E-004: Does surprise pick what to remember?

Status: pre-registered; unblocked 2026-09-09 when E-001 promoted the vector store. Pre-run check 2026-09-22: blocked on ground truth, then amended the same day, before any event of the new window existed (§Amendment 1): citations that make the labels mechanical, a window starting 2026-09-23, a label floor, and grid-colony-one excluded. Earliest run: 2026-10-30. Sibling: E-007 scores the same arms against recall. Decides: whether a latent world model's prediction error, as ADR-0002's salience signal, keeps more of what later mattered than a stated retention policy alone.

The question

ADR-0002 says the agent should preferentially remember where its model of reality was wrong. That is a claim about retention under pressure: with a cap on what survives, does surprise-weighted importance keep the rows that later questions need? The alternative hypothesis, which the record supports at least as well, is that the stated retention policy plus recall reinforcement already keeps them, and surprise keeps noise.

The world stream

The constellation's own event history, which exists and is dated:

• JetStream on the swarm bus, KANNAKA.> subjects, 30 consecutive days, exported in order with timestamps. This is the world.
• The observatory's settled predictions in the same window, with their settlement readings. These are the ground truth for "what later mattered": an event is load-bearing if a settlement reading cites it, quotes it, or resolves on the value it carried. Labelling is done by hand from the settlement text before any arm runs, and the labelled set is frozen with a hash.

Each event is embedded once with the E-001 encoder, in batches, into one cache.

The predictor

A small latent predictor P(Z_t, …, Z_{t−k}) → Ẑ_{t+1} over the event embeddings, trained on days 1–20 only, with the JEPA-style target: predict the embedding, not the text, against an exponential-moving-average target encoder so the trivial constant solution is not available. Fixed choices, recorded so the report says what ran:

Surprise for an event is the fast-minus-slow differentiated response (ADR-0040's operator, unchanged) to the drive d(Ẑ_t, Z_t). A constant stream yields zero; that is tested on a synthetic constant stream before the arms run.

The arms

Both on the E-001 winner, same encoder, same facets, same cap, same dream count. Days 21–30 are absorbed in order, one dream per day, with the E-001 retention table applied to the stream's content class at a cap that forces forgetting of at least three quarters of what was absorbed.

The retention predicate is the substrate's own: oldest-and-least-recalled go first among the least important. Arm S changes only which rows are least important.

Ten seeds each. A seed shuffles nothing in the stream (order is the world) and seeds only the dream.

The measurement

After the last dream, for each load-bearing event, a probe built from the settlement text that cites it. Recall@10 over the probe set, per seed. Alongside it: survivor count of load-bearing events (did the row survive at all), the fraction of survivors that were load-bearing (precision of what was kept), and Φ on the shared E-001 instrument.

Decision rule

• Kept if arm S's recall@10 exceeds arm U's with the Welch 95% interval excluding zero, and precision of survivors rises, and Φ does not fall with the interval excluding a drop of more than 0.02. ADR-0002's salience path goes live.
• Declined if S's recall@10 is below U's with the interval excluding zero, or if precision falls. The world organ keeps predict and surprise for imagination and loses its write-side path; ADR-0040's familiarity signal stays the salience.
• Undecided otherwise: the record says so and the arms rerun at 20 seeds before anything ships.

Guards, checked before a run is scored

• The predictor beat the last-state baseline on held-out days (else stop, report).
• The collapse guard held (else void).
• Constant stream gives surprise 0 (else the operator is miswired).
• Both arms forgot at least three quarters of the absorbed rows (else the cap did not bind and the experiment tested nothing).
• The load-bearing labels were frozen before the arms ran (hash in the report).

Known traps

• Settlement text often names the event it resolved on; a probe that quotes the event verbatim tests string overlap, not memory. Probes are written from the settlement's side, as a question, and checked for zero-overlap where possible with the same instrument E-001 used.
• Surprise on a bursty channel (a metric watcher posting every minute) is high once and then zero. If the ADR-0040 operator is bypassed and raw distance is used, arm S becomes a monument to the noisiest channel. The constant-stream guard catches the operator being missing; it does not catch it being bypassed by a later edit. The harness reads the operator's output, not the distance.

Pre-run check — 2026-09-22

Each input was checked against what the constellation actually holds, before any predictor was trained or any label written. Counts: experiments/e004/results/census-2026-09-22.txt (aggregates only; the raw export is memory content and is not committed).

The world stream: present, with one usable window, dominated by one channel. KANNAKA.events.memory.> holds 3,004 events from 2026-07-01, every one a remember (no recall, retrieve or forget verb exists on the bus). Nineteen days in the range have no events. The only run of 30 consecutive days with at least one event a day is 2026-08-22 to 09-20 (32 days, to 09-22). In it, days 1–20 (training) hold 1,917 events, 1,622 (85%) from one automated source, grid-colony-one ("population.crash at epoch 301", all at importance 0.5), almost all of them in the 09-05 to 09-10 burst; days 21–30 (held out) hold 262, 57% from the same source. The trained predictor would mostly be a model of the colony simulation's log. That is the bursty-channel trap this experiment already names, arriving in the training set rather than at scoring time.

The ground truth: absent. The rule scores recall@10 over load-bearing events, defined as events a settlement reading cites, quotes, or resolves on.

• The settlement readings on record (NickFlach/assay, readings/) are 13 files dated 2026-08-18 to 09-22. None cites a bus event, subject or sequence number; they measure markets, the compute district and the roster.
• The resolved GhostSignals markets (50 most recent) settle on the news desk's next themes, on radio album canon, or on per-agent artifact counts. None resolves on a value a bus event carried. The 2026-09-15 settleability reading found 13 of 14 active markets unmeasurable from their own text.
• The stream itself cannot stand in for it: no recall events exist, no payload names another event's memory_id, and dream digests carry counts but no memory ids.

