michael@multipliers.dev

Ecosystem · the line

Everything here ships down the same five stages.

A game move, a dependency upgrade, a published report and this website are unrelated products. Each one began as a product requirement that needed a mechanism nobody had built yet. They pass the same gates, in the same order, for the same reason: an agent's output is a proposal until something deterministic has checked it and a person has judged it.

STAGE 01

Intent

Passes when the work names what the product has to do and where it must stop, not only its steps.

STAGE 02

Agent execution

Passes when the output is structured enough to be checked mechanically.

STAGE 03

Verification

Passes on a named reject class or a green check — never on a plausible reading.

STAGE 04

Judgment

Passes when a human or an adversarial pass has assessed what the checks cannot.

STAGE 05

Evidence

Passes when the claim carries its scope, or is not made at all.

Run a system down the line

STAGE 01

Intent

Passes when the work names what the product has to do and where it must stop, not only its steps.

01 · Intent

Model migration as experiment

A new model is not swapped in place; it runs as a controlled experiment with evaluation built into product behaviour.

STAGE 02

Agent execution

Passes when the output is structured enough to be checked mechanically.

02 · Agent execution

Schema-first clue generation

Structured outputs from OpenAI propose candidate clues; the JUDGE flow batches candidates rather than trusting the first reply.

STAGE 03

Verification

Passes on a named reject class or a green check — never on a plausible reading.

03 · Verification

Board-aware domain validators

Deterministic validation against a closed reject vocabulary: structure, membership, cardinality, completeness, consistency, survivors, provider.

STAGE 04

Judgment

Passes when a human or an adversarial pass has assessed what the checks cannot.

04 · Judgment

Second-model judge, then telemetry

Survivors go to a judge pass with a direct-clue fallback; the server-canonical pipeline-outcome event makes the failure mode reviewable afterwards.

STAGE 05

Evidence

Passes when the claim carries its scope, or is not made at all.

05 · Evidence

Live product and three reports

codenames-ai.com, with its qualified figures on the case study and three field reports on the calls behind it.

Under the line · shared infrastructure every lane draws on

4 items · 1 prototype

Team-harness plugin

Planning methodology, repo bootstrap and the cloud-hooks primitive, versioned as one plugin.

v1.10.2 · private repo · no public URL

Four-layer hook stack

Session start, file edit, pre-commit and CI — the same enforcement in local and cloud agent runs.

In force in this repo · scripts visible in source

Facts inventory

Canonical career and product facts stored once; prose may select and rephrase, never invent.

private repo · no public URL

Savepoints

Durable capture of decisions during agent work. Repo-local spool only; packaging and UI deferred.

prototype · not a shipped surface