Antidote.
Work with us

Method

Evidence before conclusion.

The public surface is a reading copy of the same governed engine partners use. It does not add a universal moat score, a composite winner, or an unsupported ROI claim.

Four planes, kept apart

Mixing capability, cost, structural rights, and modeled scenarios into one number is how diligence becomes marketing. Each plane has its own status, denominator, and decision implication.

Measured workflow capability

Task performance, critical failures, reliability, latency, human correction, and uncertainty on a frozen case set.

Measured operating economics

All-attempt direct cost, tool and API charges, correction and QA labor, full cost per accepted output, and volume scenarios.

Structural defensibility

Data rights, exclusivity, flywheel evidence, rebuild path, switching costs, distribution, and services dependence. Dimensions stay separate; they are not averaged into a moat grade.

Modeled or unknown claims

Vendor statements, scenario assumptions, and unverified hypotheses stay labeled. A failed measurement stays open. Judgments stay labeled as judgments.

Three kinds of record, one engine

The client decision selects the kind. The planes, comparison conditions, and claim labels do not change; the verdict object and the role of public-benchmark evidence do. Published benchmarks attach to the job, not the vendor. Two competing products in the same category share those benchmarks and differ in their company-specific findings.

Software CDD

An investor evaluating a software target: is the asset defensible, replicable, or exposed, and on what evidence.

Verdict object. A structural finding about the asset.

Annex role. Published benchmarks bound what a generic system already does in the category. They never score the vendor. Software analyses ↗

Product advisory

A software company choosing which AI features to build: is each feature feasibility-supported today, and what must a pilot still establish.

Verdict object. A feature thesis with per-feature feasibility floors.

Annex role. Published benchmarks bound whether models can do the task type. They say nothing about the product's data, users, or exceptions. Advisory records ↗

Automation diligence

An investor or operator of a services business: which workflow components are exposed to automation, at what evidence tier, and what stays human.

Verdict object. Two modeled objects, never blended: an automatable share of minutes, and a chain-completion bound that fails closed.

Annex role. Published benchmarks supply a floor only where the task mapping is strong. Weak mappings never borrow a score. Automation records ↗

Workflow decompositions

Automation-diligence records break a workflow into components and map each to a public floor only where the mapping is honest. The rules keep a decomposition from turning into a blended score.

  • Component shares are labeled by basis: observed only when a sanctioned workflow shadow exists, modeled otherwise. Shares sum to one; nothing is dropped or renormalized silently.
  • A public score is a floor only where the component mapping is strong. Weak mappings never borrow a score; for context-heavy work a public benchmark can overestimate.
  • Two derived objects, never blended: an automatable share of minutes (exposure, not success) and a chain-completion bound (sequential work fails at the weakest link, so a weighted average is not end-to-end capability).
  • The chain fails closed: one required component without a floor-valid mapping makes the bound indeterminate, and no number is emitted.

Industry discourse

Competitive dynamics often arrive as reporting and practitioner accounts: a vendor's model strategy, a first-party product from a model lab, a customer building in-house. That material is admissible only as an attributed, dated structural judgment.

  • Every discourse item is attributed and dated: source type, reference, and the register row it informs.
  • The evidence class is pinned to structural judgment. Discourse never upgrades itself into a measured field or a report number.
  • Figures carried by discourse are acknowledged on the record and stay out of headlines, charts, and measured cells. Prefer the mechanism without the figure.
  • Each item names the corroboration that would upgrade it and the trigger that would retire it. Stale competitive color is retired, not kept.

Comparison conditions

For each priority workflow, the method asks for the current process, the raw model surface, a reproducible generic system, the target system, and authorized ablations. Missing conditions stay on the page.

  1. Current workflow

    The incumbent process, included only when it was actually observed.

  2. Generic model

    A generally available model with no product-specific setup.

  3. Named competitor

    A declared alternative with the same job, tools, and budget.

  4. This product

    The system under review, tested with authorized access.

  5. Isolated changes

    Included only when access and rights permit. Missing comparisons stay visible.

Model efficiency protocol

A cheaper model that still clears the gates is a finding. A frontier model that does not beat the generic comparator on the same frozen cases is also a finding.

  • Documented model ladder across small, mid-tier, and frontier configurations.
  • Frozen snapshot, prompt, tools, source access, token/compute/time budgets, retry policy, human intervention, and price card.
  • Lowest-cost configuration that meets prespecified quality, critical-error, latency, and review gates.
  • No attribution of proprietary orchestration, fine-tuning, or smaller-model advantage unless the comparison supports it.

Gates that cannot be skipped

The web app is a thin workspace over the engine. Browser input cannot freeze a study, approve a claim, or mark a report released.

  1. Authorization

    Access grants, rights, and spend authority are recorded before any live call.

  2. Protocol freeze

    Study definition, cases, budgets, and models are frozen. Amendments create a new revision.

  3. Evidence lineage

    Every report number resolves to a structured analysis field and then to evidence hashes.

  4. Human review

    Truth validation, claim approval, and visual inspection remain human-gated.

  5. Release controls

    Draft, verified simulation, and released are distinct. A failed gate is indeterminate.