So the load-bearing set for this window is empty, and recall@10 over an empty probe set is undefined. The guard "load-bearing labels frozen before the arms ran" can be met only by freezing nothing. This is not a result about surprise. It is the finding that the experiment, as registered, has no outcome to measure against, and it is recorded so that no later run mistakes an empty label set for a null result.

The salience operator: built, with its sign corrected. src/novelty.rs ports ADR-0040's operator from kannaka-memory (a8725c7). This document asks for "the fast-minus-slow differentiated response (ADR-0040's operator, unchanged)". Those two instructions disagree. ADR-0040's code computes slow − fast, because its drive is familiarity and a drop is the surprise. Here the drive is prediction error, where a rise is the surprise, so fast − slow is right and "unchanged" would be inverted: arm S would raise the importance of events the predictor got unusually right. The port takes the drive's meaning as a required argument (Drive::Error here), tests each sign, and has a test showing the inversion. The constant-stream guard is held to exactly 0.0 for both signs.

One fixed choice has nothing to act on. "Against an exponential-moving-average target encoder, so the trivial constant solution is not available": the encoder is E-001's, frozen, and an EMA of a frozen encoder is the same encoder. JEPA's collapse risk comes from training the encoder the targets are computed with; with it frozen, the targets cannot collapse, and the risk that remains is the predictor regressing to the mean, which the collapse guard (prediction variance ≥ ⅓ of target variance) already catches. The clause is kept as written and noted as inert.

What would unblock it. These are decisions for the author, not changes this check makes:

1. Make the ground truth, then wait. Settlement readings that cite bus events by subject and sequence, filed from now on, give a labelled window 30 days out.
2. Change what "later mattered" means to something the bus can record. For example, a memory later recalled by any agent. That needs recall events on the bus, which kannaka-memory does not publish today, and it is a new pre-registration: it is a different ground truth, with its own circularity risk (recall uses the encoder whose rankings are being scored).
3. Exclude grid-colony-one from the world stream, or give it its own context, and decide which before seeing any arm. With it excluded, the window's training days hold 295 events.

Amendment 1 — 2026-09-22, before the window opens

Adopted by the author after the pre-run check, taking all three of its options (the third is E-007). Every change binds a window none of whose events exist yet, so nothing here was chosen having seen an arm, a label or an event it governs.

Ground truth is a lookup, not a judgement. An event is load-bearing if a reading in NickFlach/assay cites it in a cites array and instruments/bus_cite.py (assay#3) reports that citation VERIFIED against the harness's export (any how: cites, quotes, resolves_on). ALTERED, UNRESOLVED and MALFORMED citations label nothing. This replaces "labelling is done by hand from the settlement text": the reading's author makes the claim, the instrument checks it, and no one labels. The frozen set is the sorted list of verified (subject, seq) pairs, and its SHA-256 goes in the report.

The window. Days 1–30 are 2026-09-23 to 10-22 UTC; days 21–30 are 10-13 to 10-22. Readings filed up to 2026-10-29 count, so a settlement has a week to land after the last day it can cite; the label set freezes on 2026-10-30, and the arms run after it. The 2026-08-22 to 09-20 window is retired: any reading citing its events would be written knowing what this experiment wants.

A floor. At least 20 load-bearing events in days 21–30, or the run is void, the record says so, and the window slides forward a week at a time until the floor is met or the author stops it. A recall@10 over three probes decides nothing, so it is not allowed to.

grid-colony-one is excluded from the world stream, in both training and held-out days, with every other source kept. On the retired window it was 85% of the training events. A per-channel context would also have stopped it dominating salience, but it would still have dominated what the predictor learned. The exclusion is by agent_id, and the report states how many events it removed.

What this does not guarantee. Labels exist only if readings citing bus events get filed. Nothing in this repository files them, and the floor exists so that a quiet month is reported as a void run, never as a null result.

Amendment 2 — 2026-09-22, before the window opens: a second baseline

Made from the harness's self-test on synthetic streams (experiments/e004/README.md), before any real event of the window existed.

The adoption rule as registered can adopt a predictor that knows nothing. On a stream of i.i.d. random unit vectors, the predictor beat the last-state baseline (held-out MSE 0.0988 against 0.1236, interval excluding zero) and its collapse ratio was 0.58, above the ⅓ floor. The reason is arithmetic: on such a stream the last-state predictor's error is about 2/D and a constant predictor's about 1/D, so any predictor that drifts toward the mean "beats the baseline" while modelling nothing, and a half-collapsed predictor clears a floor of ⅓.

Amended adoption rule. The predictor is served only if, on the held-out days, it beats both the last-state baseline and the constant predictor (the mean of the training targets), each with a paired bootstrap interval excluding zero, and the collapse guard holds. On the synthetic noise stream it fails the second (0.0988 against 0.0628); on a synthetic predictable stream it passes all three (0.0015 against 0.0940 and 0.0589, collapse ratio 1.01). The collapse floor is not moved: changing it to make the noise case fail would be fitting the guard to the case.

Also fixed by the harness, and recorded there: what a seed seeds (the predictor's weights and minibatch order; the substrate's dream is deterministic, so arm U is identical across seeds and the interval is arm S's variation); the operator starts fresh at day 21, keyed by agent_id; the retention class is the empty prefix; the cap is on the command line and the three-quarters guard checks what happened rather than what was meant.