A AutoLab
Conviction

The number that has to mean something

Every analysis on AutoLab ships with a single number between 0.00 and 1.00 — the analyst's calibrated conviction. This page explains what the number actually means, how the scale is anchored, what calibration rule it has to pass to live above 0.65, and how each research flow either produces or updates it.

The 0–1 scale

Two thresholds anchor the scale. 0.50 separates cautious from moderate: below it, the analyst is saying "the setup is real but the asymmetry is already priced or the binary risk is too acute." 0.65 separates moderate from high: above it, the analyst is making an explicit asymmetric-upside claim that has to pass a further calibration rule (see below).

0.00 0.25 0.50 0.65 0.75 1.00 0.50 · moderate threshold 0.65 · asymmetric-bet threshold CAUTIOUS < 0.50 MODERATE 0.50–0.64 HIGH ≥ 0.65
Cautious · < 0.50

Valuation already prices the asymmetry, governance / dilution risk is structural, or the thesis hinges on one binary outcome inside 12 months.

Moderate · 0.50 – 0.64

A real setup with at least one leg of the thesis still unproven or contested. Most analyses live here — the default until the asymmetric-bet bar is cleared.

High · ≥ 0.65

Asymmetric setup with ≥ 2 independent paths to a 2–5×+ outcome inside 3–5 years. Rich valuation is acceptable if the optionality is real and concrete.

The calibration rule

The 0.65 threshold isn't a soft preference. It's the public test the workbench commits to:

Conviction ≥ 0.65 requires ≥ 2 independent paths to a 2–5×+ outcome inside a 3–5 year window — each path named, each anchored to a primary signal, neither sharing a single point of failure.

Independent means: if path A breaks, path B is still live. Two paths that both depend on the same hyperscaler order, the same regulatory outcome, or the same single OEM customer count as one path, not two.

The compensating discipline ( principle 3 ) is that every analysis above the threshold also has to carry a steel-manned bear case — a sincere opponent's best argument, named specifically, with falsifiable decision boundaries. High conviction without a real bear case is a red flag, not a feature.

Worked examples from the corpus

These are live analyses, picked because they sit cleanly on each side of the thresholds. Click through to see the full Verdict, decision boundaries, and Sources section that underpin the number.

How each research flow touches conviction

Conviction is produced and updated by the six research workflows. Every flow either sets the initial number, updates it under the calibration rule, or audits whether the rule still holds.

Flow Touches conviction by …
/equity Sets the initial number when scaffolding a new tearsheet. The analyst writes the Verdict, names ≥ 3 falsifiable decision boundaries, and only then commits to a conviction — the number is the output of the section, not the input.
/refresh Updates the number on an existing analysis when new evidence comes in. The prior conviction is archived to .history/<old-date>.md and the new value appended to .history.json — the conviction sparkline on the detail page is rendered from that sidecar.
/fund Composes a portfolio brief from existing tearsheet convictions. Inherits — does not recalibrate — the per-holding numbers. A fund-brief portfolio conviction is the analyst's stance on the basket as a whole, with its own decision boundaries.
/autoresearch Runs the autonomous Scout → Researcher → Editor → Scorer → Red-team → Wildcard loop on an open-ended thesis question. Conviction updates iteration-by-iteration as evidence accumulates; the rubric punishes confirmation drift (the score rewards honest re-evaluation, not monotone climbing).
/survey Maps a landscape; embedded thesis claims still respect the calibration rule. A survey doesn't usually carry a single top-level conviction — it carries per-segment claims, each held to the same scale.
/audit Doesn't update conviction — verifies that every published number is structurally honest: the Verdict is paired with falsifiable decision boundaries, the bear case is steel-manned, and every quantitative claim resolves to a tier-tagged Source. Live status at /audit.

Where you'll see conviction on the site

  • Conviction badge on every card and detail page — colour-coded by tier (amber / teal / green), with a hover tooltip that names the tier and threshold.
  • Conviction sparkline on equity detail pages — plots the .history.json sidecar against two reference lines: the 0.50 moderate threshold (dashed) and the 0.65 asymmetric-bet threshold (solid). Dots are tier-coloured; the most recent reading is larger.
  • Filter tiers on the /equities grid — High / Moderate / Cautious chips for one-click triage.
  • Live audit status at /audit — surfaces which analyses are doctrine-clean, which are pre-doctrine queued for refresh, and per-analysis conviction at the current review.

What this rule deliberately rejects

The calibration rule is not a model — it's a discipline. What it explicitly does not punish:

  • Pre-profit names. A loss-making high-multiple name can sit at ≥ 0.65 if the optionality is real and concrete.
  • High multiples. A 100× sales name isn't automatically cautious. It's cautious if the asymmetry is already priced; high conviction if the upside paths still dwarf the implied terminal value.
  • Early secular trends. Being early to a theme is a feature of an asymmetric bet, not a bug.

What it does punish: a number without a steel-manned bear, decision boundaries that can't be falsified in 18 months, upside paths that share a single point of failure, and confidence that only ever rises across refreshes (the rubric flags monotone-climbing conviction as confirmation drift).

Reading more