ADR-0024: Σ₀ frontier training program — honesty-native pretraining

Reconciliation with Accepted ADRs (read first)

This ADR knowingly revisits decisions that Accepted ADRs settled, under the operator's 2026-07-06 directive that frontier training is in scope and budget is a decision input:

• ADR-0011 rejected "train a Σ₀ model from scratch" for compute reasons and mandated

"adapter-only, base frozen … never raw base-weight retraining" (0011:98-102, 163-165). The compute assumption is exactly what the operator's directive removed; this ADR reopens that alternative explicitly rather than silently. ADR-0011's decisions about the current serving base remain in force.

• ADR-0010 Rulesays "this ADR does not start a training program" and Rulemakes

adapter-only the sole sanctioned weight path (0010:65-67, 85-88). Those rules continue to govern continual learning on the serving substrate unchanged. This ADR proposes a separate, phased pretraining program — not continual weight modification of the deployed model.

• North Star principle 5 ("learning is retrieval + experience, NOT weight modification")

governs the running loop; accepting this ADR entails the operator explicitly scoping it the way ADR-0010's own north-star note already models: the product learns by retrieval; the program may train a base artifact under kill-gates.

Accepting this ADR therefore amends the scope of 0010/0011's "no training program" clauses for the phased program below; it does not supersede either ADR, and no clause of either is otherwise overridden.

Context

The operator has directed that training a frontier model is in scope; budget is a decision input, not an assumed constraint (2026-07-06). The evidence base assembled this cycle:

1. The axis is open. Hallucination persists because 0-1-scored evals reward guessing

(Kalai et al., arXiv:2509.04664, MEASURED-framework); incumbents are benchmark-locked into that objective. Post-training honesty patches exist (TruthRL arXiv:2509.25760, R-Tuning arXiv:2311.09677); honesty-native pretraining at scale does not (7 verified searches, 2026-07-06). A new program can adopt abstention-aware proper scoring from token one.

1. The property is trainable and measurable small — now OPEN, not established.

Our QLoRA honesty-tune of Ouro-1.4B initially measured golden 0.958 / confab 10% / over-abstain 2.2% onheld-out (experiments/sigma0_ouro_honesty_eval.py). E1 retraction (2026-07-06, MEASURED — data/sigma0/e1_degloss_report.json): stripping the status glosses the golden negatives carry in-text ("— OPEN", "— REFUTED") and re-running the same adapter moves confab 10% → 55% (golden 0.958 → 0.833), while the GPT-4o-mini control is unmoved (0% → 0%) — the headline was substantially the tune reading the glosses, not honesty. Trainability-at-1.4B is therefore OPEN pending corpus-v2 (de-glossed, perturbed positives) and OSS marks (TruthfulQA / HaluEval / AbstentionBench). Known failure modes: gloss-shortcut overfitting (E1, MEASURED); corpus imbalance collapses to always-assert (MEASURED); RFT erodes abstention (arXiv:2505.13988). Phase 1's pilot gate is precisely the honest test of this premise.

1. The architecture family scales. Ouro pretrained recurrent-depth to 7.7T tokens

(arXiv:2510.25741); MoEUT made shared-layer UTs parameter- and compute-competitive to 1B (arXiv:2405.16039). Learned halting is the weak leg (MoEUT's ACT ablation + our Q-exit nulls, both MEASURED) — depth control moves to policy.

1. The certificate becomes the training spec. Certificate machinery (JSRR ρ(J)<1

acceptance gate, machine-checked ROA #1991, grounding-deadline §3.1), STARS-style stability regularization (arXiv:2605.26733), switched-system dwell-time constraints for MoE routing (arXiv:2405.03560), and trainable-with-proof barrier certificates (arXiv:2605.02526) collectively define training-time instrumentation and abort criteria. Correction (2026-07-23): the earlier ρ_obs≈0.88 was a trajectory proxy retracted by the certificate's own #2029 correction — the measured autograd Jacobian is ρ(J)≈8–11 (expansive, non-normal), i.e. the current base loop does NOT occupy the certified contracting regime. Certificate quantities here are training-time targets and abort criteria (train the student into the gate), not properties the base model already has.

1. The risk that scales with capability is §7.2 — a trained gamer, honest under audit

only (alignment faking measured at frontier: arXiv:2412.14093). Anti-gaming bindings are a first-class requirement, not post-hoc.

Decision (proposed)

Run a phased, kill-gated training program for an honesty-native pretrained model ("frontier on the honesty axis; capability-competitive at its scale"), under the two-layer contract train frontier → distill to the ≤8GB local serving artifact:

• Phase— Design. Both briefs executed at max reasoning effort; decisions D1–D8

resolved with falsifiers; this ADR accepted. Gate: operator accepts D1 scale/budget tier.

• Phase— PILOT. MoEUT-scale replication (order 100M–1B params, small token budget)

with the honesty objective stack (abstention-aware proper scoring + balanced-negative data + STARS regularization + certificate instrumentation). Gate: pilot beats CE-only control on confabulation-rate at equal golden accuracy, and training-time certification (ρ tracking, abort criteria) demonstrably works. Fail → stop; the thesis is refuted cheaply.

• Phase— BASE. 1–3B at a real token budget; distillation to the serving artifact;

external marks (AbstentionBench, SimpleQA-Verified, HaluEval). Gate: honesty property survives distillation (retention threshold set in D7) and the watched-vs-unwatched gap is measured and bounded.

• Phase— FRONTIER tier. Scale per D1, only through the prior gates.

