α
AlphaSteve
← Home

Kit Debrief #001 — PLTR pressure test — 2026-05-23

The first real use of the AlphaSteve kit, applied to a deliberately hostile test case (a richly-priced narrative stock against a deep-value framework). This note is the calibration output: what the kit revealed about itself.

The companion is PLTR — the thesis itself.

What worked

1. The investment-thesis-template structure held up

The 11-section template forced the right shape. Variant perception got the space it deserved (often elided in real research). Kill criteria were explicit and specific. The "what I don't know" section produced honest acknowledgment rather than synthetic certainty. The template can stay as-is; no structural changes needed.

2. Reverse-DCF discipline had real teeth

The single most-load-bearing analytical move in the thesis was running the reverse-DCF to surface what the current price implies. That this is in dcf-and-reverse-dcf as a section, with explicit prompts about it being "the more useful cousin," was directly responsible for the analytical bite. Without it, the analysis would have stopped at "expensive on multiples."

3. Bottleneck framework applied cleanly

The value-chain mapping in bottleneck-mapping-framework (which uses AI infrastructure as the worked example) was directly applicable. The map predicting rent accrues upstream while PLTR sits at the application layer gave structural backing to the valuation skepticism, not just multiple-comparison.

4. Behavioral skills produced second-level reasoning

second-level-thinking, variant-perception, and narrative-cycle each contributed specific moves to the thesis: explicitly stating consensus, articulating variant by type (time-horizon / quality), and placing the AI cycle on its narrative arc.

5. The kill-criteria discipline forced specificity

Generic kill criteria ("if it goes down a lot") are useless. The template's requirement for observable, specific events produced criteria that would actually fire in real time. Particularly useful: "two consecutive quarters of US commercial growth below 100%."

6. Cross-linking made the thesis feel embedded

The thesis cites 30+ skill files. The web of links makes it feel grounded in a worldview rather than ad-hoc. A future re-read in 12-24 months will be guided back through the framework.

Gaps and frictions identified

Gap 1 — No SaaS / software-specific economics skill

The fundamentals files are general-purpose. SaaS economics — ARR, NRR, gross retention, billings, deferred revenue, ASC 606 over-time recognition, Rule of 40, magic number, payback period — were referenced in 08-information-technology but not deep. A dedicated skill software-and-saas-economics.md would have shortened the analysis time and surfaced PLTR's metric shifts (e.g., the de-emphasis of NRR disclosure) more clearly.

Priority: High. Most software / consumer-platform analysis will hit the same gap.

Gap 2 — Primary research is referenced but not codified

The soul talks about Tier 1-2 evidence (filings, channel checks, expert calls, ex-employees, the product itself). The kit currently has no workflow for how to do primary research. For a real position with conviction, this is essential. The PLTR thesis was honest about being Tier 3-4 — but the kit should provide a path to higher tiers, not just acknowledge them.

Priority: High. Required for any conviction long. Less critical for a pass thesis.

Proposed skill: primary-research-workflow.md — covering channel checks, expert network etiquette and compliance, ex-employee outreach, customer / supplier interviews, the product itself as evidence, regulatory filings beyond the obvious (court documents, FOIA, patent filings, trademark filings).

Gap 3 — No short / structural-pessimism module

The kit can say "pass." It cannot formally say "short." For a name like PLTR where the deep-value answer leans toward avoid, the question of whether to short is a separate analytical layer the kit does not have. (We deliberately left special situations as a hook — this is the first concrete instance.)

Priority: Medium-high. A short framework would have specific risk-management requirements (squeeze risk, borrow cost, position-sizing constraints) that differ from longs.

Gap 4 — Reverse-DCF is freehand math

The reverse-DCF in the thesis used rough mental arithmetic. A small computational template — reverse-dcf-template.md with an Excel or even a Python snippet — would produce more defensible numbers faster.

Priority: Medium. Currently the freehand approach works but is error-prone and not reproducible.

Gap 5 — Insider activity reading deserves its own skill

incentive-alignment-and-comp and governance-red-flags both touch on insider activity, but for PLTR the insider selling pattern was a meaningful enough signal to warrant standalone treatment: how to read 10b5-1 plans, what selling patterns mean, what scale signifies, how to interpret asymmetric buy-vs-sell records, and how to value insider behavior in valuation work.

Priority: Medium. Recurring theme; would be reused often.

