The Toyota production system and jidoka in venture capital

The line that never finishes

A weaving shed in 1924 had one problem worth solving: a broken warp thread produces bad cloth for as long as the loom keeps running, and nobody notices until the bolt reaches inspection. Sakichi Toyoda's automatic loom fixed that by giving the machine authority to stop itself the instant a thread snapped. Taiichi Ohno generalised this into jidoka — autonomation, machine intelligence plus human judgement — one of the two pillars of the Toyota Production System. The principle: detect the deviation at the point of occurrence, not at the end of the line.

Venture investing has a line too. It runs from a market thesis, through diligence, into a term sheet, through board seats, and out the far end as a return or a write-off. The defect that jidoka was built to catch — a flawed unit travelling for months before anyone flags it — has an exact analogue here: a thesis defended for a year after the market it assumed has already dissolved. The corpus that produced the thesis was frozen at the point of investment. Everything after that is inspection at the exit, and the exit is a write-down.

What arrives at the loom

Four streams feed a working thesis, and each has its own cadence. Filings — Form D notices, UCC liens, state registrations, patent continuations — arrive in batches, weeks behind the events they describe, but they are structured and cheap to parse. Hiring signals move faster: a company's own careers page, LinkedIn transitions, a sudden run of postings for a role the org never needed before, or the quieter tell of a VP quietly de-listing their title. Product telemetry, where it is visible at all — app store rank, API call volume through a public gateway, GitHub commit velocity on an open-core repo — updates daily and is the noisiest of the four. Market structure is the slowest and the most decisive: a competitor's Series C at a valuation that implies a different unit economics story than the one in your memo, a platform's API change that removes a category of company's entire value proposition overnight, an incumbent's acquisition that consolidates the buyer side of the market you were betting would stay fragmented.

None of these streams closes. A Large Language Model's corpus would take a snapshot of all four at diligence and never look again; that snapshot is the memo. A Large World Model would keep one scene live — say, the product telemetry — and revise happily within it. Venture's actual condition is closer to the third rung: every stream keeps running for as long as the position is held, and the thesis is a belief carrying provenance, not a fact carrying a timestamp.

What the thesis holds

A thesis, written properly, is not "this company will win." It is a chain of load-bearing claims: the market is fragmented, the buyer's switching cost is low, the founding team's velocity outpaces the incumbent's response time, the unit economics improve with scale rather than degrading. Each claim has a source and a date. Provenance is not paperwork here — it is the only thing that makes revision possible without re-deriving the thesis from scratch. A partner who cannot point to which claim rests on which stream cannot know which incoming signal is relevant to which belief, and ends up either ignoring everything or reacting to everything, which is the same failure from two directions.

The standard misreading of jidoka is to treat it as constant alarm — watch every metric, flag every wobble. Andon did the opposite: Toyoda's loom called a human only when something was actually wrong, which is precisely why one weaver could tend dozens of machines unattended the rest of the time. Applied here, that means the four streams are not a dashboard to be stared at. They are inputs to a small number of load-bearing claims, and silence is the default state. The signal-to-noise problem in venture is not too little data; four public data brokers will happily sell a partner more hiring-signal noise than they can read in a week. The discipline is deciding, in advance, which specific deviations count as a snapped thread for this thesis.

The snapped thread

Two events count, and only two. The first is mutual contradiction between streams: hiring signal says the company is scaling sales into enterprise, product telemetry says usage is flat among accounts above 500 seats, and market structure says the incumbent just shipped the exact feature this company was selling as a wedge. None of those three facts alone proves anything. Together they contradict the claim in the memo that reads "enterprise motion is working." That contradiction is the pulled cord.

The second is calibration failure — a belief that predicted an observation which then did not arrive. The thesis assumed a Series B within eighteen months at a step-up multiple consistent with the comparable set; twenty months in, no B, and two comparables in the set have down-rounded. That is not new information exactly; it is old information about the future finally overdue, and it is detectable the day the eighteen-month mark passes, not the day the company runs out of runway.

Neither test requires knowing, in the moment, which side of the contradiction is correct. That is the honest limit of the analogy: a broken thread on a loom deviates from a written tolerance, and there is no tolerance band on a claim about a market. What survives the disanalogy is narrower but still load-bearing — mutual contradiction and calibration failure are both computable at the point of intake, without adjudicating the truth, and both were exactly what a human andon pull encoded on the factory floor: not a certified defect, but a warranted surprise.

What the partner sees

The operator's view is not a wall of red lights. It is closer to a graded andon: most pulls resolve before anything actually stops. On a real Toyota line, a fixed-position andon gives the team leader the rest of the cycle — thirty to sixty seconds — to fix the issue before the line halts; most pulls never stop production at all. The venture equivalent is quarantine, not liquidation. A contradiction flags the specific claim it undermines, not the whole position. The claim is marked contested, withheld from the next allocation decision, and routed to the partner with a note on which streams disagree and since when. Most quarantines resolve in a partner meeting: the enterprise dip was a single large account's renewal timing, not a trend. Some do not, and the thesis gets rewritten with the contested claim replaced, not silently absorbed. What the partner never sees, if the system works, is dozens of the four-stream feeds updating in real time. They see the handful of moments when two of those feeds actively disagree about something the fund's capital is resting on.

The objection that this produces permanent stoppage — that an open, noisy market throws statistical anomalies constantly, and a system built to halt on all of them would never write a term sheet — is correct as stated, and it is why the graded response matters more than the detection. The claim under defence here is about where detection happens, at the point a stream updates rather than at the point an LP asks a hard question in an annual meeting, not about how forcefully the process must react once it happens.

What it costs

The cost of detection latency compounds the way rework costs compound on a factory floor: caught one station downstream, it is a correction; caught at the exit, it is a write-off; caught by the LP, it is a reputational one. A thesis defended for a year past the point its market assumption failed does not cost a year of drift. It costs the follow-on check written at month eight on the strength of a stale thesis, the board seat spent defending a position that a competitor's Series C had already quietly falsified, and the opportunity cost of the capital that sat committed to a dissolving market instead of moving to one still forming.

None of this makes the four streams sufficient. Filings lag, telemetry is gameable, hiring signal is thin for private companies, and market structure reads clearly only in hindsight for the cases that already blew up. The disputed lineage of the Toyota Production System itself is a caution here too — TPS succeeded inside levelled production schedules, supplier keiretsu, and low product variety, conditions nothing in venture resembles, and lean transplants elsewhere often failed outright. What survives the transplant is not the manufacturing system. It is the narrower claim about where a defect is cheapest to catch, and that claim needs no factory to be true: the earliest possible detection point is at the source, continuously, while the position is still open — and there is no earlier station than that.