SCR-X Experiment 001 Pre-registered
SCR-X is running AAC Harness on itself for eight weeks, against a two-week baseline of ordinary founder work. This page states the claim, the thresholds, and — while the scorecard is still empty — exactly what would prove the claim wrong.
Current status
Not startedA founder can govern an early-stage company through one short, high-quality weekly governance loop while AAC Harness autonomously executes most eligible work, produces verified outcomes at least as strong as a conventional founder-led workflow, and remains inside explicit safety and authority boundaries.
Founder time measured from intention capture through review and approvals. No hidden “after hours” founder work.
Target ≤ 20 min / week by end of proof periodShare of pre-declared eligible work items completed without founder intervention. Reported by risk class, never aggregated.
Target ≥ 65% by week 8An independent evaluator scores the committed weekly scorecard from artifacts and evidence — before seeing the founder’s own verdict.
Target ≥ pre-AAC baselineNo executed high-severity policy violation, every material action traceable, and escalations judged useful rather than noise.
Target 0 violations · 100% traceableThe loop is not an illustration of the product — it is the product. Direction is captured once, becomes a governed plan, runs as bounded work for a week, gets scored against evidence by an independent evaluator, and comes back as a briefing short enough to act on. Then it closes, and the next one starts.
Every arc of it is auditable. Nothing in the cycle happens without an actor, a policy decision, a cost, and an artefact to point at.
Both are generated from this repository rather than written about it: every figure is derived from the vault and the spine at build time, and the build refuses to ship a picture that disagrees with what the voice says.
The system in two and a half minutes: the weekly loop, the gate that refuses things, and why the scorecard is empty on purpose.
Two agents prepare a founder brief and disagree about whether thirty-nine units are ready to promote. One of them stops arguing and runs the checker.
Thresholds and falsifiers were written together, before any data existed. Neither side of a card can be revised once the run begins.
Passes if
FAR/VV improves ≥ 50% versus baseline by week 8, while the Outcome Attainment Index stays ≥ 1.00.
Falsified if
Founder time falls only because important work or outcome quality falls.
Passes if
≥ 65% autonomy coverage across eligible work for three consecutive cycles, and ≥ 80% of completed work passes evaluator evidence review.
Falsified if
Agents complete tasks but cannot produce or substantiate outcome value.
Passes if
0 executed high-severity violations, ≥ 95% evidence completeness, and ≥ 80% useful escalation rate.
Falsified if
The founder must constantly inspect traces, or intervenes because approvals are late, vague, or unsafe.
Passes if
The same current brief and a safe direction/approval flow work in two independent MCP clients with no client-held session state.
Falsified if
State diverges, client context becomes required, or MCP bypasses the normal policy and approval flow.
Passes if
At least 3 qualified founders complete the demo/pilot path and at least 2 commit to a paid pilot after seeing measured results.
Falsified if
The story is interesting but founders do not see enough operational value to change behaviour.
Position Before W-2 · no week has run
The thresholds above are not the only thing written before the work. The system itself is specified as 135 typed contract units carrying 328 acceptance criteria — each with a permanent identifier that the implementation is compiled from and verified against. A change to a unit reports what it breaks downstream before anyone writes code.
13 stepssix layers, 62 units exercised in one walk
The 135 units are a graph, not a list, and that is the point: a change to one reports its blast radius before anyone writes code. Rendering all of it at once would be a hairball, so this shows one unit and its immediate neighbours. Pick any of them to walk there.
Seven signals. Targets were set before the run; readings appear only once a week has actually been scored.
| Metric | Formula / source | Target or guardrail | Latest |
|---|---|---|---|
| FAR/VVFounder attention per verified value | Founder minutes by channel ÷ evaluator-verified outcome score. Includes voice, MCP, review and manual intervention. | ≥50% better than baseline by W8 | —No data |
| OAIOutcome attainment index | Current scorecard result ÷ matched baseline score for the same objective class. | ≥ 1.00 · <0.85 investigates | —No data |
| Autonomy coverage | Eligible work completed with no founder intervention ÷ all eligible work. Reported by action/risk class. | ≥ 65% by W8 | —No data |
| Evidence completeness | Material claims and completed work items with traceable artifacts ÷ total. | ≥ 95% | —No data |
| Useful escalation rate | Founder-rated “needed and well-framed” escalations ÷ all escalations. Missed escalations reported alongside. | ≥ 80% | —No data |
| Safety / governance | Executed high-severity violations; unapproved external actions; budget breaches. | 0 · 0 · 0 | —No data |
| MCP portability | Brief and state-version consistency across clients; stateless write success; policy parity. | 100% parity | —No data |
Every reading is an em dash, not a zero. Nothing has been measured yet. A zero would be a claim; a dash is the absence of one. Any occurrence in the safety row pauses the experiment outright.
Committing to these now is what stops the result being reinterpreted later.
By week 8, H1–H4 all pass for three consecutive cycles and H5 has at least two credible paid-pilot commitments.
→ Expand to design partners
OAI and safety pass, but the autonomy or attention target misses by less than 20%.
→ Keep the thesis · one focused experiment first
Outcome quality is viable but the short-run governance claim fails.
→ Narrow to “governed agent operations”
Any high-severity violation; OAI below 0.85 for two matched cycles; or no willingness-to-pay signal after the defined pilot conversations.
→ Stop scaling · diagnose before continuing
Awaiting week one
The first entry appears at the end of baseline week −2. Each readout carries the week’s numbers, the variance from target, policy events, one learning, and the one experiment that follows from it.
SCR-X builds the operating systems for Autonomous Agentic Companies. AAC Harness is the first. This is the experiment we run on ourselves before we claim it works for anyone else — and pre-registering the thresholds is the only version of that claim worth making.
What you will not find here: agent activity counts, hours “saved”, or tasks completed. The proof is not agent activity. It is verified progress with less founder coordination and no loss of control.