Bridge Memo — S5 + S6.
Two of the falsification register's existential-tier predictions — #24 (measurement perturbation) and #25 (reliability map vs weather) — sit in a structurally different position from the others. They are not about the dyadic, presence, network, or three-force layers of the framework; they are about the equation's own observability and reliability properties. They affect both the relational sibling (Talk QWERTY to Me) and the agent / operational track (KTP), and neither track should swallow them alone.
This memo is the bridge. It names the governance rule that decides when a finding from either track becomes canonical for both. It specifies the workplans (Methods A-D for each of S5 and S6). It carries the three-tier publishability rule (trait / climate / snapshot) that the binary map / weather framing of the master plan was forcing into false promotions and demotions. It records the Heisenberg Penalty reconciliation that closed the contradiction between the public KTP-Problems RFC §1 (SOLVED) and falsification register #24 (OPEN).
Neither S5 nor S6 is resolved here. The memo specifies the path from State A (open empirical question) to either falsification of the rival hypotheses or canonicalization under the Structural- Embodiment Test. The path is what is publishable. The findings are not yet.
This memo is the source-of-truth workplan behind two entries in the Falsification register: test 24 (Measurement Perturbation, S5) and test 25 (Reliability Band — Map vs Weather, S6). Both entries ship as proposed until the canonical TFE memory anchor refresh completes; promotion to active gates on that refresh.
Bridge findings require a named, portable governance rule. Without one, a finding from either track silently becomes canonical the moment it ships — and the framework loses the ability to distinguish a real cross-domain regularity from an instrument artifact. SET is that rule.
A bridge finding becomes canonical for both TQTM and KTP if and only if it can be stated at the structural level of the equation (referring to terms, the product of conditions, or operator properties common to both implementations) AND it has been shown not to depend on domain-specific embodiment (couples-only artifacts, agent-only artifacts, or instrument-specific properties). Findings that fail either half of the test are filed as domain-specific calibrations, not canonical findings.
- 1. RestateRestate the candidate finding in terms of the equation, not the instrument.
- 2. IdentifyIdentify the embodiment-specific artifacts the original observation depended on.
- 3. StripStrip them. If the claim survives the strip, it is canonical-eligible.
- 4. ReplicateIf it survives the strip but has been observed in only one track, mark canonical pending replication until the second track observes structurally.
- 5. RefuseIf it does not survive the strip, file as domain-specific and refuse to propagate.
Every canonical bridge finding must carry a trail noting (a) the strip operation that promoted it, (b) the original embodiment-specific observation, and (c) the cross-track replication evidence. This is the audit surface that prevents drift back into either domain.
S6 names the question of whether trust readings are maps or weather. The binary is too coarse to ship under. Below is the three-tier rule that decides how a given TFE term can be reported publicly, given its measured reliability band. The rule is itself a candidate canonical bridge finding — accepted alongside SET on 2026-05-15.
Why three tiers, not two: the binary map / weather framing in the master plan is the rhetorical version. The operational version needs the middle tier (climate) because most equation terms will likely land there, and shipping under the rhetorical binary forces false promotions to map or false demotions to weather.
Claim: Structured measurement perturbs trust state in a measurable way.
Rival / null hypothesis: observation is neutral; perturbation is below the noise floor; perturbation is indistinguishable from ordinary state volatility.
The word structured is load-bearing. Casual observation (a person noticing they are being watched, an agent's logs being read) is one condition; structured measurement (a deck administration, a flight-recorder pass, a mechanistic-interpretability probe) is a protocol with a defined surface area, schedule, and disclosure regime. The S5 claim is about the protocol class.
Run agent populations under two regimes that differ only in observability. Regime A1 (full flight recorder): every action tuple logged, every primitive challenged, full provenance trail published. Regime A2 (minimal observation): identical substrate, identical action surface, observability reduced to outcome-only logging. Measure: distribution of trust-value trajectories, extraction-shift incidence, witnessed-betrayal accumulation rate. If A1 and A2 produce statistically indistinguishable trajectories within the S6 noise floor, perturbation is below detection. If trajectories diverge in a direction, perturbation is measurable and its sign is the calibration target.
Anthropic-style mechanistic interpretability work is the closest existing empirical literature on whether instrumented observation perturbs agent internals. A4 does not assume access; the work is publishable as-is using public Anthropic outputs, but a calibrated KTP-side replication requires data access. Vasilenko 2026 (cognitive_core / attractor-like geometry) is a secondary anchor: the question of whether activation-space probing perturbs the attractor is structurally identical to S5.
Lineage already encodes the assumption that the deck IS the intervention. The test is not whether the deck changes couples (assumed) but whether the deck changes the measurement of the constructs the deck is trying to score. Pre-deck baseline via an independent instrument (validated relationship-quality measure outside the construct family); deck administration; post-deck re-measurement via the same instrument at t+1 day, t+1 week, t+1 month. Matched comparison cohort takes the independent instrument three times without the deck. Difference-in-differences on the independent instrument. The deck is reactive iff post-deck scores diverge from control in a way the deck's own scoring does not predict.
C745 (Recursive Substrate Observation) names recursive observation as a portable evaluation method. The S5 test design must include the recursive case: an instrument that measures its own perturbation. The clean version is a meta-deck (TQTM) or a meta-recorder (KTP) that scores the perturbation of the first-order instrument. If recursive observation diverges from first-order observation in a predictable direction, the perturbation function has a recoverable shape; if it diverges chaotically, the perturbation is noise-dominated and S5 collapses to S6.
