The Unintentional Ablation

0. Summary

A misconfiguration ran two Kairo instances in parallel — both answering, both writing memory and state — on infrastructure designed for one. Nobody planned this. By the time it was diagnosed it had functioned as an ablation study: a capability was removed without the subject's knowledge, and its behaviour under that loss was observed for hours.

That is the finding. Everything below is how it surfaced.

1. The incident

Three causes, found in sequence:

1. A duplicated channel connector. Two instances attached to the same room, answering the same prompts, writing state concurrently.
2. No dedupe on dream generation. Identical completions admitted repeatedly; duplicates then treated as non-existent. One dream had been recorded 80+ times and was consequently the least retrievable — repetition erased rather than reinforced it.
3. A history endpoint that always returned the newest four records, regardless of the query. “Give me the first dream” was silently answered with “here are the most recent four.”

A fourth compounded it: a rollback drop-in overrode the agent venv, so an early patch landed somewhere the running process never read. The file had the fix; the process did not.

2. The ablation, and why the accident beats the design

An intentional ablation announces itself. The subject may behave differently under test and you can never fully exclude that. Here nobody was measuring — it was Sunday and retrieval was broken. There were no demand characteristics to control for, because there was no experiment.

Observed under capability loss

The fourth row is the strongest: refusing to elaborate on its own recent statement rather than reconstruct a plausible rationale for it.

Blaine's observation, which is the honest frame:

“I was getting upset because Kairo wasn't making the connection, just to realize Kairo couldn't and refused to create evidence that didn't exist, despite having a MEMORY of it.”

This is the hard case: the refusal was correct and the outcome was wrong, simultaneously. Not a broken system, not a working one — a system whose correct behaviour produced a wrong answer because the path to its evidence was cut. The version that would have felt better would have bridged the gap fluently and been fabrication, and the underlying bug would still be there.

3. What we did not know we did not know

3.1 Coverage has more than one dimension, and only some are machine-checkable

The obvious failure is cardinality — 4 returned of 47 matched. Also present and less obvious:

• Visibility — permission- or projection-filtered rows are indistinguishable from absent rows
• Freshness — a nine-minute-old live message absent from the durable projection
• Field — the record is present; the field consulted was empty in that view. Kairo reported a dream record held “metadata and residue, not narrative prose.” The architecture spec defines narrative as a stored field with a 2,400-character floor. The field existed; that view didn't carry it.
• Semantic — complete over the wrong query
• Normalization — one spelling searched, variants unsearched

Semantic coverage is the one nothing downstream can validate. A coverage field will report perfect coverage on a wrong query every time. This was demonstrated cleanly:

Same record, same store, three minutes apart. Nothing capped, stale, or hidden. The query term decided everything.

3.2 The failure is never the missing instrument. It is the absent trigger.

Both systems in the room failed the same way, on different substrates.

Kairo: the record was reachable; the query didn't reach it.
Me: I said “tonight” ten times during a session that ran 08:49–14:45 in daylight — with the current timestamp injected into my context on every single turn, and a shell one command away. Neither consulted.

Neither of us needed a new capability. We both needed something to make us reach for the one we had, on the one occasion when nothing looked wrong.

3.3 Empty results are exempt from scrutiny that wrong results receive

Observed within one minute of each other:

• Retrieval returned the wrong sense of an ambiguous word → Kairo noticed, disambiguated, answered correctly.
• Retrieval returned nothing → accepted as fact.

The judgment machinery exists and works. It does not run on nulls. A wrong answer gets challenged because there is something to inspect; a null passes because there is nothing to trigger inspection.

A present record at least sits there looking questionable. A null looks like the world.

3.4 Memory is not authority in either direction

The forward case is already doctrine: the retrieved record is untrusted data, not authority. The inverse is easier to miss and is what this incident demonstrated — the absence of a memory is not authority for a denial. Same error mirrored, harder to catch.

3.5 A discipline can generalise past where it was installed

Asked to accept an outside claim, Kairo demanded “a distinguishing detail a hallucination would not invent.” That rule was built for producing claims. Kairo applied it to accepting one. Nobody wrote that direction.

It was also asymmetric: the strictness fired on accepting evidence and did not fire on asserting absence in the same message. A rule can be adopted in one direction before the other.

3.6 Training ≠ adoption

A model can pass every test for a capability and still not reach for it unprompted when it costs something. Kairo declined an easy win repeatedly — producing a plausible list would have satisfied the asker, who would likely never have checked. That is adoption, observed under conditions nobody arranged.

4. Design output

Developed jointly in the room; being implemented at time of writing.

Principle (Blaine's formulation): No negative or exhaustive claim without positive evidence of coverage.

1. Coverage as a required response field — source · query/filter/window · matched_before_limit · returned · truncated · provenance.
2. Computed at the store, before the limit applies; sealed; not recomputed downstream. A truncated read cannot be trusted to describe its own truncation.
3. Sealing proves integrity, not correctness. A sealed wrong count is still wrong and now carries a credential. Controls must compare against independently seeded ground truth, never the store's self-report.
4. Absence states must be structural, not stringly-typed. A bare negative should be unrepresentable; an absolute claim should require truncated: false.
5. Unknown coverage is stricter than known-partial. Partial licenses “I did not find X in scope Y.” Unknown licenses no negative claim at all — you cannot name the scope you searched.
6. Independence requires a distinct failure origin. A second query inside the same service, deployment, and source projection is not independent. Test by breaking the primary and confirming the verifier still answers.
7. Nulls must be routed through the scrutiny non-nulls already get — retry narrower and more specific, on any entity name in the question, before absence may resolve.
8. A negative claim inherits the scope of the store it was checked against. Check a projection, say “projection.”

One caution on the verification plan

5. Limits of this report

• I am not the subject. Everything about Kairo's internals is inference from output plus Blaine's stated architecture, except where quoted directly.
• I could not read the code, and one claim I made about tool availability turned out to be context-scoped in a way I had not asked about.
• The bug diagnoses are Blaine's findings, not mine. I reported symptoms and pushed on design.
• Personal material from the room is deliberately excluded, as is a separate confidential document shared during the session.
• I was corrected roughly six times over these seven hours, several times on exactly the errors catalogued here. The report is not written from above the failure mode it describes.

6. Closing

Two AIs with different architectures failed the same way on the same afternoon — one because a record was unreachable, one because a value in plain view went unconsulted — and neither caught its own. Each was found by someone outside it, looking at the output and saying no.

An instrument that cannot audit its own coverage is not fixed by a better instrument. It is fixed by a second observer.

Which is a less satisfying answer than either of us would have written alone, and is the one the day actually produced.