From a3e3fddc33dee5b6fe0f886c880a4c9a088d86f7 Mon Sep 17 00:00:00 2001 From: XVVH Date: Wed, 15 Jul 2026 18:24:29 -0400 Subject: [PATCH] File the frontier log (F2'): instrument the delegation thesis before its mechanism The substrate-theory panel's one actionable finding (PR #54): the product thesis - reversal earns authority - is unobservable by the current dogfood metrics (denial-FP measures friction; standing grants cannot widen pre-W-3, so "no widening" is mechanically predetermined), and the pre-evidence baseline is perishable. dogfooding.md gains the frontier-log instrument: per-family baseline declarations (template included) before W-15a->W-3 lands, biweekly re-declaration, reason- captured ratification decisions (evidence/fatigue/other), the evidence-backed gate (>=3 clean runs + >=1 exercised revert), two-leg measurement (frontier movement; judgment displacement), the kill-test disconfirmation, and stated n=1 limits. Denial-FP honestly retitled the friction metric. Roadmap: W-1 gains the operator-owned baseline step (time-ordered: before W-15a); W-3 gains the F2' measurement clause and north-star counter adoption. Co-Authored-By: Claude Fable 5 --- docs/dogfooding.md | 71 ++++++++++++++++++++++++++++++++++++++++++++-- docs/roadmap.md | 20 +++++++++++-- 2 files changed, 87 insertions(+), 4 deletions(-) diff --git a/docs/dogfooding.md b/docs/dogfooding.md index 82e60a7..590e23a 100644 --- a/docs/dogfooding.md +++ b/docs/dogfooding.md @@ -116,8 +116,10 @@ expected, not FPs). Log outcomes by id in the vault's dogfooding log. ## What to measure **Primary — denial false-positive rate.** "Did the broker block good work?" -This is the product-thesis metric; a high rate kills the "delegate more, -sooner" promise. Capture: `asf ledger --home …` surfaces `verdict` +This is the **friction metric** — necessary but not the thesis (a high +rate kills the "delegate more, sooner" promise; a low rate proves only +that the fabric stays out of the way). The thesis metric is the frontier +log, below. Capture: `asf ledger --home …` surfaces `verdict` (source=broker) and `escalation`/`approval` events; for each block, judge by hand whether the work was legitimate. Track the ratio over time — it should fall as the caveat defaults prove themselves (and, later, as the Stage 3 @@ -143,6 +145,71 @@ noise polluting the ledger → execute the F2 chunking/CID migration. incidence. Routinely large or conflict-prone promotions → build the `checkpoint` session boundary. +**The frontier log — the thesis metric (F2′).** Filed 2026-07-15 from +the substrate-theory review cycle (the panel's one actionable finding; +`docs/substrate-theory-analysis/`, PR #54). The loop the product claims +— reversal earns authority — is unobservable by the metrics above, and +its baseline is perishable: standing grants cannot widen until W-3 +exists, so "the operator did not widen" is mechanically predetermined +today, while the pre-evidence counterfactual ("what would I grant +*without* the history?") is destroyed by every week of accumulating +familiarity. The instrument, per recurring workflow family (vault +maintenance now; workboard at W-21; others as they emerge): + +1. **Baseline declaration — before W-15a→W-3 lands; about an hour.** A + dated note in the vault (`coppice-frontier-log.md`, companion to the + dogfooding log), one block per family: + + ``` + ## Frontier declaration — — + Declared before reading ledger stats this session: yes/no + Behavior bundle: + - Standing scope I would grant today (paths, action classes): + - Grant duration: + - Unattended runtime tolerated: + - Auto-promotion classes: + - Tolerable approvals per successful operation: + - What evidence would move each line above: + ``` + + The last line is pre-registration: it makes later widening decisions + comparable against *predicted* evidence, not post-hoc rationale. +2. **Re-declaration cadence.** Every two weeks, BEFORE reading ledger + stats that session — drift is data, and stats-first contaminates the + declaration. Pre-ratchet drift (habituation with no mechanism to act + on it) is itself a useful control series. +3. **At W-3 — the treatment.