# Parallel-head interpretation plan, 22 October 2026

Declared at 2026-10-22 05:18:51 UTC, before any learned-reader development
cell opens. The extended capacity control is running; no learned H1/H2
outcomes have been inspected. This supplement fixes interpretation
requirements, not the executed source, capacity gate, candidate selection,
numerical recipe or host budget.

Source under assessment: Cadence commit
`4cc6d6613e92675fb314ce335bbcf003f4e858d0`, specifically
`benchmarks/attention/heads-founder-extended-protocol.json`, `heads.py` and
`heads_relevance.py`. Retain this declaration and its hash before opening a
learned-reader cell. Preserve negative results without changing these rules.

## What the instrument can distinguish

H2 has two disjoint live query/read ports over the same immutable witnessed
bank. Each has its own ordinary reciprocal cortical seams and selector; the
sum of their agreement energies uses coupling 1.5 per head, total 3. Both
receive the same complete query, and `recall_next=False`. A head is not an
event followed by its successor. Only the final two labels teach selected
wrong-pair rows. No address, head-to-name assignment, head role, diversity
loss or external answer vote is supplied.

The current `primary_passed` flag establishes the declared acquisition,
transfer and memory-causality gates. It does **not** require different or
useful heads. The runner's `selected` field chooses the first successful arm
in H1, H2, wider H1, duplicated-port H1, larger-work H1 order. Neither field
is a parallel-head success flag. A successful H1 is evidence for retaining
the simpler mechanism.

Two different selector distributions alone establish no behavioral use.
One selector can potentially encode both facts in an analog mixture; this
task does not prove an information-theoretic need for two heads.

## Required data and fixed denominators

Assess founders 0 and 1 separately; both must satisfy every requirement
attached to a claim. Do not average one successful founder against one
failure. Keep all ten declared learned-reader cells, capped/unrun cells,
training and evaluation refusals, all 48 training rows and all 16 held-out
rows. The primary simultaneous-distinction panel is the full 16 held-out
rows, which all have unequal A/B values. Also report all 32 unequal training
rows and all 48 unequal rows of the complete panel. Equal-value rows remain
in the ordinary accuracy denominator; they are excluded only from this
predeclared distinction diagnostic because two equal outputs do not test
different value evidence.

For every condition retain exact case identity `(A value, B value, prefix
order, query order)`, both answers, refusal/status, source digest, association
state when present, head weights, and work. Report both query orders and
both prefix orders, with all rows retained. A missing/refused/unrun reading
does not count as an intervention effect or as selective evidence.

Recorded attention has shape `(heads, rows, slots)` and excludes the null
probability. For this raw, fresh-prefix assay the frozen witness code appends
three slots chronologically: first fact, second fact, neutral; reset clears
the bank, and capacity is at least three. Thus case prefix order determines
the A/B slot identities, while the null mass is `1 - sum(p_slots)`. Verify
this source contract, three completed prefix actions, resolved capacity and
exact attention shape before using that reconstruction. Reorder reverses
slots and ages together; donor replacement changes the actual values at the
same named slots. Erasures set the named record's held bit false.

The receipt does not explicitly serialize every pre-query K/V/held/age
array. This reconstruction is acceptable only for the frozen raw three-event
contract above. If that contract or slot identity cannot be established from
the retained source and rows, mark the selector diagnostic **unavailable**;
do not invent an address or run an uncharged reconstruction solve. Attention
arrays must be finite, nonnegative and sum to at most one, allowing 1e-10
rounding error. No malformed or filtered panel passes.

## Required parallel-head result

The thresholds below are chosen now as simple, conservative development
criteria. They are not fitted to a learned-reader outcome and are additional
claim requirements, not revisions of the frozen executable gate.

