EPIC: the NAR ledger — compute Net Attention Returned from traces, not estimates

[ssdavidai]

Design epic, 2026-08-13. Implements the measurement behind Attentionmaxxing
so the monthly Attention Statement is computed from traces rather than
estimated in a prompt. Nothing here is committed work.

The number

NAR = accepted work displaced
      − prompting − review − correction − repair − AI-caused interruption

Measured monthly, in hours, against a baseline agreed up front. Only accepted
work counts. Every supervision minute subtracts. Rejected work scores zero and
still pays its supervision cost.

The thesis sets its own evidentiary standard — "trace-based estimates, hours
inferred from work signals, not a controlled experiment" — and is explicit that
receipts only count if you can argue with them. That is the bar this builds to.

What already exists — measured on the dev tenant, 2026-08-13

Term	Signal	State
AI-caused interruption	alfred_journal outbound by source_kind	exists: 1,422 HA replies, 18 voice-call, 10 cron, 1 system, 33 null
Review	time_to_decision_ms on decisions	field exists, populated on 42 of 500 (8%)
Prompting	Hermes main session store — messages(session_id, role, timestamp)	trace exists, nothing computes it
Correction	reversed_at on decisions	0 of 500 — never used
Repair	—	nothing
Accepted work displaced	1,566 decisions; commitments; hours approvals	no duration estimate anywhere

Decision intents over a 500-row sample: noise 307, done 80, defer 42,
delegate 33, take_mine 1, unset 37.

The missing atom

Nothing records what a displaced piece of work was worth. We can count 33
delegations; we cannot say what they would have cost the principal. Every other
term subtracts from an unknown.

One record, written close to the action: Alfred did X · it would plausibly
have cost the principal ~N minutes · the principal accepted or rejected it ·
here is the provenance. Everything else in the formula subtracts from that.

How it gets computed — the nightly recap

Sir's design, and it is better than a monthly reconstruction: a nightly job
reads the day's sessions and formalises what actually happened.

From one pass over the Hermes main session store plus the day's Desk activity
it can derive both halves:

• the counterfactual — what this task would have cost done by hand
• the real cost — start-to-finish wall clock through prompting, from message
  timestamps, and whether the thing actually completed

Doing it nightly rather than monthly matters: the sessions are still fresh, the
outcome is knowable, and a monthly reconstruction is exactly the "story" the
essay attacks.

Precedent for the read: #556's cron-journal reconciler already opens the Hermes
session store read-only from ctrl-api. Same posture — we do not own that
schema, we never write to it, and a missing column degrades rather than crashes.

Suppression — counted, but not by estimate

Sir's ruling: suppression does count. Filtering costs cognitive load and
preventing work is worth something. But it must not go soft.

So it is not estimated by the model. Suppression earns a fixed, stated,
conservative per-item rate, published on the statement as its own line rather
than blended into displaced work.

The arithmetic shows why. 307 suppressed items at a defensible 30 seconds each
is ~2.5 hours. The same 307 with a model guessing "about five minutes" is ~25
hours — a quarter of the monthly claim, invented. A fixed low rate cannot
inflate, is arguable in one sentence, and survives the sceptical question.

Interruption — solicited turns are excluded

1,422 HA conversation replies are the largest line in the journal and every
one is a reply to a turn the principal started. Counting them would swamp the
term and flatter the number.

Only unsolicited outbound counts: cron deliveries, proactive notifications,
anything Alfred initiated. On the dev tenant today that is roughly 10–40 items,
which is small and honest. The solicited/unsolicited split has to be explicit in
the data, not inferred at report time.

What counts as accepted

Two paths, per Sir:

1. Explicit — done or delegated through the Desk; a commitment reaching
   accepted; an approved hours proposal. Unambiguous, already recorded.
2. Inferred — the nightly recap concludes from the session that a task was
   carried out and completed. Weaker evidence, and it must be labelled as
   inferred on the statement, never merged with the explicit kind.

Rejected work scores zero displacement and still contributes its review and
correction minutes to the subtraction. That asymmetry is the point.

Storage

Per the four-store contract (CLAUDE.md §5.1):

• Per-action NAR entries → alfred-state.db. Machine working memory; the
  principal never edits a row. Needs a migration, therefore phase0.
• The monthly Attention Statement → a vault record. The principal reads it,
  argues with it, and signs it.

The signing pattern already exists and should be reused verbatim: #468's
hours-reconstruction produces an auditable per-period proposal and never
accepts hours automatically, at any confidence. The statement is the same
shape — Alfred proposes, Sir signs, nothing is claimed unilaterally.

