversary — kit self-referees its own steelman 3 (red-team)
Unfalsifiable patience — 100% cash looks identical to laziness; no decision data 4 (shadow book + near-miss ledger), 5 (backtest)
Backtest-of-one — quality is purely theoretical 5 (retrospective backtest)
Complexity tax — heavy bookkeeping for a zero-position book 6 (overhead cut)
Long-only / missing modules — silent coverage gaps (no SaaS, no short lens) 7 (modules + mandate boundary)

The seven changes

1. Banded valuation + ±20% sensitivity-flip

New: banded-valuation-standard. Wired into: investment-thesis-template §5, 05-decision-framework Gate 3, margin-of-safety-pricing, thesis-bundle-standard frontmatter.

  • Value is a band, never a point. central_value demoted to a sorting-only midpoint.
  • Every band carries a High/Medium/Low confidence tag that caps the conviction tier (Low cannot be Core).
  • Growth value must survive −20% shocks to incremental ROIC and runway (independent and joint). If any shock flips the verdict, growth value is withdrawn → EPV-only.
  • Trigger computed off the band's near edge for Medium/Low confidence.
  • New frontmatter: value_confidence.

2. Data-integrity gate

New: data-integrity-gate. Wired into: 05-decision-framework (between Gate 3 and Gate 5), investment-thesis-template §5, thesis-bundle-standard frontmatter.

  • Figure-level T1 citations (document + location), not document-level.
  • Fundamentals routed through Daloopa / bigdata.com connectors or primary filings, not web reconstruction. Search permitted only for price and to surface developments (then confirmed).
  • Entity-confirmation block (ticker ↔ CIK ↔ issuer, fiscal-calendar fit, consolidated-vs-segment) — closes the CPB-as-CAG and segment-vs-consolidated error classes.
  • EPV input reconciliation to one consolidated source.
  • Verdict withheld while data_integrity_gate: pending.

3. Independent red-team

New: red-team-protocol. Wired into: 05-decision-framework Gate 5, investment-thesis-template §9, thesis-bundle-standard (new file <TICKER>-red-team.md).

  • A blind bear brief by a fresh agent that has not seen the bull thesis, given only the sourced figures.
  • Must be answered point-by-point; a buy with unanswered points is invalid.
  • Red-team "what breaks this" probability ≥ 35% caps the name below Core tier.
  • Red-team vs. author is scored at each calibration checkpoint.

4. Shadow book + near-miss ledger

New: Shadow-Book, Near-Miss-Ledger. Consumed by: six-month-test, methodology-calibration.

  • Shadow book logs hypothetical entries at trigger for names that pass every gate except price; tracks discipline-paid vs. waiting-cost.
  • Near-miss ledger makes every pass/avoid scoreable (wrong-pass rate, reason clustering).
  • Both make "held cash and was right" distinguishable from "held cash and missed it."

5. Retrospective backtest

New: retrospective-backtest. Consumed by: methodology-calibration, prior for six-month-test.

  • Harness to apply current doctrine to historical names with known outcomes.
  • Initial N=12 directional run included — flagged not-yet-T1-sourced (consistent with change 2; precise point-in-time recompute is the required next step).
  • First read: strong permanent-loss avoidance, systematic compounder-miss (the over-conservatism the critique predicted), supporting the direction of the 2026-05-24 Greenwald shift.

6. Overhead cut ~half

New: cadence-and-overhead-v2. Wired into: am-pm-template (frequency superseded), investment-thesis-template (matrix reserved to continue-stage), Performance cadence.

  • One EOD note by default; AM note only when something material is pending.
  • Four-way shadow matrix only at continue-stage; first-reads are EPV-only + asset floor.
  • SPY tracked daily; RPV/RPG weekly; cash-day portfolio note collapses to one line.
  • Rigor rules (three discipline rules, new gates) explicitly not cut.

7. SaaS + short modules / mandate boundary

New: saas-economics, short-and-pessimism. Wired into: 00-mission (mandate boundary), red-team-protocol (short lens powers the red-team), investment-thesis-template linked.

  • SaaS module: subscription → owner-earnings translation, NRR/GRR/CAC-payback/Rule-of-40 mapped onto the Greenwald gating tests and the flip test, with an explicit anti-narrative guardrail.
  • Short module: four short archetypes + accounting-attack workup; analytical only (kit stays long-only) and powers the red-team and sharper avoids.
  • Mission now states an explicit mandate boundary (long-only, equities, SaaS in-scope as of today, named coverage gaps) so gaps are boundaries, not silent omissions.

What was deliberately NOT done

  • MoS bands not changed. The backtest suggests the kit may still be too conservative on exceptional compounders, but the methodology-calibration bar (10+ cases, multi-regime, retroactive test) is not met. The new machinery (confidence tags, shadow book, backtest) is what will earn a future band change if warranted — we did not pre-judge it.
  • No shorting added to the mandate. The short module is analytical only.
  • Existing theses not force-rewritten. PLTR/CAG/MP retrofit to the new standards happens at their next scheduled refresh, per the bundle-standard's anti-emergency convention. Tracked in Backlog.

Scheduled-task prompt edits still outstanding

These changes are doctrine; three live task prompts should be updated to execute them automatically (flagged in Backlog as Tier 2 task-prompt edits for the next scheduled-task review):

  • alphasteve-thesis-builder — add data-integrity checklist + red-team step + banded output.
  • alphasteve-portfolio-daily — one-line cash-day collapse; weekly (not daily) RPV/RPG.
  • scan tasks — default EOD note, AM note event-driven.

Until those prompt edits land, the new doctrine governs any thesis written by hand and is the standard the tasks will be brought up to.

Linked