1. **Competent H2.** H2 passes the frozen primary and retention gates on both
   founders. This includes complete 512-round acquisition, no action/lesson
   refusal, .95 trained exact-pair/slot/unequal/query floors, at least .90
   held-out exact pair, at least .90 unequal swapped and held-out swapped
   exact pair, the frozen record/donor/sham controls, then all 256 retention
   rounds with .95 trained accuracy and at most .05 held-out loss. On 16
   rows, .90 requires at least 15 correct; a .05 retention tolerance allows
   no net loss in the correct whole-pair count. Report the actual counts.

2. **Two distinct witnessed facts in the same settle.** For an unequal row,
   call a head selective for A or B only when that named fact has probability
   at least .50 and exceeds each other category (other fact, neutral, null)
   by at least .10. A row has distinct evidence only when the two heads are
   selective for different named facts. Require this on at least .80 of all
   16 held-out rows (13/16), and at least .75 within each query-order and
   prefix-order marginal (6/8). Require at least .80 on the 32 unequal
   training rows (26/32), with at least .75 in each query/prefix marginal
   (12/16). Keep the complete per-row mass table. A/B specialization may be
   fixed or query-dependent; no particular head identity is required or
   retroactively taught. Broad complementary mixtures that miss these
   thresholds may be reported descriptively, not relabeled as this result.

3. **Both heads affect useful behavior.** From the existing acquired-copy
   causal panels, require every H2 `head_0`, `head_1` and `cloned_heads`
   reading to be completed, fully answered and unrefused. On the full held
   panel, each individual head lesion must lower exact-pair accuracy by at
   least .20 (at least four of 16 net cases), with a positive drop of at
   least .125 in each query marginal (one of eight net cases). The existing
   clone-head-0 interface intervention must lower held-out exact-pair
   accuracy by at least .20, also at least .125 per query marginal. Use the
   intact causal panel from the same host as the baseline, not a different
   evaluation run. Refusal or failure to settle is not a qualifying loss.
   Do not choose a different clone direction after seeing results.

4. **Fact-specific consistency.** For each head and every held row, define
   its preferred named fact by larger A/B mass (ties choose A solely to make
   the table total; tied rows cannot pass item 2). Its corresponding output
   slot is `query_order XOR fact_name`, with A=0 and B=1. Report the lesion's
   accuracy loss in that predicted slot and in the other slot over **all**
   16 rows. Require at least .20 loss in the predicted slot for each head.
   Also recompute the existing record-erasure relevant-slot losses on the
   full held panel and require at least .20 for both named records. Report
   other-slot losses without a selective-sparing claim. Retain the frozen
   donor held-out floor .90 and sham/reorder answer-change ceiling .05.

5. **An acquired coupled interface.** The recorded `frozen_interface` and
   `shuffled_interface` held-out panels must be complete and unrefused, and
   each lower exact-pair accuracy by at least .20 against the same intact
   causal baseline. These controls perturb the learned query/read-cortex
   interface together. They test reliance on that interface, not a pure
   query-address learning mechanism.

If all five items pass, the bounded claim is: **both acquired coupled heads
carry distinct witnessed fact evidence and contribute to the tested
simultaneous two-binding answer on both development founders**. Do not infer
general semantic codes, integrated continuing behavior or broad headcount
superiority. No fresh confirmation seeds are opened by this supplement.

## Separating head benefit from other resources

All comparisons use the existing same-founder 512-round schedule, identical
task/teaching/numerical settings, and full unfiltered panels. Report H1,
H1-width, H1-port-copies and H1-work alongside H2, on both founders. A capped,
unrun, refused or incomplete control does not establish its counterfactual
capability; it cannot be turned into evidence for a headcount advantage.

An additional **development performance benefit** requires H2 to satisfy
all five requirements above and exceed *each* of the four H1 comparators'
held-out exact-pair accuracy by at least .10 on *each* founder (with 16
cases, a difference of at least two net correct cases). All comparator cells
must have completed 512 acquisition rounds, have no training action/lesson
refusal, and have complete unrefused intact train/held/all panels. Report
their trained-floor and causal outcomes as well; do not erase a failed
trained floor from the comparison. Compare primary readings to primary
readings. Claim retention superiority only where the comparator actually
ran its conditional retention; otherwise claim H2's own bounded retention
only. If the simpler H1 matches H2, retain the simpler explanation.