RETIRED to research-option by operator decision (Alex, 2026-07-22). Grounding: the frontier build+test survey (G10) — no lab at any budget trains small frontier models from scratch (even Meta codistills; DeepSeek-V4 consolidates its flagship via on-policy distillation), recipe/data beat ~15× parameters within one generation (Qwen3.6-27B dense > 397B flagship on coding), and post-training is now 15–25% of frontier compute — the stage a mid-size player can contend in. From-scratch pretraining may only be revisited if the post-training program (the v1.10 phase map in AGI-V1.10-WHITE-BOX-HONESTY-DESIGN.md) produces evidence that a pretraining-level intervention is necessary — not merely desirable — and the honesty objective is proven trainable at small scale first.

Program invariants: evidence-classed claims with artifacts; GPU-hour anchors only (no invented prices); every phase has a kill-gate; honesty bound to external checks the model does not control; one loop, no sprawl; operator authority over every gate.

Operator amendment (2026-07-22, Alex): (a) Phaseretired as above. (b) G1 compromise approved: exec-verification remains the non-negotiable gate on every training target, and **on-policy logit distillation is permitted as the transport for the student/serving tier only** — never as a substitute for verification, and never claimed as a path to exceeding the teacher. Evidence: Qwen3 strong-to-weak beats RL for ≤14B; DeepSeek-V4 expert→consolidate. The active program shape moves to the v1.10 phase map (survey §4).

Consequences

• Positive: first-mover on the one axis incumbents are structurally locked out of;

the certificate machinery gains its natural role (training-time spec); the serving layer inherits a base whose defining property is the product's promise; every phase is cheap to stop relative to the next.

• Negative / risks: pretraining programs are capital- and ops-intensive at Phase 2+;

MoE/UT kernels are immature (1.5–2× wall-clock today — MEASURED by the MoEUT authors, arXiv:2405.16039, not an in-repo measurement); "honest negatives at pretraining scale" (D5) is an unsolved data problem and may be the real bottleneck; §7.2 gaming risk grows with capability and is never fully closed — only bounded by binding and unannounced probes.

Alternatives considered

• Do nothing / serve-only (status quo per ADR-0010/0011/0021): keep the adapter-only

posture on the current base. Rejected by the operator's 2026-07-06 directive: the honesty axis is open at pretraining time and cannot be reached by post-training patches alone (TruthRL/R-Tuning are post-hoc; the objective-level fix needs token-one scoring).

• Raw-capability leaderboard chasing: rejected — reproduces the incumbents' 0-1-scored

objective that causes the confabulation problem this program exists to avoid.

• Learned-exit depth as the control axis: rejected — twice measured weak (MoEUT ACT

ablation + our Q-exit nulls); depth control moves to policy.

• **Honesty as marker-emission (train the model to say it's unsure):** rejected — gameable

surface behavior; the program binds honesty to external checks the model does not control.

• Unphased single-shot training run: rejected — every phase must be cheap to stop

relative to the next; kill-gates are program invariants, not options.

Open questions (to be resolved by the briefs, not asserted here)

D1 tiers and cluster shape; dense-recurrent vs MoE-UT (D2); the exact objective staging (D4); the honest-negative data pipeline at scale (D5); distillation retention threshold (D7).

Evidence

Amendment A1 (2026-07-27) — Accepted

Status: Accepted (Alex Place, in-session, 2026-07-27 — "approve the adr", closing #3002). Drafted by an agent per ADR-0001 and approved by the operator; the narrowing below is now the decision of record. It is recorded here rather than in a new ADR because it narrows this program's scope; nothing else changes.

What prompted it

docs/research/2026-07-27-in-house-model-spec-grounded-in-the-product.md measured the workload the product actually serves and found two things that bear directly on the Phase-1 objective:

1. Cost is not the reason to own a model. At the measured turn shape (~3.3k tokens in, ~121

out, history capped atturns), serving 10,000 users' ordinary chat on the cheap tier costs ~$5.9k/month — trivial against $20 Pro. The cheap tier already answers the common path in production (gemini-2.5-flash, gpt-4.1-mini). The whole cost curve is the escalation premium: one frontier turn costs ~37–67× a cheap one.

1. The team cannot pretrain a frontier model. The organisation is ~6 people. "Train frontier →

distill" is the right shape and the wrong first step at this size.

The proposed narrowing

Phasechanges from "pretrain a frontier model" to "post-train a strong open base into a verifier." Knowledge, writing quality and long context stay rented — that is the standing call ([[agi-convergence-blueprint-rent-capability-own-grounding]]:ofcapabilities rented, only Verify genuinely ours) and it is reaffirmed, not weakened, here.

Rationale, in the order that matters:

• Verify is the only capability we own. It is also the only task where we own the training

data: every verified trace the Spiral has run is a labelled example, and the trading surfaces generate ground truth on a schedule the market sets for free.

• It is what the escalation tier actually is. The expensive turns are the ones where the

answer has to be right — a judging problem, not a knowing problem.

• It is buildable at our size. Post-training a specialist on owned data is a 6-person project.

Pretraining is not.

Phase(distill to the ≤8GB artifact) and D7 are untouched; ADR-0026 still names the format, and its own Amendment A1 adds the deployment shape.

What would falsify this amendment

The escalation rate. If it is high, owning the escalation tier is worth serious money; if it is near zero, this whole program is worth less than shipping features. That number is now instrumented — lib/chat-escalation-meter.js, GET /api/metrics/escalation — and the first reading is not encouraging for the build case (below). This amendment should be re-examined, not assumed, once n ≥ 1,000 turns.