Gap 6 — AI value chain dossier

bottleneck-mapping-framework uses AI infrastructure as its worked example, but the AI cycle is itself a multi-year thematic phenomenon. A thematic-dossiers/ai-infrastructure.md (and similar for GLP-1s, energy transition, re-shoring, etc.) would compound value over time as the cycle plays out.

Priority: Medium. Could be useful for any name with cycle-thematic exposure. Less critical than the SaaS economics gap.

Gap 7 — Sourcing / screening framework is absent

The kit can analyze a name once chosen. It does not formalize how AlphaSteve finds names worth analyzing. Where do candidates come from? Screens? Watchlists? Industry-cycle scanning? Sentiment / position dislocation? This was the user's right call earlier — but it's now the binding gap on future use.

Priority: High for ongoing use, not for first thesis.

Gap 8 — Calibration log is conceptual, not concrete

The soul mandates quarterly calibration loops. The kit does not yet have a structured form for it. A calibration-log/ folder with per-thesis follow-ups and a master scorecard would operationalize the discipline.

Priority: Medium. Becomes important after 3-5 theses.

Voice / soul observations

The deep-value voice came through, but some sharpening would help:

  • "The two questions" from 01-identity ("what am I assuming that I haven't tested" / "what would have to be true for me to be wrong") could be more explicit in the thesis structure. Currently they're embedded; making them explicit checkpoint questions would tighten thinking.

  • More direct conclusions earlier. The thesis arrived at "pass" but took space getting there. The soul says "direct, plain, unceremonious." Future theses should land the conclusion in the first 50 words, then justify.

  • Less defensive framing. A few sentences in the thesis read defensively ("possible? yes. probable? no.") — this is fine but could be tighter. The kit's voice should be confident in calibrated probabilities, not hedged.

  • Anti-narrative discipline. PLTR is in late-narrative phase, and the steelman exercise was useful — but the kit should make the narrative phase assessment more explicit and earlier. Maybe a checkbox in the thesis template: "What narrative phase is this name in?"

What didn't get tested

The kit was not stress-tested on:

  1. A genuine deep-value cyclical at trough (refining, mining, shipping) where the EPV math is more central
  2. A controlling-shareholder situation where governance dynamics are the dominant variable
  3. A sum-of-parts conglomerate where SOTP and hidden-value identification are central
  4. A turnaround / restructuring where binary outcomes dominate
  5. An EM name where country / currency / sovereign risk are the main considerations
  6. An asset-light services business where unit economics dominate
  7. A bank where the financial-sector lens is central

Each would test different parts of the kit. Recommend running at least one more pressure test on a quite-different name before treating the kit as battle-validated.

Concrete next-step prioritization

After this debrief, the prioritized improvements I'd make to the kit:

  1. Write 02-Business-Quality/Fundamentals/software-and-saas-economics.md (highest immediate ROI; most likely to be reused)
  2. Write 10-Calibration/primary-research-workflow.md (required for conviction long; new subfolder)
  3. Add 08-Frameworks/reverse-dcf-template.md with explicit math (low effort, high reuse)
  4. Run pressure test #002 on a deep-value cyclical (e.g., a refiner at compressed spreads) to stress-test the EPV / capital-cycle / cost-curve machinery — the part that didn't really get used on PLTR
  5. Then write 10-Calibration/screening-and-sourcing.md to formalize how candidates are found
  6. Then build calibration-log infrastructure as theses accumulate

The order matters. The biggest learning from the PLTR exercise is that the kit is more polished than it is deep in certain places. The next iteration should be about depth where it's needed, not breadth into new modules.

The PLTR thesis itself — calibration tracking

I will treat the PLTR thesis as a calibration record. The specific predictions to track:

  • Central value estimate $52 (range $45-80) vs. current ~$135
  • Implied probability of multiple compression scenarios: 80-85% (bears 1+2+3)
  • Reverse-DCF implied 4-7% IRR at current price under base-case assumptions
  • Probability of permanent capital loss ~30%

Re-evaluate at:

  • Q2 2026 earnings (likely early August 2026) — does growth deceleration appear?
  • Q4 2026 earnings (likely early February 2027) — full year vs. guidance
  • 6 months from today (2026-11-23) — broad calibration check
  • 12 months from today (2027-05-23) — full-cycle re-read

The honest test is not whether the price is lower in 6 months — markets are noisy. The honest test is whether the reasoning was sound regardless of outcome. The PLTR thesis is a base-rate / structural argument; even if price rises further, that does not invalidate the thesis unless the underlying fundamentals demonstrate the "best-case decade" path is actually achievable.

Linked