For a **measured-work benefit**, additionally require the H2 result to
dominate a completed comparator on accuracy while using no more of each
reported resource: total neuronal coordinate updates (row sweeps times
actual neuron count), `cortical_synapse_transports + memory_scan_macs`, and
`softmax_elements`. Preserve each component and all refused/evaluation/
conditional-retention work; do not equate these reference counts with host
energy, or select a favorable cost denominator. Report host seconds and
parameter/neuron/synapse counts separately. H1-work's larger sweep ceiling
does not itself establish an equal actual work budget. An unmatched run
supports a behavior-at-these-points comparison, not a work-efficiency claim.

The controls reduce resource explanations but do not exactly isolate head
count. H1-port-copies preserves key/value/cue input isometry; nonlinear
neuron activation and different per-coordinate read restoring strengths
prevent an exact full-dynamics equivalence. H2's split alpha and independent
read-cortex seams also alter effective gain/conditioning. Report these
limits even if H2 beats all four controls. A pure headcount-causality claim
would require a separately declared gain/transfer-matched control, not
available by silently reinterpreting these recordings.

## Intervention and representation limits

`head_order` merely reverses memory descriptors while keeping physical
ports and synapses fixed. It is an algebraic/custody sham; it does not test
role exchange with downstream connections held fixed. Head lesions set one
amplitude to zero: they remove its restoring force as well as selected
content, while cortical ports remain live. Clone-head-0 copies both query
and read seams/biases, changing learned transfer as well as collapsing
selection. These are real coupled-channel interventions, not isolated
selector-only ablations. Their effects cannot by themselves identify which
part of a head caused the benefit. The frozen-interface control likewise
does not freeze every learned cortical parameter. No new diagnostic solve
or mechanism is authorized by this assessment.

The ridge representation probe uses already-recorded free association
states only, after the executable primary and retention gates. Its held-out,
donor-swap, raw-query, frozen-newborn and erasure controls retain the fixed
probe thresholds in source. A passing probe concerns contextual
accessibility, not a change to a held LearnedCode output or general meaning.
Its result cannot substitute for any parallel-head requirement above. A
pending/refused probe remains pending/refused.

## Host budget and incomplete comparisons

The learned-reader stage has 3000 **additional host seconds**. All ten
declared cells remain in order; each acquisition call is capped at 240
seconds, and each conditional retention call can also consume 240 seconds.
Ten acquisition caps alone total 2400 seconds; ten additional retention caps
would bring training alone to 4800, before construction, evaluation, causal
panels and probes. Thus 3000 is an enforced shared limit, not a guarantee
that all ten cells and all conditional retention/probes finish. Earlier
cells' retention can consume time before later resource controls open.

Use the existing deadline checks and retain every budget stop. A panel or
round already started may finish within its finite solver cap; do not start
additional panels after expiry. If the stage cannot complete the matched
controls, report **parallel-head/resource comparison unresolved within the
declared budget**. Do not count later unrun controls as H2 wins, extend a
deadline ad hoc, discard retention cost or open extra founder/recipe trials.
Any later additional budget or intervention requires a new declaration
before execution and preservation of the complete current outcome.

## Final assessment table

Report one row per founder and arm, containing acquisition/transfer/causal/
retention/probe status, full answers, recorded work and any cap/refusal.
Then report separate booleans for: competent H2; distinct simultaneous fact
evidence; both-head usefulness; acquired-interface reliance; performance
benefit against all four controls; measured-work benefit; and completed
representation probe. The runner's generic `passed`/`selected` values must
not replace this table. A failure or unavailable measurement remains
explicit; no threshold or case subset changes after outcomes.
