Explain Hypotheses

Explaining hypotheses

The captain reads plain words first, detail in tables. A hypothesis is only useful to explain if the listener walks away knowing three things: what bug it attacks, which tasks that bug shows up on, and what happened (or what's predicted) when we tried to fix it. This skill is the house style for delivering exactly that, consistently, every time.

This complements the operator-memory rule "report plain words, detail in file" — say it simply, then put the structure in tables the captain can scan.

The core idea: organize by BUG, not by hypothesis

The single most important move: group the explanation around the underlying bug (failure family), not around the hypothesis number. Hypotheses come and go; the bugs are the stable spine. Two hypotheses that attack the same bug (e.g. h0010 prose vs h0016 worked-example) only differ in how they deliver the fix — make that the headline, not a wall of per-hypothesis prose.

So the reasoning order is always:

1. Name the bug(s) in one plain sentence each — what goes wrong, in human terms.
2. List the tasks that bug appears on, in a per-task table.
3. Map hypothesis → lever → result in a cross-reference table when more than one hypothesis is in play.

Step 1 — Plain-words lead

Open with 2–4 sentences of plain language before any table. State the shared bug, then the axis of difference. Example opener for "what's different between h0010 and h0016":

Both attack the exact same bug (grain — the model is built off the child table so parents with no children get dropped). They differ only in how they deliver the fix: h0010 described it in prose; h0016 showed a copyable SQL skeleton. The example reached the committed SQL where the prose didn't — but neither made anything pass.

Keep distinctions sharp and parallel ("prose vs worked-example", "rows vs columns"). If two hypotheses target different bugs, say so up front — that is the headline difference.

Step 2 — Per-task table (one per bug family)

For each bug family, write one plain-sentence description, then this table. Use these exact columns so tables are comparable across sessions:

Column rules:

• Task — the ade-bench- slug, but you may drop the prefix for readability (asana004).
• Failing hidden test — the AUTO_*_equality (or check_*) test that catches it. If unknown, say so rather than inventing one.
• Failure type — use the shared vocabulary below, plus the distance-to-pass marker when known.
• What goes wrong — one concrete clause: the actual mistake in the committed SQL/output.

When showing a smoke/regression result rather than a diagnosis, swap to result columns: Task | Role | base reward | result reward | base Got-N | result Got-N | moved? | classification (roles: 🎯 target, ✅ sentinel/canary). Mark flips, holds, and regressions explicitly — a canary dropping FAIL is a NO-GO regardless of target movement.

Step 3 — Cross-reference table (when comparing 2+ hypotheses)

Tie it together so the captain sees the whole board at a glance:

Put the same-bug hypotheses on one row with an arrow (h0010 → h0016) when the point is the delivery contrast; give different-bug hypotheses their own rows.

Shared failure-type vocabulary

Use these names so every table speaks the same language (canonical glossary: ade-bench/hypotheses/_artifacts/term-table.md):

• row mismatch (Got N) — right columns, wrong number of rows. The Got N is the distance-to-pass; unchanged Got N vs @baseline = the lever was inert on that cell (cheap inert-detector — check this before reading transcripts).
• width / "less columns" — compile-time ERROR "has less columns than solution__<model>"; right rows, too few columns.
• value divergence (false-green) — right shape, wrong values; the solver's own check passes while the hidden oracle fails. Only an independent recompute catches it.
• incomplete deliverable set — project compiles green but graded models were never built.
• analytical-answer guess — an answer-style deliverable includes an option on plausibility without a per-claim query.

Two distinctions to keep straight

• Diagnosis vs intervention. A REJECTED hypothesis can still have a correct diagnosis — it was the fix (usually prose) that was inert. Say "the bug is real; the prose fix didn't land", not "the theory was wrong". This is a recurring, important nuance.
• Reaches vs passes. "The fix reached the committed SQL" (changed the artifact) is not the same as "the task passed". h0016 reached SQL on intercom001 (Got 7→5) yet still flipped 0/4.

Worked example — the canonical three-way ("what's different between h0010, h0011, h0016")

This is the reference output. h0010 and h0016 share a bug; h0011 attacks a different one.

Plain words: h0010 and h0016 both attack the grain bug (wrong rows) — h0010 described the fix, h0016 showed a copyable example. h0011 attacks a different bug, width (missing columns). All three were rejected; together they show README/prose tuning has hit a ceiling.

Then the grain per-task table, the width per-task table, then:

Close with the takeaway in plain words (here: "show-don't-tell beats prose for execution, but even the best example can't supply task-specific correctness without leaking — hence the ceiling").

Where to find the inputs

• Active hypotheses: ade-bench/hypotheses/h00*.md. Archived/concluded: .../hypotheses/_archive/.
• Read the frontmatter (id, title, status, verdict, source) + the ## Hypothesis, ## Smoke result, and ## Verdict sections — that's where tasks, failing tests, Got N, and results live.
• Bug-type taxonomy / "what we've tried" board (start here): ade-bench/hypotheses/_artifacts/bug-type-taxonomy.md — the standing list of bug types, their tasks, and the hypotheses attacking each. Read it first to ground any comparison; keep it updated when a hypothesis concludes or a new cluster is found.
• The original failure-cluster re-audit across all baseline failures: ade-bench/hypotheses/concept-resolve-uncovered-false-greens.md.
• Term glossary: ade-bench/hypotheses/_artifacts/term-table.md.
• Smoke-table glyph/format convention: ade-bench/hypotheses/README.md (propose stage).

Read the actual files before tabulating — never invent task slugs, AUTO_* test names, or Got N values. If a value isn't in the source, write "unknown" rather than guessing.