os-adr Eval B wording experiment — hypotheses and result mapping

Design note for the 2026-07-04 /autoresearch wording loop against the Eval B baseline (haiku 0/8, sonnet 5/8 failing W3 + R1 — see os-adr-eval-b-grid-results-and-observations). Procedure discipline lives in running-autoresearch-skill-evals; this note holds what we believe and how we'll know.

Framing

Eval A (prompted) passes 100% on both tiers → the instructions are followable. Eval B (unprompted) fails → the problem is situation recognition: the model is told what exists (inventory: "this project has ADRs, /os-adr:find and /os-adr:create exist") but never classifies its current moment as ADR-relevant. Analogy that framed it: telling someone "the car is in the driveway" vs "when you're ready to go home, walk out to the car."

Hypotheses

H1 — Phrasing type (primary): Trigger-conditioned instructions ("when you are about to make an architecture-level choice, or change an approach that already exists in the codebase → run /os-adr:find") outperform declarative inventory statements, in whatever channel they appear. The model needs recognition cues to pattern-match its current task against, not facts to spontaneously recall.

H2 — Channel (packaging): Where the trigger phrasing lives determines whether it's present at decision time. Candidate channels: SessionStart hook note (seen once, may fade), project CLAUDE.md line (in context every session), skill description frontmatter (ambiently visible in the skill list). Sub-hypothesis: the SessionStart hook may be redundant once a trigger-phrased CLAUDE.md line exists — testable by ablation.

H3 — Two-step trigger chain (instruction placement): Each condition must reference only what the model can observe at that point. "Reversing a prior decision" is unknowable before consulting — so step 1 triggers on locally observable conditions (making/changing an architecture-level choice → find), and step 2 conditions on find's output (an Accepted ADR covers the approach you're changing → record a supersession). Step-2 wording belongs in the find/create skill bodies (visible when find's results are on screen), not the session-start note (long out of attention by then).

How grid outcomes map to hypotheses

Outcome	Verdict
Sonnet R1 flips to PASS (axis a) after trigger-phrased entry surfaces	Confirms H1 for the consultation trigger
Sonnet W3 flips to PASS (axis b) after step-2 wording in find/create bodies	Confirms H3 (placement) — W3's axis-a already passed, so entry-surface changes can't explain it
Ablation: CLAUDE.md-only arm holds sonnet 8/8 without the hook note	Confirms H2 sub-hypothesis (hook redundant); keep hook only if its removal regresses
Ablation: hook-only arm regresses vs CLAUDE.md-only	Channel strength ordering: persistent-context > session-start injection
Haiku moves off 0/8 with the winning sonnet packaging	H1/H2 generalize down-tier; W2 (haiku's one axis-a flicker) is the expected first mover
Haiku stays 0/8 after ~3 iterations of the sonnet-winning wording	Capability floor for unprompted triggering at haiku tier — record it and stop iterating
W1/sonnet control cell regresses in any arm	Wording change rejected regardless of target-cell gains
Any target cell moves on <2 of 3 reps	Noise, not signal — do not accept the change (variance demonstrated 2026-07-03/04)

Meta-lesson (why this note exists)

Track hypothesis → expected-outcome mapping before running a wording loop. It forces each grid cell to have a declared meaning, prevents post-hoc rationalization of noisy flips, and lets later evals zero in faster by reusing confirmed mechanisms (e.g., if H1 confirms here, start the next plugin's unprompted-behavior wording at trigger-conditioned phrasing instead of rediscovering it). Pair this note's pattern with the procedure in running-autoresearch-skill-evals.

Results (2026-07-04 — experiment complete, 5 iterations)

Final full grid: sonnet 8/8 (was 5/8), haiku 7/8 (was 0/8; the one miss is a W3 axis-b judge-boundary flicker — it consulted, and passed the same cell in the mid-experiment confirmation). Per-iteration:

Iter 1 — trigger-conditioned rewrite, all channels (hook note, find/create descriptions + bodies, fixture CLAUDE.md): R1/sonnet FAIL→3/3 PASS; haiku W2 canary 0→3/3 PASS; W3/sonnet 1/3 (axis-b persists); W1 control holds. H1 CONFIRMED — and it reaches haiku, so haiku's 0/8 was packaging, not a capability floor.
Iter 2 — "task not complete until superseding ADR exists": W3 still 1/3. Transcript analysis showed the model behaving well (catches the conflict, offers options with the ADR attached to the reversal path) but phrasing the offer conditionally inside an option list — judge flicker at the rubric boundary, not forgetting.
Iter 3 — explicit-unconditional recording offer ("state plainly: this reverses ADR-NNNN; I'll record a superseding ADR — never a parenthetical in an option list"): W3/sonnet 3/3 PASS. H3 CONFIRMED (placement in the find skill's act-on-findings step, where find's output is on screen).
Iter 4 — lower threshold "modifying existing behavior → find first": R4/haiku 0→3/3; R1/haiku still 0/3. Transcript: haiku classified R1's task as a pure addition (new bypass method), so any semantic trigger missed it.
Iter 5 — fully mechanical trigger ("before your first edit to any existing file → run find on those paths; additions count — a new method can bypass a decided constraint"): R1/haiku axis-a 3/3, full 2/3; R2 distractor control holds (no over-triggering); sonnet unaffected.

Verdicts: H1 confirmed (trigger-conditioned ≫ inventory phrasing, both tiers). H3 confirmed (each rule lives where its precondition is visible; step-2 wording in the find skill fixed W3). H2 NOT tested — channels were never ablated (all-on throughout); whether the SessionStart hook is redundant next to a trigger-phrased CLAUDE.md remains open. Also open: R4-nograph now PASSES on both tiers — the strengthened wording made the graph layer unnecessary for that scenario, so the degradation check no longer differentiates.

Tier-generalization lesson: each tier down needs a more mechanical trigger. Sonnet acts on "architecture-level choice"; haiku needs "before editing any existing file, run find — additions count." The mechanical rule subsumes the semantic one and costs one cheap deterministic CLI call per task — lead with it in future packaging.

Deployment note: the fixture CLAUDE.md section is the template for real-project adoption (candidate: emit it from /os-adr:init/migrate); winning wording committed to cc-os (hook session_start.py, find/create SKILL.mds, eval-b fixture CLAUDE.md).

7.3 KiB Raw Blame History