RubyEyedReaper Learn — what is worth remembering, how sure we are, and when it stops being true
A record earns its place by clearing a capture bar, states its confidence as a rung on a ladder rather than a feeling, and carries the command that would re-check it — so a later session can re-run the evidence instead of taking the record's word for it.
This skill adds no store. The memory tiers already exist and are described in
.claude/docs/odin-memory-standards.md. What it adds to a
record in one of them is two fields and a discipline for moving them.
What this is not for
| The task | Its owner | Why not this skill |
|---|---|---|
| Where a record physically lives, and which class it belongs to | odin-memory-standards.md + odin-memory-guard.sh |
The tiers, the operational/task split and the write guard are settled. This skill is a reader of them. |
| Recording a decision and the reasoning behind it | architecture-decision-records |
An ADR is a decision record. This is about knowledge that was discovered, not chosen. |
| Domain vocabulary and its definitions | domain-modeling |
CONTEXT.md is the glossary. A term is not a finding. |
| A mistake that needs a mechanical guard | oops, then mistake-to-gate |
Those turn an incident into a check. This one decides whether the knowledge is worth keeping. |
| A principle recurring across skills, headed for a rule file | rules-distill |
A rule is always-on text. A record is recalled on demand. |
| Context bloat, cache buildup, memory surfacing from the wrong project | leek |
Those are hygiene failures in the environment, not questions about a record's truth. |
| What to work on next | roadmap |
Knowledge is not a backlog. |
The two fields
Added to a record in a tier that already exists:
confidence: gated
verified_by: bash .claude/tests/safety-guard.test.sh
verified_on: 2026-08-27
verified_by names a command, not a person and not a date. A date says somebody looked. A
command says what would have to be re-run to look again — the difference between a record that can
be re-checked and one that can only be trusted.
The always-loaded memory index has a 300-character-per-line ceiling (memory-index-check.sh,
DEC-0053) because it is read whole into every session in its scope. So the index line carries the
short half only — · gated 2026-08-27 — and the command lives on the record it links to, which is
unbounded on purpose. The engine renders that marker; it is never hand-assembled.
The six outcomes
Exactly one per pass over a candidate or a record.
| Outcome | Means |
|---|---|
record |
The candidate cleared the bar; write it into the tier it belongs to |
merge |
An existing record already covers it; fold the new evidence into that one |
promote |
New evidence lifts an existing record a rung |
refresh |
The verifying command still passes; the rung holds, the date moves |
demote |
The evidence for the current rung no longer holds; the record falls one rung |
retire |
The premise is gone; remove the record and say why |
retire deletes knowledge on purpose. A record kept past its truth is worse than no record — it is
read with the same confidence as one that is still correct, and nothing about it says otherwise.
The five confidence tiers
asserted → observed → reproduced → gated → enforced. Every rung states the evidence that
promotes a record into it and the trigger that demotes it out; a ladder whose every rung is
a feeling is a vocabulary, not a mechanism. Rung by rung, with the promotion evidence each one
demands and the mistake each rung invites:
confidence-tiers.md.
The three rungs from reproduced up require verified_by. A record claiming one without a
command is malformed, not merely unverified.
The capture bar
Five refusal classes, evaluated highest-harm first, each with its own exit code so a caller can tell them apart — five refusals sharing one non-zero code report "no" and nothing else, and the remedy for a secret is not the remedy for a duplicate.
| # | Class | Exit | Refused because |
|---|---|---|---|
| 1 | credential | 3 | Secret material — matched by key name, or by value shape |
| 2 | personal | 4 | Personal data — an address, an identifier, a phone number |
| 3 | chatter | 5 | True only inside the conversation that produced it |
| 4 | duplicate | 6 | An existing record already covers it → merge |
| 5 | derivable | 7 | A restatement of a line the repository already carries |
A refusal never prints the matched credential. Not in the message, not in a verbose mode, not in a fixture's expected output. The caller is told the key name or the pattern name — what they need in order to remove it — and nothing more. A refusal that echoes the secret has published it to every log the caller writes, which is worse than the record it prevented. Masking is not a substitute: a masked secret is still a secret in every one of those logs.
What each class is keyed on, what it deliberately does not catch, and the false-positive it was tuned against: capture-bar.md.
The staleness pass
verify re-reads a record and decides its rung from the command, not from the calendar. Age
flags; the command decides. A record past its max age is marked stale — which means "re-check
this", never "this is wrong".
Reading a record never executes the command inside it. Execution is opt-in via --run; the
default reports what would run and changes nothing. How a demotion is chosen, why an unresolvable
command falls to observed rather than to nothing, and when a pass ends in retire:
staleness-pass.md.
Quick reference
cd .claude/skills/learn
python3 -m scripts.capture_bar check --candidate cand.json --root ../../.. --corpus <dir>
python3 -m scripts.capture_bar verify --record rec.json # reports; runs nothing
python3 -m scripts.capture_bar verify --record rec.json --run --cwd ../../..
python3 -m scripts.capture_bar tiers
| Exit | Means |
|---|---|
| 0 | accepted, or a pass that held the tier |
| 1 | malformed record |
| 2 | usage error |
| 3–7 | refused: credential · personal · chatter · duplicate · derivable |
| 8 | the pass demoted the record one rung |
| 9 | the pass retired the record |
--corpus and --root are optional, and a class without its input is reported "skipped" on
stdout and announced on stderr. A check that quietly passes when its subject is missing is
indistinguishable from one that agrees with every input.
Common mistakes
| Mistake | What happens |
|---|---|
A verified_on with no verified_by |
The record can be aged but never re-checked. Nothing can ever demote it, so it stays at its rung forever. |
Claiming gated because a check exists |
gated requires that the check fails when the fact stops being true — proved by breaking the fact and watching it red. An unmutated gate is an observed. |
| Masking a secret instead of withholding it | Still a secret, now in every log. Report the key name. |
| Refusing every finding as "derivable" | The class is keyed on restatement of one line, not on inference. A fact assembled from three files is a finding. |
Reading a record with --run by habit |
Reading a record then executes text stored in it. Default to the reporting pass. |
| Treating a stale flag as a demotion | Age says re-check. Only the command decides the rung. |
| Keeping a record whose premise is gone | It is read with the confidence of a true one. retire it and record the reason. |
Verifying a change to this skill
cd .claude/skills/learn && python3 -m unittest discover -s tests -t . -v 2>&1 | tail -5
Confirm the collected count is non-zero. tests/__init__.py is what makes it so; without that
file discovery collects nothing and the suite passes having examined nothing.