Non-negotiables

• A defensible small number beats an impressive one. This is shown to
  someone deciding whether to keep paying. One inflated month costs more trust
  than a year of conservative ones earns.
• No baseline, no claim. If step 2's baseline was not captured, the
  statement says so rather than comparing against an assumption.
• Every line traceable to its evidence — which session, which decision,
  which commitment. A number you cannot drill into is marketing.
• Estimated and inferred lines are labelled as such, always.

Open questions

1. Where do counterfactual estimates come from? A model estimating "this
   would have taken 25 minutes" is the softest part of the whole scheme. Options:
   a per-task-class rate card agreed with the principal at onboarding; a model
   estimate the principal can correct, with corrections training the rate card;
   or model estimates shown but excluded from the headline until calibrated.
2. Repair has no signal at all. Either define one or drop the term and say
   so on the statement.
3. Correction is theoretically covered but unused — reversed_at is zero
   across 500 decisions. Is that because corrections are rare, or because the
   reversal path is not reachable from where mistakes are noticed?

Sequencing

#	Work	Lane
1	Populate time_to_decision_ms reliably — the one term that is 92% missing from an existing field	I
2	Solicited/unsolicited flag on journal outbound	I
3	NAR entry table + migration	phase0
4	Nightly recap workflow reading the session store read-only	II
5	Suppression rate card + fixed-rate line	II
6	Monthly Attention Statement as a signed vault record	I + II
7	Baseline capture at onboarding	III

Items 1 and 2 are small, independent, and useful before anything else lands.

[github-actions]

This issue is missing prioritization dimensions, so it will not
appear correctly in Now/Next/Later portfolio views.

#563 EPIC: the NAR ledger — compute Net Attention Returned from traces,
    missing:     kind, impact, effort, priority, horizon, customer, theme, area

0/1 open issues fully classified.

Rubric: docs/ISSUE-TAXONOMY.md

[ssdavidai]

Repair, defined — resolves open question 2

Sir's definition:

Repair is the back-and-forth between v1 and the last accepted version. Alfred
does the thing, I don't like it, we keep refining, eventually it's good
enough. If the first and last version are an hour apart, that hour is repair.

This is computable from the same session trace the nightly recap already reads,
and it gives a clean boundary against prompting, which was previously the
vaguer of the two:

prompting  = session start        → first substantive deliverable (v1)
repair     = v1                   → the accepted version

One split point, one pass, no new instrumentation. Both terms come out of
messages(session_id, role, timestamp) in the Hermes main store, which the
recap is already opening read-only.

Why this term matters more than its size

Repair is not just a subtraction. It is the leading indicator for the entire
thesis.

Ideally we will spend a ton of time with repair in the beginning and then that
will gradually reduce as Alfred gets better tools, skills and context.

That makes repair-over-time the falsifiable claim underneath the 30/60/90 arc.
If repair does not decline across a client's first quarter, Alfred is not
learning, and no amount of gross displacement makes the engagement work. It is
the one number that predicts whether the monthly statement will keep improving
or plateau.

So it should be plotted on the statement as a trend, not only subtracted as
a monthly total. A month with high repair and honest disclosure is a better
artefact than a month where repair quietly vanished into the aggregate.

The hazard: accepted versus abandoned

A thread that stops does not tell you why it stopped. "Good enough, done" and
"gave up, did it myself" look identical in the trace — same shape, same silence
at the end.

They must not score the same. Per the thesis, rejected work earns zero
displacement and still pays its full supervision cost. So an abandoned thread is
the worst case for NAR — maximum repair, no credit — and misreading it as
accepted would inflate the headline in precisely the direction that destroys
trust in the statement.

Distinguishing them needs a signal the trace alone may not carry. Options, in
order of how much they cost the principal:

1. Downstream evidence — the artefact appears in a send, a commit, a
   commitment moving to delivered. Strongest, free, but only covers work that
   leaves a trace.
2. The principal doing it again by hand shortly after — a strong abandonment
   signal, inferable from the following session.
3. Asking. One tap on the Desk the next morning: did this land? Cheap, and
   it is the same propose-and-confirm pattern P2: ship evidence-based hours reconstruction as a per-matter opt-in — propose, never auto-accept #468 already uses for hours.

Defaulting to "accepted" when unsure is the one option that must be excluded.
When the evidence is ambiguous the item should score zero displacement and
full repair — the conservative direction — and be listed on the statement as
unresolved rather than silently counted.

Consequence for the ledger

The per-action record needs three timestamps rather than one: task start, v1,
and accepted-or-abandoned — plus the outcome and how it was determined
(downstream evidence / inferred / asked).