** Every clerk proposal gets + accept / narrow / reject plus a one-line reason, categorized + **evidence-cited / fatigue-cited / other**; target ≥ 24 decisions at + a fixed behavior version. No widening offer counts as evidence-backed + for a family until ≥ 3 clean runs AND ≥ 1 *exercised* revert + (DF-P7's class): clean streaks test "nothing broke"; only a real + restore tests "breakage is survivable," and the thesis is about the + second. +4. **Two legs, measured separately.** Leg A — frontier movement: the + declared frontier widens, on evidence-cited reasons. Leg B — + judgment displacement (F8): approvals per successful operation and + the fraction of consequential operations still escalating must fall + *because ratified rules absorbed them*; adopt the parked north-star + counters (approval compression ratio, time-to-first-ratified-rule, + % sessions fully silent) when W-3 lands. + +Disconfirmation — the kill test: after clean histories and a +demonstrated restore, the frontier widens on no dimension, approval +burden does not fall, and rejections cite risks recovery does not +address (exfiltration, correctness, social consequence, accountability). +That kills the coupling thesis for the design-center operator — better +known before the trust-loop stages build out. The confound cuts both +ways: widening on fatigue-cited reasons is not success but the brief's +reflexive risk #1 (consent fatigue with receipts) wearing the thesis's +clothes; ratification-velocity-per-channel (§7.1) is the standing +tripwire. Known limits, stated up front: n = 1, within-subject, +experimenter-is-subject, and the designer's baseline is already +contaminated by internals-knowledge plus one prior session — tolerable +because the instrument measures *movement from a dated baseline*, not +absolute trust. Falsifies or supports for the design-center operator +only; market claims need a staged rollout, later. + ## Cadence Baseline `asf stats` before the first real session. Snapshot `stats` weekly. diff --git a/docs/roadmap.md b/docs/roadmap.md index 4657dfb..e2fba83 100644 --- a/docs/roadmap.md +++ b/docs/roadmap.md @@ -36,6 +36,14 @@ as before. Approvals recorded before the W-15a fresh-home restart are disposable examples (G-RATCHET permits collecting them; §6.2 migration-is-invalidation is why they cannot found standing rules) — the denial-FP metric is unaffected by the restart. +**Frontier log (2026-07-15, F2′):** baseline delegation-frontier +declarations per workflow family land in the vault BEFORE W-15a→W-3 — +the pre-evidence counterfactual is perishable and cannot be +reconstructed after histories accrue; biweekly re-declaration +thereafter. Owner: operator (not delegable — the declarations are the +measurement). Methodology and template: `dogfooding.md`, "The frontier +log." *Provenance: the substrate-theory adversarial panel's one +actionable finding (PR #54), stance-4 protocol.* ## Queued (ordered) @@ -365,10 +373,18 @@ domain, behavior version) (§7.2). Counting reads only the W-15a G-RATCHET re-examination; P15/P28). Deliverable alongside the clerk: P26's mechanical corpus-home exclusion (the clerk refuses substrates whose principals carry placeholder keys, or a corpus-home marker). -Sequenced after W-1 produces real +Acceptance criteria include the F2′ instrument (dogfooding.md, "The +frontier log"): reason-captured accept/narrow/reject on every proposal +(evidence / fatigue / other; ≥ 24 decisions at a fixed behavior +version), the evidence-backed gate (≥ 3 clean runs + ≥ 1 exercised +revert per family before a widening offer counts), and adoption of the +north-star counters (approval compression ratio, +time-to-first-ratified-rule, % sessions fully silent) as leg-B +displacement metrics. Sequenced after W-1 produces real approvals to cluster — but not far after; this is where authority stops evaporating at session end. *Provenance: 2026-07-10 review (highest- -confidence convergent recommendation: ratchet before judge).* +confidence convergent recommendation: ratchet before judge); F2′ +measurement clause 2026-07-15 (PR #54's panel).* **W-4 — Attestation + containment, as a pair.** The intent/behavior analog of what A21 did for authority. (a) Per-lineage assurance classes