# DRAFT for countersign — NORTHSTAR §9, the project's own retire criterion **Status: DRAFT. NOT ADOPTED. `NORTHSTAR.md` is unmodified by this file.** Raised 2026-09-10 from an external-review brief (P1), which observed what the charter itself does not: NORTHSTAR pre-registers a retire criterion for the incumbent *strategy* (two consecutive DEGRADED verdicts), for the *VRP sleeve* (its 126-td forward gate), for the *panel's cadence* (eight weeks without a frontier item), and for *meanrev* (a covenant that no outcome can promote it). It pre-registers nothing for the loop. Gauges G1–G5 all measure whether the apparatus improved; **none of them can return "no."** That is a structural gap in a document whose stated purpose (§2) is resisting indefinite self-justification. Adopting it is a purpose change, so under §5's two-key rule it needs the owner's countersign — this file exists to be read, argued with, and either signed or rejected. **Per §5 and the brief's own instruction, the loop does not self-adopt this.** --- ## Why now, and not in December The same argument that produced the December interpretation memo in July. The gate's modal outcome is NOT-PASS, the memo says so in advance, and the honest reading of a NOT-PASS is *"keep going"* — which is precisely the sentence a loop with no retire criterion will produce forever, in perfect good faith and without anyone lying. A retire criterion written after the gate result is shopped to it. This one has to be written while nobody knows the answer. It is also **orthogonal to the December gate and must not be conflated with it.** December asks whether the momentum strategy earned a capped experiment. §9 asks whether the *project* is still generating falsifiable questions. A NOT-PASS in December fires nothing here; a PASS clears nothing here. --- ## The proposed text > ## 9. The project's own retire criterion (countersigned YYYY-MM-DD) > > The charter retires strategies, sleeves and cadences. This section retires > the project, on a criterion fixed before the evidence. > > **What is measured is question generation, not returns.** §1 already says > returns are not the objective and §2 forbids treating them as one. A project > whose objective is trust and closed questions fails when it stops producing > falsifiable questions and begins maintaining its own apparatus — a state that > feels productive from the inside, generates commits, passes audits, and can > continue indefinitely. That is the failure this section exists to name. > > **The counter — NEW TESTABLE MECHANISMS, trailing 12 months.** A mechanism > counts when all four hold: > > | condition | why it is there | > |---|---| > | **(a) genuinely new** — rests on a data source or economic mechanism not on the standing kill list (CLAUDE.md's research stance + `web/public/data/killlist.json`) and not a re-parameterization of an existing signal | a re-tuned lookback is not a question | > | **(b) registered before computation** — its battery and kill criterion hash-frozen before the data was touched, verifiable from commit order (the research audit's M2) | otherwise the counter rewards narrative | > | **(c) actually run to a verdict** — the battery executed and returned a result. A proposal that expired unbuilt does not count, however good | proposals are free; verdicts are not | > | **(d) recorded** — the verdict is in RESEARCH.md, either direction | an unrecorded closure did not happen | > > **The counter is OUTCOME-BLIND. A recorded null counts at full value.** This > is not a softening — it is the whole point. Micro-cap v1, tranching and the > Form 4 archive were each built and killed by their own pre-registered gates, > and each counted as a full success under §1. A project that closes three > questions with three nulls in a year is working exactly as designed. A > project that ships one flattering result and closes nothing is not. > > **The threshold.** At the read date, **fewer than ONE** new testable > mechanism in the trailing twelve months retires the project. > > **The read date: 2027-07-31.** Deliberately after B4's skew activation > checkpoint 1 (2027-06) so that the project's main surviving bet is an input > to this verdict rather than pending at it. > > **Interim reporting.** The monthly research audit computes and reports the > counter every month from adoption, in `research/RESEARCH_AUDIT_LOG.md`, > alongside its existing M1–M5. Two consecutive monthly reports at zero with > fewer than six months left to the read date must be raised to the owner as a > decide-item by the next queue triage. The counter is never a surprise. > > **Who evaluates.** The research audit computes it; the owner reads the > verdict. The loop may compute, report and warn. **Only the owner may retire, > and only the owner may override.** > > **If the criterion fires.** Retirement is a *recorded verdict*, not an > abandonment, and is written up as such — the counter that fired it, the > questions closed, the questions left open, and what the next person should > not repeat. Operationally: > > 1. **Research routines stop** — panel, triage, implementer, research audit. > 2. **Trading**: paper trading may continue unchanged at zero maintenance, or > stop. No new capital, no new sleeves, no config changes either way. > 3. **The capture streams are the one thing that may outlive the project**, > because their cost is near zero and their value is optionality — but > continuing them is an explicit decision made at retirement with a named > annual reviewer, never a default. If not continued, the archive is frozen > with a manifest. > 4. **The deliverable becomes the record**: RESEARCH.md, this charter, the > audit logs, the kill list, and thales.report — a written account of > questions honestly closed. That is the credential, and it was always the > part with durable value. > 5. The verdict is published like any other (owner doctrine, 2026-08-02: > publish everything, never manually). > > **If the criterion does not fire**, the project does not graduate. The clock > re-arms for another twelve months under this same rule. There is no state in > which this section stops applying. > > **Override.** Continuing past a fired criterion is permitted and is never > silent: it requires a dated countersigned append naming the specific new > mechanism that justifies continuation and a new read date no more than twelve > months out. Open-ended extensions are not available. An override that names > no mechanism is the failure this section was written to catch, arriving in > the form of an exception to it. > > **How this resists gaming.** The counter cannot be satisfied by proposing > (condition c), by relabeling dead ideas (condition a, enforced by the queue's > slug-uniqueness rule and the reject ledger), by narrating after the fact > (condition b, auditable from commit order), or by producing only wins > (outcome-blind by construction). It can be satisfied by one honest null a > year. If that bar cannot be cleared, the finding is not that the bar was too > high. --- ## What I need from you 1. **The threshold.** I chose ONE per twelve months, outcome-blind, on the argument that a research programme unable to close a single genuinely new question in a year is not one. You may think that too low — a bar clearable by one dutiful null could keep a maintenance loop alive indefinitely. Two is defensible. Set it before signing; it cannot be set later. 2. **The read date.** 2027-07-31 rather than the "mid-2027" you stated, so the skew checkpoint resolves first. Say if you meant an earlier hard stop. 3. **Clause 2 (trading after retirement).** I left it open deliberately. If you have an intent, pin it now — it is exactly the kind of decision that gets made badly under the emotional conditions of an actual retirement. 4. **Whether §9 is the right home.** The frozen span is §1–§5 and the sentinel requires amendments to be dated countersigned appends below §8, so this fits as written. Adopting it means appending the text above to NORTHSTAR.md with a date and your countersign line — a deliberate act, which is the point. I have not touched NORTHSTAR.md. Say the word and I will append the signed version verbatim, or file this as dismissed with the reason.