Workflow - how the Investor Market & Venture Opportunity Analysis was produced

The case in two minutes #

What the client asked

The production trail of the investor study: a contract issued before any action, four research lanes that grew into 36 accounted agents, a first-party measurement from the ledger, a market-size model built by two independent method families, a live re-verification of every published source, and two accounting defects found and corrected. The findings live on the Research page; this page carries only the process: steps, roles, time, cost, evidence.

Who executed

44 execution entries · 4 roles · 39 agent sessions

Capabilities

created: 53 · reused: 0 · modified: 9

Checks

validated: 44 · rejected: 0 · reworks: 1

Result

From framing to finished document the study took 13.47 hours of wall time and cost $402.92 in model spend across 43 accounted execution units, 36 of them agents. The main lesson of this case is an accounting one: the production contour turned out to be nine times wider than assumed at launch (4 lanes became 36 agents), and the token field in completion notifications excludes cache reads and understated cost by 2-7x. Both defects were caught by the project's own gates and corrected with disclosure rather than silently. On substance, the principal result is a 2.5x disagreement between market-sizing methods with an identifiable cause, and a first-party measurement that refutes the project's own earlier claim about reuse economics.

Acceptance status

in progress

Coverage and currency

44 execution entries · snapshot 3db1db19aea3

The entry count and digest state which set the totals were computed from. The digest is taken over EVERY figure in this block rather than over selected columns, so it changes when membership changes or when any figure beside it changes. An attestation narrower than its apparent scope is worse than none, because it buys trust it has not earned.

Every figure above is derived from the execution ledger and the artifact matrix below. The conclusions text is written by hand and is pinned to the entry count: if the ledger moves past it, the projection refuses to load rather than publish a stale claim.

Failures and rework #

Derived directly from this case's execution ledger: no failed or reworked unit is hidden. Zero is a statement too.

Failed: 0 · Reworked: 1 · 44 of accounted units

Case economics #

Where the money and the time went #

Between one grand total and a ledger of hundreds of rows sits the breakdown: by role, by stage and — the one that matters — how much bought KNOWLEDGE versus how much was EXECUTION on knowledge that already existed. Every bar carries its own confidence marker: measured and reconstructed figures are never added into one number.

Confidence of these sums

Dashed bars are fully reconstructed after the fact; solid bars contain a measured part. The marker at the end of each bar names its measured share.

By role

By stage

A stage’s money comes from the ledger entries of the tasks that stage declares it covers, not from the caption under the timeline step. Stages sharing the same task set are merged into one bar: the ledger does not resolve finer than the task there, and splitting it further would be an invented allocation.

Producing knowledge versus executing on it

Classification rule: an execution unit counts as "producing knowledge" when at least one artifact it created or modified resolves to a Capability Atlas entity — by the same match the artifact matrix below uses. A unit that both produced knowledge and delivered work counts wholly as knowledge production. That is deliberately conservative: the rule can only inflate the knowledge side, so the published knowledge cost is an upper bound and the reuse claim is never flattered by it.

Created and reused

Knowledge elements engaged

8

Uses of them by execution units

174

Average units per element

21.75×

Elements used by more than one unit

6

Knowledge-artifact touches (created / modified / reused)

0 / 0 / 0

Delivery-artifact touches (created / modified / reused)

53 / 10 / 0

Two different quantities, and not interchangeable: "touches" count what a unit did to an artifact, "uses" count how many units executed on that knowledge element. The reuse claim rests on the second.

Timeline caption reconciled against the ledger

The timeline step captions reconcile with the ledger entries: no divergence.