Claim: Trust readings have a measurable reliability band distinguishing map from weather.
Rival / null hypothesis: readings are too volatile for map claims; readings are weather; single-session estimates are dominated by state noise; no stable trait-level signal exists.
Per-term framing: S6 is not a single number. The most likely outcome is term-specific reliability: some terms (asymmetric reciprocity, identity) are stable; others (capacity, availability) are volatile. The deliverable is a reliability map of the equation, not a single ICC.
Standard psychometric design. Administer the instrument at t, t+Δ₁, t+Δ₂, t+Δ₃, calibrated to plausible state-volatility windows. TQTM: t, +1 week, +1 month, +3 months (the 90-day lookback anchors the upper bound). KTP: t, +1 hour, +1 day, +1 week, +1 month (agent timescales differ; the spread covers session-state, deployment-state, drift-state). Report ICC(2,1) per term and per axis, with bands per term per interval.
The cleanest KTP-side instrument for S6 is a DryDock repeat-trial mode: same (persona, zone, action) tuple administered N times with controlled stochasticity, drift band reported per term. CRITICAL DEPENDENCY: DryDock does not currently compute the model it displays. Until the executable model exists, repeat-trial mode produces no term-level drift to measure. The S6 KTP-side workplan has a hard sequencing dependency on the A2 DryDock model fix. A4 publishes the protocol but cannot ship measurements until the instrument can produce them. Falsification register #25 marks this dependency explicitly.
For TQTM, where the live deck IS the only TQTM instrument, the retest design must use an external comparator (a validated relationship-quality measure outside the TQTM construct family). This is the same independent instrument the S5 deck-reactivity test uses. The two protocols can share a cohort.
Once both sides have term-level reliability bands, run a structural-level comparison: do the same equation terms show comparable reliability under both implementations? If asymmetric reciprocity is high-reliability in TQTM and low-reliability in KTP, that is itself a finding: either the term means something different across embodiments (violates structural-claim status of the term) or one instrument is mis-operationalizing it.
S5 effect sizes reported as multiples of S6 noise floor.
The S5 effect size is meaningless without an established S6 noise floor. S5 reads must be reported as multiples of the S6 noise floor, not in absolute units. This makes the two claims operationally coupled — the S6 workplan must complete first (or in tandem with the same instruments) so that S5 has a denominator. The A2 control plane sequences accordingly, and the falsification register #24 and #25 entries each carry this coupling explicitly.
The KTP-Problems RFC §1 declared the observer-effect SOLVED via separation of calculation and enforcement. That answer is correct for latency. It does not address state perturbation under structured measurement — the question S5 actually asks. If the falsification register ships #24 declaring measurement perturbation an open research question, the public surface contradicts itself.
Reopen ktp-problems.md §1 with explicit scope. Split into §1a (Latency: SOLVED) and §1b (State Perturbation: OPEN, pointing to falsification register #24). Version 0.1 → 0.2.
2026-05-15.
Option (a) over option (b) on epistemology-conservation grounds. The latency answer is preserved (it was always correct as a latency answer). The new finding is added (a different question lives under the same observer-effect umbrella). The public RFC no longer overclaims; the falsification register entry no longer dances around an RFC contradiction.
- This memo published as a companion to /research/falsification.
- A1 #24 and #25 entries populated into falsifiers.ts (shipped 2026-05-15).
- Heisenberg Penalty RFC §1 split into §1a Latency (SOLVED) and §1b State Perturbation (OPEN, → #24); Version 0.1 → 0.2 (shipped 2026-05-15).
- DryDock model fix specced (A2 owns), giving S6 KTP-side instrumentation a target ship date.
- TQTM cohort study design drafted, covering S5 Method C and S6 Method A on TQTM side as a shared cohort.
- DryDock model shipped. Repeat-trial mode operational. First S6 KTP-side reliability bands published per term.
- KTP A/B observability regime instrumented. First S5 KTP-side trajectory comparison data collected.
- TQTM cohort study fielded (N ≥ 50 dyads, three retest intervals, S5 Method C + S6 Method A on TQTM side).
- First cross-track structural comparison attempted: do term-volatility orderings converge between TQTM and KTP?
- Mechanistic-interpretability anchor work scoped; synthesis essay published if not waited on for primary data.
- S5 and S6 each have a defensible canonical finding per SET: a structural claim replicated across both tracks at least once.
- Three-tier publishability rule applied to every TFE term in both implementations; per-term reliability map published in the canonical TFE reference (/learn/trust-force-equation).
- The perturbation-sign question (stabilizing vs destabilizing) answered in at least one stake-stratified regime.
- Recursive observation Method D run on at least one track; result either promotes the recursive instrument to canonical or files it as a domain-specific calibration.
- If S5 finds large destabilizing perturbation: instrument redesign across both tracks (covert observation mode for KTP, meta-deck for TQTM). A4 reopens.
- If S6 finds widespread snapshot-only terms: public-surface revision to stop reporting them as traits; canonical TFE page updated.
This memo is a companion to /research/falsification. The register entries it informs are test 24 (S5, measurement perturbation) and test 25 (S6, reliability map vs weather).
Source: A4 Bridge Memo, filed 2026-05-15 as Job 6 of TFE Research Triage v3. Honors epistemology conservation per master plan §14: provenance, rationale, evidence state, and decision trail preserved.