grahama.co

Proof Method route

One real run, from goal to receipt.

Each step resolves to a real artifact and states what it does not prove.

Inspect Repository path

Ask an agent to trace the work through source.

  1. Startskills/*/SKILL.md
  2. Checkskills/monitor-website/run.sh
  3. Inspectsite/design-roundtable/*
  4. StopNOT_TESTED stays visible

  1. Goal

    The objective is fixed and hashed before anything runs. Every later step is checked against this exact hash.

    invariantgoal_hash is frozen; any drift trips goal_hash_mismatch and the run fails closed.

    provesThe work was pinned to one stated objective, identified by a content hash.

    does not proveThat the objective itself is the right thing to build.

    dag.json · sha256 8387f6624b5a79b5…
    immutable_goal
    A prioritized, executable elevation plan for the live grahama.co: harsh critique of the current site, restrained 2026 motion language, unified art direction for the 10 project imag…
    goal_hash
    sha256:e31c26f1586bbab75304fd79d4eb1cb8e5a9622fa192201b3c562ab83fa59f64

  2. Bounded context

    Limits, required evidence, and fail-closed conditions are declared up front — the agent operates inside a fence.

    invariantMissing any required-evidence artifact aborts the run rather than reporting success.

    provesPermissions, timeouts, and stop conditions existed before the agent acted.

    does not proveThat the chosen bounds are sufficient for every risk.

    dag.json · sha256 8387f6624b5a79b5…
    limits
    {"default_timeout_seconds":300,"max_concurrency":4,"max_total_attempts":5,"resume":true}
    required_evidence
    ["handler_response_receipt","normalized_handler_receipt","roundtable_join_receipt"]
    fail_closed_on
    13 conditions incl. goal_hash_mismatch, target_changed, unexpected_node

  3. Constrained execution

    The exact command Tau dispatched, with its side-effect declaration, recorded verbatim.

    invariantmutates:false is a declared contract — a write would contradict the receipt and fail review.

    provesWhat ran, whether it could mutate state, and its timeout — captured, not narrated.

    does not proveThat the reviewer's output is semantically correct.

    tau-dispatch-command.json · sha256 73733cb2c87f6df8…
    mutates
    false
    compile_only
    false
    requires_network
    true
    timeout_s
    3780

  4. Captured evidence

    The handler receipt records that a real provider answered live — not a mock, not a claim.

    invariantlive:true with mocked:false is required; a mocked lane cannot satisfy the evidence gate.

    provesA real model produced this seat's response, on the record, at a stamped time.

    does not proveThat the response's content is right — only that it genuinely occurred.

    node-receipt.json · sha256 e697899eb7dcc021…
    handler
    webclaude
    live
    true
    mocked
    false
    browser_model_preference
    Fable 5 High
    created_at
    2026-08-05T20:58:04Z

  5. Deterministic judgment

    A join gate — not a model — counts usable seats and emits a PASS/FAIL verdict against a typed schema.

    invariantThe verdict is computed from seat receipts; it cannot be set by a model asserting success.

    provesA deterministic gate ran and recorded how many seats actually returned usable work.

    does not proveThat the panel's collective judgment is correct.

    node-receipt.json · sha256 b2e310580f9ef327…
    schema
    ask.tau_dag_roundtable_join_receipt.v1
    status
    PASS
    topology
    concurrent
    usable_response_count
    4
    failed_seat_count
    0

  6. Receipt

    The top-level receipt carries its own explicit boundary — what it proves and, just as loudly, what it does not.

    invariantThe receipt states its own does-not-prove list; the claim can never exceed the evidence.

    provesThe run reached a typed terminal state with a scoped, self-limiting proof statement.

    does not proveThat PASS means the design advice was good — only that the process is accountable.

    execution-status.json · sha256 1f65826b88b3fa7d…
    status
    PASS
    live
    true
    mocked
    false
    proves
    Tau's real CLI was invoked against the emitted DAG artifact.
    does_not_prove
    Provider calls occurred unless provider_live is true and provider receipts are present.

tau node receipt — the run that designed this site

PREFLIGHT: PASS

A real model produced this seat's response, on the record, at a stamped time.

That the response's content is right — only that it genuinely occurred.

node-receipt.json · ask-tau-roundtable-round-2-webgpt-webcla-ce79c8ef960c · sha256 e697899eb7dcc021…. The judgment, proof boundary, and raw payload stay together; the summary is never the only evidence.

node-receipt.json · ask-tau-roundtable-round-2-webgpt-webcla-ce79c8ef960c · sha256 e697899eb7dcc021…

monitor-website audit — this page, checked against the repo

LOCAL DRIFT: 0 · PUBLIC ENDPOINTS: 6/6

The checkout's generated surfaces agreed with their source files, excluding artifacts.json while it was being regenerated, and the named public endpoints returned expected markers at capture time.

That the deployed HTML was built from this exact commit, or that every interaction works.

skills/monitor-website/run.sh audit --ignore-surface artifacts.json --json. The judgment, proof boundary, and raw payload stay together; the summary is never the only evidence.

skills/monitor-website/run.sh audit --ignore-surface artifacts.json --json

inventory provenance — where the numbers come from

BUILD: e6395236

The numbers above came from checked source state, not marketing copy.

That every skill is complete, useful, or production-ready.

site/inventory.json · sha256 d7dd84e2e5178a3c…. The judgment, proof boundary, and raw payload stay together; the summary is never the only evidence.

site/inventory.json · sha256 d7dd84e2e5178a3c…