Stage timeline #

1. 2026-08-20

   contract-assembly

   resolver-operator — Knowledge Contract issued before any research action: K_id sha256:f98de211, receipt sha256:7e1cdc4a, envelope sha256:920dac58, state Unbound by design, ontology validator PASS. A new contract-grade skill (investment-analysis) authored by expansion; the Gap run that correctly refused it before approval is preserved. · public repository · 51m / $6.65

2. 2026-08-20

   first-party-measurement

   evidence-auditor — The only measured rather than cited dataset: 284 ledger entries. Produced the load-bearing negative finding that reuse economics measures 0.71x, refuting the earlier 9x figure taken from 22 entries. · public repository · 4m / $0.42

3. 2026-08-20

   desk-research-topdown-sizing

   researcher — Lane A: 168 classified figure rows across 11 intersecting markets; the ~25x embedded-vs-standalone definition gap identified as a measurement problem; Upwork's $10T knowledge-work figure shown absent from its own SEC filings. · public repository · 30m / $18.69

4. 2026-08-20

   desk-research-bottomup-sizing

   researcher — Lane B: entity counts and per-entity spend computed from Census SUSB, Eurostat SBS, ONS IDBR and BEA input-output microdata; no credible global count of firms above 50 employees or $10M revenue exists at any source class. · public repository · 32m / $13.61

5. 2026-08-20

   desk-research-funding-comparables

   researcher — Lane C plus two sub-lanes: 80+ verified rounds 2024-2026, 11 published investor theses quoted verbatim, and the bear-case evidence that comes from category insiders rather than critics. · public repository · 1.3h / $53.02

6. 2026-08-20

   desk-research-matrix-and-exits

   researcher — Lane D: competitive-matrix facts per company with unknown written wherever evidence was absent, the exit and consolidation sweep, and the category failure-mode record including the Big Four AI-defect wave. · public repository · 45m / $54.14

7. 2026-08-20

   research-fanout-subagents

   researcher — Thirty further sub-researchers spawned by the four lanes, each accounted as its own unit. Two produced work that never reached its parent and was recovered from transcripts, filling a gap the parent had declared unfillable. · public repository · 7.9h / $250.80

8. 2026-08-20

   telemetry-correction

   evidence-auditor — Two accounting defects found and corrected: an invented wall time caught by the ledger validator, and a cost basis that understated the tree by 2-7x because the notification token field excludes cache reads. All 36 agents rebooked from measured transcript telemetry. · public repository · 7m / $0.50

9. 2026-08-20

   model-and-synthesis

   evidence-auditor — The sizing model built by two independent instrument families whose 2.5x disagreement is reported with its cause, plus graded defensibility, the bear case, scenarios, exit paths, evidence gates and the recommendation. · public repository · 51m / $1.75

10. 2026-08-20

    publication-under-contract

    knowledge-curator — Bilingual publication of 21 sections under /research, the separate public Workflow, the sources manifest with every citation re-fetched live, and the Investor Room digest and link. · public repository · 23m / $2.20

11. 2026-08-20

    rework-reuse-economics

    evidence-auditor — Recomputation of the reuse-economics ratio under four explicit estimators after the customer returned it for not reproducing; the method note, the snapshot digest, and the correction applied in six places. · public repository · 13m / $1.14

12. 2026-08-20

    contract-repin-after-merge

    resolver-operator — The knowledge half of this research was merged by union rather than by choosing a side, and the Knowledge Contract was re-pinned afterwards: issuance had stopped because the graph snapshot moved, not because the merge was wrong. A field-by-field diff established that only snapshot_digest differed. · public repository · 8m / $0.59

Conclusions #

From framing to finished document the study took 13.47 hours of wall time and cost $402.92 in model spend across 43 accounted execution units, 36 of them agents. The main lesson of this case is an accounting one: the production contour turned out to be nine times wider than assumed at launch (4 lanes became 36 agents), and the token field in completion notifications excludes cache reads and understated cost by 2-7x. Both defects were caught by the project's own gates and corrected with disclosure rather than silently. On substance, the principal result is a 2.5x disagreement between market-sizing methods with an identifiable cause, and a first-party measurement that refutes the project's own earlier claim about reuse economics.

Research #